Operating by John Brewton

Subscribe to Alex’s Substack, The Strategy Stack, here.

IV. Explanation Is Becoming Cheap. Judgment Is Not.

Every major technological shift changes the economics of something. The steam engine reduced the cost of physical labour. Industrial manufacturing reduced the cost of production. Computing reduced the cost of calculation. The internet reduced the cost of accessing information.

Generative AI appears to be changing something altogether different. It dramatically lowers the cost of producing explanations.

That distinction deserves careful attention because explanations occupy an unusual position within organizations. They satisfy a deeply human need to make events appear coherent. A strategy failed because the market shifted. A customer left because pricing was too high. A competitor succeeded because its product was superior. Most organizational life is surrounded by narratives that transform complex, multidimensional events into stories that feel understandable enough to support action.

The problem is not that explanations are inherently misleading. Organizations cannot function without them. Every roadmap, investment proposal, product strategy and board presentation depends on a shared interpretation of reality. The difficulty lies elsewhere. Coherence and accuracy are related, but they are not identical. A persuasive explanation may still overlook critical variables, underestimate uncertainty or mistake correlation for causation. History is filled with decisions that appeared perfectly reasonable because the underlying narrative was internally consistent, only for reality to reveal later that the organization had misunderstood the problem from the very beginning.

Generative AI introduces a fascinating dynamic into this relationship because it produces coherence with remarkable ease. Faced with an unfamiliar question, the system does not experience hesitation, uncertainty or intellectual discomfort. It generates a plausible interpretation almost immediately, drawing together patterns from enormous volumes of text into a response that feels structured, complete and often highly persuasive. Even when the underlying reasoning is imperfect, the explanation itself carries many of the characteristics that humans naturally associate with understanding: fluency, confidence, logical progression and linguistic precision.

From the perspective of organizational psychology, this matters because human beings rarely experience uncertainty as an abstract philosophical concept. We experience it emotionally. Uncertainty creates tension. It interrupts momentum. It delays decisions. It exposes the possibility that our current understanding may prove inadequate. Throughout history, that discomfort encouraged investigation. The effort required to reduce uncertainty often became the mechanism through which better knowledge emerged.

AI changes that emotional sequence.

The question is asked.

An explanation appears.

The feeling of uncertainty recedes.

Whether uncertainty itself has actually diminished remains an entirely separate question.

This difference between objective uncertainty and subjective certainty may become one of the defining management challenges of the coming decade. Organizations have always operated in environments characterised by incomplete information, delayed feedback and complex causal relationships. None of those conditions disappear because explanations become easier to generate. Markets do not become more predictable because summaries become more elegant. Customers do not become more consistent because presentations become more convincing. Competitive advantage does not emerge because every employee can now produce a beautifully written strategic analysis in under five minutes.

Reality has not become simpler.

Only our interface with it has.

Perhaps the most significant consequence is not technological but social. Every organization depends on countless small signals that reveal where understanding remains incomplete. Someone asks another department for help. An engineer escalates an issue to an architect. A consultant admits they need to investigate further before making a recommendation. A physician requests another opinion. These interactions do more than exchange information. They reveal where uncertainty currently resides inside the organization, allowing expertise to flow toward it.

If explanations become permanently available, many of those signals begin to disappear. Employees no longer need to expose uncertainty before receiving a plausible answer. Teams become increasingly capable of operating independently, but independence and understanding are not synonymous. An individual equipped with an excellent language model may appear highly informed while remaining disconnected from colleagues whose experience would have fundamentally altered the interpretation of the problem. Knowledge becomes more accessible at precisely the moment collective sensemaking risks becoming less frequent.

This is an irony worth reflecting upon. Organizations adopted digital collaboration platforms to connect people more effectively across functions and geographies. Generative AI may inadvertently reverse part of that dynamic by allowing individuals to resolve questions privately that previously required conversation. The immediate productivity gains are obvious and, in many cases, entirely justified. The longer-term organizational effects are more ambiguous. Every question answered without conversation is also a conversation that never takes place, an assumption that is never challenged and a perspective that never enters the decision-making process.

The implications extend beyond collaboration into the architecture of organizational learning itself. Much of what organizations know is not stored in documents, databases or process manuals. It exists in the interactions through which people compare interpretations, negotiate meaning and gradually refine their understanding of ambiguous situations. Organizational theorists have long distinguished between explicit knowledge—the things that can be documented—and tacit knowledge—the intuitions, experiences and contextual judgments that resist formalization. AI dramatically improves access to the former. Its influence on the latter remains far less certain.

Perhaps the greatest risk, then, is not that AI produces incorrect answers. Incorrect answers have always existed. Organizations have developed countless mechanisms for discovering and correcting them. The more subtle possibility is that explanations become so immediate, so abundant and so persuasive that they quietly reduce the number of occasions on which people realise they have reached the boundary of their own understanding. If that boundary becomes less visible, organizations may discover that they have become exceptionally efficient at producing explanations while simultaneously becoming less effective at recognising when explanation should give way to further inquiry.

V. Designing Organizations That Preserve Uncertainty

The discussion surrounding agentic AI often begins with capability. How many tasks can an agent perform? How many systems can it connect? How much human intervention can be removed from a workflow? These are sensible engineering questions, but they say surprisingly little about organizations themselves. Enterprises have never struggled because they lacked the ability to execute instructions. They struggle because they repeatedly execute assumptions that later turn out to be incomplete.

Viewed through that lens, the next generation of organizational design may depend less on what AI can accomplish autonomously and more on the role it plays within the organization’s broader process of judgment. Every technology eventually becomes embedded in an institution, and institutions shape technology at least as much as technology reshapes institutions. The real design challenge, therefore, is not simply building more capable agents. It is deciding which organizational functions those agents should strengthen and which human capabilities they should deliberately leave intact.

Current development trends reveal a strong preference for completion. Agents draft reports, prepare presentations, write code, summarize meetings, coordinate workflows and increasingly orchestrate other agents. The underlying objective is remarkably consistent: reduce the amount of work that remains unfinished. Yet completion has never been the defining characteristic of high-performing organizations. A completed analysis that frames the wrong problem creates more risk than an unfinished one that still invites challenge.

That distinction suggests a different design philosophy.

Instead of asking how an agent can replace the next human task, organizations might begin by asking which forms of uncertainty disappear too early in their existing decision processes. The objective shifts from maximizing automation toward improving the quality of organizational sensemaking.

An agent designed around that principle behaves very differently.

Rather than presenting a single recommendation, it identifies competing interpretations.

Rather than optimizing for confidence, it estimates the quality and completeness of the available evidence.

Rather than summarizing agreement, it highlights disagreement.

Rather than producing the shortest path to execution, it identifies where additional observation would meaningfully change the decision.

The result is not a slower organization.

It is an organization that allocates its attention more intelligently.

That distinction becomes easier to appreciate when viewed through familiar organizational practices.

Conventional Agent

Epistemic Agent

Produces answers

Maps uncertainty

Completes tasks

Identifies missing evidence

Optimizes efficiency

Optimizes decision quality

Reinforces existing assumptions

Challenges existing assumptions

Maximizes confidence

Calibrates confidence

Ends the workflow

Decides whether the workflow should continue

This is not merely a different interface.

It represents a different philosophy of organizational intelligence.

Much of today’s AI landscape assumes that knowledge work consists primarily of producing outputs. Reports, presentations, analyses and recommendations become the measurable products through which productivity is evaluated. Yet organizations rarely derive competitive advantage from documents alone. They compete through the quality of the decisions those documents support. A beautifully written recommendation based on an incomplete understanding of the environment remains strategically inferior to a less polished analysis that exposes uncertainty before resources are committed.

Seen from this perspective, the role of agentic systems begins to resemble something closer to institutional architecture than digital labour. Good institutions have never existed to eliminate disagreement. Their purpose has been to ensure that disagreement appears early enough to improve judgment. Constitutional courts, scientific peer review, investment committees and aviation safety investigations all slow the journey from observation to action in carefully chosen places because experience has demonstrated that unchecked certainty becomes extraordinarily expensive.

Perhaps AI should be designed with the same principle in mind.

Instead of asking whether an agent can make a decision independently, organizations might ask a different set of questions.

Which assumptions is this recommendation built upon?
What evidence would most likely change this conclusion?
Which perspectives are currently absent?
Where does confidence exceed available evidence?
Which part of this decision still depends on human judgment rather than computational capability?

Notice that none of these questions reduce the value of automation. They simply recognize that automation and judgment solve different organizational problems. One increases operational capacity. The other determines whether that capacity is directed toward the right objective.

This distinction becomes increasingly important as organizations evolve from isolated AI tools toward networks of collaborating agents. A single language model answering questions for an employee introduces one set of challenges. Hundreds of specialized agents coordinating product development, legal review, procurement, finance and customer operations introduce another. Once agents begin exchanging conclusions with one another, organizations face the possibility of creating systems that reinforce certainty at machine speed. If every agent accepts the outputs of every other agent without introducing mechanisms for challenge, contradiction or independent verification, the organization risks scaling coherence faster than understanding.

Perhaps the most valuable agent inside an enterprise will therefore not be the one that writes the report, negotiates the contract or schedules the meeting.

Perhaps it will be the one that interrupts.

The agent that notices conflicting assumptions between departments before anyone else does.

The agent that detects unusual confidence unsupported by evidence.

The agent that identifies the missing customer perspective in an otherwise compelling strategy.

The agent that quietly asks whether the organization has begun answering a question it never properly understood.

Those responsibilities sound remarkably familiar because organizations have always depended on people who performed exactly that role. The difference is that they rarely occupied positions of formal authority. They were the experienced architect who challenged an elegant design, the analyst who questioned the assumptions hidden inside a financial model, the physician who requested one additional test despite overwhelming consensus, or the strategist who noticed that everyone had become unusually comfortable with the current narrative.

The future of agentic AI may ultimately depend on whether we choose to replicate only the visible work those people performed or whether we also preserve the invisible function they served within the organization. The first path produces faster execution. The second strengthens institutional judgment itself.

VI. Judgment as Competitive Advantage

Every generation tends to believe that the defining technologies of its time fundamentally alter the conditions under which organizations compete. Sometimes they do. Railways changed geography. Electrification transformed manufacturing. The internet redefined distribution. Artificial intelligence will undoubtedly reshape many aspects of organizational life in ways we can only partially anticipate today.

What history also suggests, however, is that enduring competitive advantage rarely emerges from technology alone. Technologies diffuse. Best practices spread. Capabilities that once distinguished a handful of organizations gradually become available to everyone else. What remains difficult to replicate are the institutional habits through which organizations interpret reality, revise their assumptions and make decisions under conditions that remain fundamentally uncertain.

This observation deserves greater attention because it changes the way we think about intelligence itself. The coming decade will almost certainly produce organizations capable of generating extraordinary volumes of analysis, recommendations, simulations and strategic options. Access to explanation will continue to expand, while the cost of producing sophisticated reports, market assessments and technical evaluations will continue to fall. In many industries, these capabilities will become expected rather than exceptional. They will shape the baseline from which organizations operate, not the advantage that separates one organization from another.

The differentiator will increasingly shift elsewhere.

Not toward who has the most explanations.

Nor toward who automates the largest number of workflows.

But toward those organizations that remain capable of recognizing when their own explanations have become inadequate.

That capability has always occupied an unusual position within strategy. Markets rarely reward organizations for answering yesterday’s questions more efficiently than their competitors. They reward organizations that notice emerging questions while others continue refining outdated answers. Every major strategic inflection point contains this asymmetry. One organization continues optimizing an existing model because the evidence still appears internally coherent. Another begins questioning assumptions that have not yet visibly failed. Looking backward, these moments are often described as innovation or visionary leadership. Looking more closely, they frequently begin with something less dramatic: an institution preserving enough epistemic humility to recognise that reality may already have moved beyond its current understanding.

Seen in this light, judgment begins to look less like an individual attribute and more like an organizational capability. It is embedded in governance structures, decision processes, incentive systems and cultural norms that determine whether uncertainty is explored or suppressed. Organizations do not become adaptive simply because they employ intelligent people. They become adaptive because they create environments in which intelligent people can revise their thinking without treating revision as failure. Learning is not the accumulation of correct answers; it is the continuous recalibration of mental models as reality changes around them.

Generative AI enters this landscape as both an extraordinary opportunity and a subtle design challenge. It dramatically expands humanity’s capacity to produce explanations, synthesize information and coordinate increasingly complex work. Used well, these capabilities can elevate the quality of organizational decision-making in ways that were previously unimaginable. Used without careful attention to organizational design, they may also encourage institutions to mistake fluency for understanding and completion for judgment. The technology itself remains neutral. Its consequences depend on the architecture into which it is introduced.

Perhaps that is the deeper lesson emerging from this moment.

Organizations have spent decades investing in systems that improve access to information. We are now entering an era in which access to explanation is becoming equally abundant. The question quietly moving toward the centre of management is no longer whether people can obtain answers. Increasingly, everyone can. The more consequential question is whether organizations continue to preserve the conditions under which answers are challenged, refined and occasionally abandoned altogether.

That challenge reaches beyond management practice. It touches the way societies organize expertise, distribute authority and cultivate trust. Universities, scientific communities, engineering disciplines and democratic institutions all evolved around the recognition that certainty benefits from structured disagreement. Their objective was never to eliminate uncertainty but to prevent confidence from outrunning evidence. As organizations become increasingly AI-native, they inherit the same responsibility. The technologies may be unprecedented, yet the institutional challenge remains remarkably familiar: preserving the quality of judgment while expanding the scale of intelligence.

John and I approached this question from different directions. One of us spends much of his time thinking about how organizations build strategic capability, how decisions shape competitive advantage and how institutional learning becomes an enduring source of resilience. The other explores the architecture of agentic systems, orchestration and the operational models that will increasingly define AI-native enterprises. Those perspectives converge on a common conclusion. The future of organizational intelligence will not be determined solely by the sophistication of the systems we build, but by the quality of the environments into which we introduce them. The design of an agent and the design of an institution are becoming inseparable questions.

Perhaps that is where this conversation ultimately leads.

The defining organizations of the AI era will almost certainly produce better analyses, faster recommendations, and more efficient workflows than any generation before them. Those achievements will matter, but they are unlikely to be sufficient. Long after explanation becomes inexpensive and ubiquitous, organizations will continue competing through something that cannot be downloaded, prompted, or automated in isolation: the collective discipline of recognizing where their understanding remains incomplete, resisting the temptation of premature certainty, and creating institutions in which better questions continue to emerge before better answers are accepted.

Subscribe to Alex’s Substack, The Strategy Stack, here.

The Infinite Value of a Meeting Transcript for the AI-Fluent Operator: How to 100X the Value of Meeting Transcripts in Your Workflows and Business

John Brewton — Tue, 18 Aug 2026 10:22:18 GMT

Most companies are using AI to solve the least interesting problem with meetings.

They record the call. They generate a transcript. They ask AI for a summary. Someone skims it, closes the window, and the team moves on with their business.

That is useful. But it dramatically underutilizes the opportunity at hand. A meeting transcript is not merely a record of a conversation. It is a structured source of decisions, commitments, customer intelligence, objections, deadlines, requirements, ideas, risks, owners, and institutional knowledge.

In an AI-NAtive company, all of those things can become inputs into other systems.

The transcript can update the CRM. It can create tasks. It can draft the follow-up email. It can write the project brief. It can generate a proposal. It can prepare an executive summary. It can tell someone who missed the meeting exactly what they need to know. Months later, an AI agent can search thousands of previous conversations and answer a question nobody remembers discussing.

That changes the purpose of the transcript entirely.

The transcript should not simply tell us what happened in the meeting. It should reduce the probability that another meeting is required.

That may eventually become one of the most important operating advantages of AI.

A Practical Example: Turning One Transcript Into a Finished Asset

Watch the Loom video here.

I wanted to show one small example of what this looks like in practice.

Every Friday, I host Operating Founder Office Hours for members of my paid Substack community. We spend the session working through questions from operators, builders, solopreneurs, and small-business owners. Those conversations can move across dozens of ideas, resources, frameworks, and problems in a single hour.

Historically, the recording itself would have been the primary asset.

Now the transcript becomes the starting point.

In the Loom above, I show Claude pulling the recording and transcript directly from Zoom and creating a comprehensive, anonymized set of session notes. I then have it run my Elite Research skill across the conversation, identifying the concepts, books, academic research, articles, and thought leadership connected to what we discussed.

That produces an expanded research library members can use to keep learning long after the call ends.

Finally, I use my JB Personal Brand skill to turn everything into a polished PDF that can be distributed alongside the recording.

One conversation becomes a summary, a knowledge resource, a research library, a learning tool, and a finished community asset without requiring me to manually reconstruct the session.

Members can then feed the document back into Claude or ChatGPT, ask which resources are most relevant to them, explore an individual topic, or have an extended conversation about the research.

Record it once. Structure it once.

Reuse the intelligence many times.

The same principle becomes much more powerful when applied across an entire company. Here is a sample of the weekly session notes that all Operating Founders receive from our weekly calls:

What You’ll Learn in This Article

Why is the meeting transcript becoming an operating layer rather than just a record of what happened?
How transcripts can automatically improve CRM quality by capturing decisions, objections, commitments, stakeholders, and next steps.
How AI can turn meetings into execution by creating tasks, assigning owners, identifying deadlines, and surfacing blockers.
How better transcript workflows can eliminate future meetings, especially recap, handoff, status, and information-sharing calls.
How one conversation can become an asset factory, producing follow-ups, briefs, proposals, project plans, and internal documentation.
Why thousands of transcripts can become proprietary institutional intelligence that teams and AI agents can search and learn from.
The new KPI for meetings: not how well they were summarized, but how much work they created—and how many future meetings they made unnecessary.

What You’ll Get With This Article

The Meeting Transcript Operating System: Paid subscribers get a complete interactive workbook designed to be uploaded directly into Claude or ChatGPT. Bring one real meeting transcript, and the workbook walks you through building a repeatable system for extracting decisions and commitments, routing intelligence into your CRM and other tools, creating high-value assets automatically, identifying meetings that can be eliminated, and turning the finished workflow into reusable AI instructions, skills, scheduled tasks, and agents. You leave with a working Transcript-to-System Map, CRM Extraction Schema, Asset Factory, Meeting Elimination Report, Transcript Agent Instruction Set, Transcript Value Scorecard, and a 30-day implementation plan. This is not simply another set of prompts.

Become a Paid Subscriber

The Quiet Builders Are Coming. Great Companies Will Build Around Them.

John Brewton — Sat, 15 Aug 2026 12:30:06 GMT

A few years ago, we became obsessed with the concept of quiet quitting.

People stopped answering emails at night. They stopped volunteering for work that was not theirs. They stopped treating every request from an employer as an opportunity to prove how badly they wanted the next promotion.

We interpreted much of this as a decline in ambition.

I think we got that wrong.

The ambition did not disappear. It shifted directionally.

The person who no longer wants to spend Saturday preparing another presentation for their employer may be spending Saturday building a product of their own.

The marketer who stopped obsessing about becoming VP may be building an audience. The accountant may be learning how to code with AI. The salesperson may be testing a consulting business. The executive may be discovering that becoming CEO of a company she owns sounds more interesting than becoming EVP of a company she does not.

This is not quiet quitting.

It is quiet building.

Quiet quitting was about withholding additional effort from someone else’s company.

Quiet building is about redirecting that effort toward your own.

And I increasingly believe that quiet building will become one of the defining professional behaviors of the AI economy.

Jobs are becoming bundles of tasks that machines can increasingly help us complete. Skills are changing faster. Organizational capability is becoming less dependent on organizational headcount. Individuals are gaining access to capabilities that once belonged almost exclusively to companies.

The professionals who recognize this have an opportunity to become dramatically less dependent on a single employer.

The companies that recognize it have an opportunity to build organizations filled with some of the most curious, commercially minded and ownership-oriented people in the workforce.

Companies are becoming tech stacks.

We are all becoming companies.

And both sides of the employment relationship need to start acting in accordance with the reality of this change.

The Safe Job Is Becoming a More Complicated Idea

For most of the last century, one of the safest economic strategies available to an ambitious person was relatively straightforward.

Get educated.
Develop a valuable skill.
Find a good company.
Become increasingly valuable to that company.
Get promoted.
Earn more.
Save.
Invest.
Retire.

There is nothing wrong with that strategy. Millions of people will continue to build wonderful lives that way. But I think it is dangerous to assume the structure of the job sitting underneath that strategy will remain as stable as it once appeared.

The International Monetary Fund estimates that roughly 60 percent of jobs in advanced economies are exposed to AI. Importantly, exposure does not mean elimination. In roughly half of those exposed jobs, the IMF believes AI could complement human work and increase productivity.

The International Labour Organization reached a similarly important conclusion. Its latest global work on generative AI says roughly one in four jobs worldwide has some degree of GenAI exposure, but job transformation is more likely than wholesale replacement because most occupations still contain tasks requiring human involvement.

The question employees should be asking is the following:

How much of the work currently inside my job still needs to be done by me three years from now?

The U.S. Bureau of Labor Statistics now explicitly incorporates AI into its employment projections. BLS expects AI and other automated systems to increase productivity in some occupations while reducing labor demand in portions of office support, administrative work, sales and other task-heavy categories. At the same time, it projects substantial growth in technology-intensive occupations such as data science.

I do not think the intellectually serious argument is that there are no safe jobs anymore. It is the following: It is becoming less safe to assume that today’s job description, today’s skills, and today’s workload will remain intact.

Increasingly, I think career security will be about becoming valuable to a market, not a single company.

Part of Your Job Will Disappear

We keep discussing AI at the occupational level.

Will AI replace programmers?

Will it replace marketers?

Lawyers?

Accountants?

Consultants?

Writers?

That framing misses the way work actually happens. A job is not one thing. It is a collection of things.

Researching. Writing. Scheduling. Analyzing. Calculating. Communicating. Reporting. Creating presentations. Updating systems. Answering questions. Coordinating people. Making decisions.

AI does not need to eliminate the entire collection to fundamentally change the economics of the job. It simply needs to begin compressing enough of the tasks inside it.

In one of the most important real-world studies of generative AI at work, researchers from Stanford and MIT studied thousands of customer-support workers. Access to a generative AI assistant increased productivity by roughly 14 to 15 percent, with the largest improvements concentrated among newer and less-skilled workers.

The broader Stanford AI Index now documents productivity improvements across multiple types of knowledge work, though the gains vary considerably depending on the task and the worker. That variability is important. AI is not magically making every worker twice as productive, and we should be suspicious of anyone pretending that it is.

But even modest productivity gains create an organizational question that becomes more important as the technology improves:

What happens when a role designed around 40 hours of tasks no longer contains 40 hours of economically valuable human work?

Maybe it contains 35.

Then 30.

Maybe a year later it contains 20.

The traditional organization has an instinctive response to this.

Fill the space. Add responsibilities. Add meetings. Add reporting. Add another project. Turn a productivity gain into more activity.

I think the better organizations will eventually ask a harder question.

Do we actually need this job to remain exactly what it was?

Companies Are Becoming Tech Stacks

This is the organizational transformation underneath all of this. The traditional company accumulated capabilities largely by accumulating people.

If you needed more sales capacity, you hired salespeople.
More research required researchers.
More design required designers.
More content required marketers.
More analysis required analysts.
More software required developers.
More administrative capacity required administrative employees.

There was a fairly direct relationship between the amount an organization could do and the number of qualified humans it could employ.

According to the OECD’s 2026 research on skills in the AI age, the share of firms using AI across OECD countries with available data rose from roughly 7 percent in 2021 to about 20 percent in 2025. The OECD attributes part of that acceleration to the arrival and diffusion of general-purpose generative AI tools. (OECD)

We are moving toward organizations composed of some combination of:

Employees + AI models + agents + software + data + automation + contractors + systems.

The company itself is starting to look more like a technology stack. Adding organizational capability no longer always requires adding organizational headcount.

A ten-person company can increasingly perform work that once required 20.

A five-person company can access research, design, development and marketing capabilities that once belonged to organizations with entire departments.

A single extraordinary operator can increasingly orchestrate systems that perform work previously distributed across several specialists.

If that continues, we should expect two things to happen at the same time. (1) Companies will need fewer humans for certain categories of work. (2) And the humans they do employ will need to become much more capable.

That is where the quiet builder becomes incredibly interesting.

The Employee You Are Afraid of Losing May Be Exactly the Employee You Need

Think about the characteristics companies say they want from their employees.

Ownership.
Curiosity.
Initiative.
Commercial awareness.
Comfort with ambiguity.
Problem solving.
Adaptability.
Technology fluency.
Continuous learning.
Resourcefulness.

The World Economic Forum’s Future of Jobs research found that employers expect 39 percent of workers’ core skills to change by 2030. AI and big data are among the fastest-growing areas of skill demand, but employers also place tremendous value on analytical thinking, creative thinking, resilience, flexibility, agility and leadership.

We have another word for a person obsessed with developing most of those characteristics. Entrepreneur.

Companies are desperately trying to make their employees think more like owners, yet often become deeply uncomfortable when those employees begin to act like owners.

They want entrepreneurial thinking.

Just not too much entrepreneurship.

They want curiosity.

As long as that curiosity stays within the organization's walls.

They want employees obsessed with learning AI, understanding customers, improving margins, finding opportunities and creating systems. Then the employee starts a small company on the weekend and suddenly everyone gets nervous.

I think that mentality is increasingly backward.

Because the quiet builder may be exactly the employee you want.

Would you like to work with John?

A Side Business Can Be the Best Leadership Program Your Company Never Built

Something changes when you try to build a company.

I experienced it myself as an operator, and I see versions of it constantly in the founders and professionals I work with.

Learning stops being book-based and theoretical.

You do not learn sales because someone told you sales is important.
- You have to sell something.
You do not learn financial management because your company added a finance module to its leadership program.
- You are spending your own money.

You begin to understand ownership by owning something.

A quiet builder starts noticing costs. They start recognizing inefficient processes. They ask why the customer wants something rather than merely completing the request. They understand that revenue does not magically appear. They start thinking about distribution. They understand the value of speed. They begin to ask whether a process really needs seven steps, because, in their own company, every unnecessary step costs them time.

Quiet building can become an employee-development program the employer does not have to design or pay for.

Yes, companies need sensible restrictions.

Employees cannot steal intellectual property. They cannot misuse confidential information. They cannot build competing products on company time. They cannot allow an outside project to destroy their performance.

Those are reasonable boundaries.

But there is an enormous gap between protecting the business and believing the business should own someone’s entire professional ambition.

The companies that understand that difference will have an advantage.

Stop Treating Every Future Founder Like a Retention Failure

Companies have spent decades becoming obsessed with retention.

In most organizations, the logic is straightforward:

Great employee stays = win.

Great employee leaves = loss.

But what if that is the wrong time horizon?

Imagine a talented employee who has spent five years at your company. During the final two years, she begins building something on the side. The business grows. Eventually she reaches the point where keeping a full-time job no longer makes sense.

The traditional company views what happens next as attrition.

I think the more interesting company asks:

What should our relationship become now?

Maybe she becomes a consultant.
Maybe her new company becomes a vendor.
Maybe you become her first enterprise customer.
Maybe she develops technology your company licenses.
Maybe she helps train your people.
Maybe she becomes a distributor.
Maybe she sends extraordinary candidates back to you.
Maybe five years later you invest in her company.
Maybe you acquire it.

The employment relationship can end without the economic relationship ending.

That suggests a different philosophy of retention. Do not retain the employment contract at all costs. Retain the relationship.

The best relationship between two people today may be between an employer and an employee. Five years from now, it may be customer and vendor. Ten years from now, it may be two CEOs sitting on each other’s boards.

Organizations built for the next era should be capable of letting that relationship evolve.

Ten Ways Companies Can Support Quiet Builders

If quiet building is going to increase, companies need more sophisticated ways to work with it.

Here is where I would start.

1. Make responsible outside building culturally acceptable

Stop automatically equating a side business with disloyalty.

Create clear rules around conflicts, intellectual property, time and competitive activity. Then give responsible adults room to build.

2. Measure output instead of theatrical busyness

AI makes this increasingly important.

If an extraordinary employee can perform excellent work in 25 hours that once required 40, do not immediately reward that person with 15 hours of invented activity merely to preserve the appearance of a full calendar.

Reward outcomes.

3. Give builders problems instead of tasks

People developing an ownership mentality should be given opportunities to use it.

Tell them what needs to become true.

Give them appropriate authority.

See what they build.

4. Teach employees the economics of the company

Show people how revenue works.

Teach margins. Pricing. Customer acquisition. Retention. Cash flow. Capital allocation.

The more employees understand how the machine makes money, the more capable they become of improving it.

5. Let curious people cross functional boundaries

A future founder should understand more than marketing or finance or operations.

Let strong people learn how the other pieces work.

The resulting employee will become more valuable to you even if they never leave.

6. Subsidize AI fluency

If AI is going to change the structure of work, your people should become unusually capable with it.

Do not limit AI education to a two-hour compliance course.

Give people models. Tools. Sandboxes. Training. Time to experiment. Problems worth solving.

7. Create places to build inside the company

Not every entrepreneurial person wants to leave.

Give employees opportunities to launch products, automate workflows, test new businesses, create internal tools and own new initiatives.

A builder with enough surface area to build may happily keep building for you.

8. Experiment with flexible employment

A full-time job does not need to be the only relationship available.

Certain roles can evolve into four-day schedules, fractional positions, reduced-hour arrangements and project-based relationships.

Not every job allows this.

More jobs will than companies currently assume.

9. Build an employee-to-vendor pathway

When a quiet builder reaches the point where the outside company deserves full-time attention, have a mechanism for asking whether your company should become a customer.

Graduating from payroll to accounts payable does not have to be a breakup.

It can be a promotion in the relationship.

10. Build an extraordinary alumni network

Today’s employee could become tomorrow’s customer, supplier, consultant, founder, investor, referral source, strategic partner or acquisition.

Think in decades.

A company that produces successful founders should celebrate that reputation, not hide from it.

Quiet Building Is Career Insurance

Your side project does not need to become a $10 million company to completely change your career.

Maybe it creates $25,000 a year in additional income.
Maybe it gives you 10,000 people who pay attention when you publish something.
Maybe it teaches you to sell.
Maybe it turns you into the best AI operator in your field.
Maybe it gives you 20 customers who know that you can solve a specific problem.
Maybe it builds relationships with people across your industry.
Maybe it establishes your reputation around an expertise that exists independently of your employer.
Maybe it eventually becomes a company.

Every one of those outcomes creates leverage. And leverage reduces dependence. For most of the industrial era, the safest career strategy was to become increasingly indispensable to a single powerful institution.

That institution had the capital, the people, the technology, the customers, distribution, and brand. You needed access to the institution to access those capabilities. That is changing.

The old career insurance policy was becoming indispensable to one employer.

The new one may be becoming valuable to a market.

We Are Already Becoming Companies

America already contains an enormous economy of companies without employees.

According to the latest U.S. Census Bureau data, there were approximately 30.4 million U.S. nonemployer businesses in 2023 generating nearly $1.8 trillion in receipts. (Census.gov)

These are businesses with no paid employees.

Consultants.
Creators.
Tradespeople.
Independent professionals.
Small merchants.
Owners of rental businesses.
Freelancers.
Specialists.
Entrepreneurs of every imaginable kind.

The important point is not that all 30.4 million of those businesses resemble the AI-powered companies I am describing. They clearly do not. The point is that operating as a company without building a traditional employee base is already an enormous part of the American economy.

Now place AI on top of it. One person can increasingly access capabilities that once required entire teams, from sophisticated research, software development, design, financial modeling, writing, and video production to marketing, automation, data analysis, customer support, sales enablement, and strategy.

And increasingly, AI agents can do more than assist with individual tasks. They can execute multi-step workflows, coordinate work across tools, and give a single operator leverage that once belonged almost exclusively to larger organizations.

You no longer necessarily need to hire a company before you can become one.

Quiet Building Is Not Quiet Quitting 2.0

There is an important difference between these ideas.

Quiet quitting says:
My employer does not own more of me than it pays for.
Quiet building adds something much more important:
And I am going to invest the rest of me somewhere.

Quiet quitting asks how much work you are obligated to give. Quiet building asks what you are going to do with the ambition, curiosity, technology and time that remain.

That is why I find quiet building so much more exciting.

It is not anti-company.

It is not anti-career.

It is not anti-work.

It is pro-ownership.

And the best companies should recognize the value they can capture by embracing this changing dynamic in the market for work.

Companies Are Becoming Tech Stacks. We Are All Becoming Companies.

This may be the larger transition hiding underneath everything else. Companies are becoming smaller collections of extraordinarily capable people orchestrating increasingly sophisticated technology stacks. At the same time, extraordinarily capable people are accumulating enough technology, distribution and expertise to function increasingly like companies.

Those two trends are going to collide.

And when they do, I do not think the future needs to become some simplistic competition between employers and employees.

A much more interesting arrangement is possible.

Companies can become places filled with people learning how to build. Some will remain employees for decades, some will build new businesses inside the organization, some will create side businesses that never become their primary work, and some will eventually become founders, vendors, customers, or partners. Occasionally, one of them will build something enormous, and the company that helped them along the way should not see that as a failure, but as evidence that extraordinary people want to spend part of their journey there.

For professionals, the lesson is equally important: do not wait until your job disappears to discover whether you can create economic value without it. Do not wait for the layoff, the reorganization, the moment your skills become obsolete, or the day your company realizes one AI-enabled employee can do what once required three people. Build, learn, experiment, create, develop relationships, and find customers while you are still employed, so you develop the muscle of ownership before your life requires you to use it.

The corporation once possessed capabilities an individual could never reasonably reproduce. Those advantages have not disappeared.

But technology is steadily unbundling many of them and making pieces available to individuals one subscription, model, agent and platform at a time.

That changes what a company can be.

It changes what an employee can be.

And I think it fundamentally changes what ambition can look like.

The safest future may not belong to the person with the safest job.

It may belong to the person with the greatest number of ways to create value.

Quiet quitting asked how little of yourself you owe your employer.

Quiet building asks a much better question:

What are you going to build to become an owner?

Are You a Quiet Builder or CEO that Would Like to Work with John?

Click Here.

Research Appendix

Artificial Intelligence and Employment Exposure

International Monetary Fund — AI Will Transform the Global Economy
The IMF estimates that almost 40 percent of global employment is exposed to AI, rising to approximately 60 percent in advanced economies. Importantly, exposure can lead to either displacement or productivity-enhancing complementarity.
Read the IMF analysis. (IMF)

International Labour Organization — Generative AI and Jobs: A Refined Global Index of Occupational Exposure
The ILO’s 2025 task-level analysis concludes that approximately one in four jobs worldwide has some exposure to generative AI while emphasizing that job transformation is generally more likely than complete occupational replacement.
Read the ILO research. (International Labour Organization)

U.S. Bureau of Labor Statistics — Industry and Occupational Employment Projections
The BLS incorporates AI-related productivity effects into its employment projections, anticipating declining demand in some occupations while projecting growth in others as AI changes the composition of labor demand.
Read the BLS analysis. (Bureau of Labor Statistics)

AI and Productivity

Erik Brynjolfsson, Danielle Li and Lindsey Raymond — Generative AI at Work, NBER
This field study examined the deployment of a generative AI assistant among thousands of customer-support agents and found an average productivity improvement of roughly 14 percent, with disproportionately large benefits for less-experienced workers.
Read the NBER paper. (National Bureau of Economic Research)

Stanford Institute for Economic Policy Research — What Is Really Happening to Jobs?
Stanford’s 2026 review separates emerging evidence from stronger claims about widespread AI unemployment. It finds substantial evidence of productivity effects alongside early employment pressure in certain AI-exposed categories, particularly among younger workers.
Read the Stanford analysis. (SIEPR)

Stanford HAI — 2026 AI Index Report
Stanford’s annual AI Index compiles research on the economic effects, adoption and capabilities of artificial intelligence, including measured productivity improvements across multiple categories of knowledge work.
Explore the 2026 AI Index. (Stanford HAI)

AI Adoption and Changing Skills

OECD — Skills in the AI Age
The OECD reports that AI use among firms in countries with available data increased from roughly 7 percent in 2021 to approximately 20 percent in 2025, illustrating how quickly AI is moving from experimentation toward ordinary business infrastructure.
Read the OECD report. (OECD)

World Economic Forum — Future of Jobs Report 2025
Based on a survey of more than 1,000 major employers representing over 14 million workers, the report estimates that 39 percent of workers’ core skills will change by 2030. AI and big data skills are among the fastest-growing areas of demand, while analytical thinking, resilience, flexibility, leadership and creative thinking remain important.
Read the Future of Jobs Report. (World Economic Forum)

The One-Person Business Economy

U.S. Census Bureau — 2023 Nonemployer Statistics and Business Owner Characteristics
The Census Bureau reports approximately 30.4 million U.S. nonemployer businesses in 2023 generating nearly $1.8 trillion in receipts, demonstrating the enormous scale of economic activity already occurring through businesses without paid employees.
Read the Census Bureau release. (Census.gov)

U.S. Census Bureau — The Small Business Story
The Census Bureau’s 2026 review reports 30,427,808 nonemployer establishments in 2023, up from 29,811,495 in 2022, with total revenue approaching $1.8 trillion.
Read the Census Bureau analysis. (Census.gov)

Stop Paying For Advice. Start Investing In Operating Systems.

John Brewton — Wed, 12 Aug 2026 10:28:36 GMT

Most coaching ends the same way.

You leave with notes, a few good ideas, and the intention to do the work.

Three weeks later, the notes are cold. The ideas are still ideas. The thing that was supposed to change inside the business has not actually evolved.

You paid for advice. But what you needed was a system.

That realization has shaped a lot of how I think about coaching, consulting, and advisory work now.

The traditional model has an implementation problem.

The coach explains. The client understands. The call ends. Then responsibility for translating that understanding into something operational gets handed straight back to the person who was already too busy, too close to the problem, or too unsure of the solution to fix it alone.

There is nothing inherently wrong with advice. I have benefited enormously from good advice throughout my career.

But increasingly, I do not think advice should be where the work stops.

Especially now.

AI gives us the ability to take what used to live in a strategy document, a set of notes, or somebody’s head and turn it into a working system.

That changes what I think clients should expect from me.

And it is why I built The Operating Project.

Why I Built This

Before I became a coach, advisor, writer, and investor, I was an operator.

I spent years helping operate, optimize, and scale my family’s B2B industrial distribution company from $27.5 million to $75 million in revenue before we sold the business in August 2021.

This year alone, I have spent more than 400 hours working directly with more than 100 founders, entrepreneurs, operators, and professionals.

Those engagements have taught me an enormous amount.

They have also made one pattern impossible to ignore.

The people I work with rarely lack ideas.

They usually lack a reliable way to turn an important idea into something that works repeatedly inside the business.

Sometimes the problem is an offer.
Sometimes it is pricing.
Sometimes it is positioning.
Sometimes the founder knows they need to publish consistently but have no content system.
Sometimes there is simply one stubborn operating problem inside a larger company that everybody can see and nobody has been able to solve.

The answer is rarely another 60-minute conversation about the problem.

What people need is a way to move from:

problem → decision → design → build → operating system.

That is what The Operating Project is designed to do.

One Problem. One Outcome. One Working System.

The Operating Project is a four-session engagement in which we take one important problem in your business from unresolved to solved—and build the system that helps keep it solved.

You bring the problem.

I bring the structure, the operating frameworks, the worksheets, the thinking, and the build.

Before we begin, we define what finished looks like.

Then we work toward it.

This is not open-ended coaching.

It is not a standing consulting retainer.

It is not four calls where I give you ideas and send you away with homework.

It is one operating problem, scoped to one outcome, worked on until something useful exists.

Bring Me the Problem You Keep Circling

You do not need to arrive knowing exactly what the solution is.

You need to know which problem matters.

Here are four kinds of projects I expect to work on most often.

1. Your offer or business model is not working

Maybe the pricing is wrong.
Maybe the offer is confusing.
Maybe customers do not understand what they are buying.
Maybe you have several products but no coherent architecture connecting them.

We can work through the economics, offer design, pricing, packaging, positioning, and go-to-market logic until the model makes sense.

The finished result might include an offer map, a pricing model, a product architecture, or a decision system for evaluating future offers.

2. Your brand feels different every week

Your website says one thing.
Your LinkedIn says another.
Your pitch deck says something else.
Every new piece of content feels like another attempt to rediscover what the company stands for.

We can build the positioning, messaging, brand architecture, and identity system that creates consistency across the business.

The goal is not simply a prettier brand.

It is a system that helps you make better brand decisions repeatedly.

3. Content keeps becoming a weekly emergency

You know you should publish.
You may even know what you should publish.
But every week still begins with the same question:
- What am I going to make this week?

We can build the strategy, recurring formats, editorial calendar, research process, production workflow, and AI infrastructure that turns content from an act of recurring inspiration into an operating system.

4. Something inside the company is simply broken

Maybe it is a workflow.
Maybe it is a decision process.
Maybe it is a sales problem, an operating bottleneck, a reporting problem, or a process everybody has learned to work around.

You do not need to rebuild the company.

You need to isolate the problem, redesign how the work should happen, and create a practical system around the fix.

You Do Not Leave With Homework. You Leave With Infrastructure.

This is the part of The Operating Project that matters most to me.

Most advice dies because it stays a document.

A document still depends on you remembering it.

A framework still depends on you applying it.

A strategy still depends on somebody translating it into work every time the work needs to happen.

A machine is different.

If we redesign your content strategy, I do not want to stop at a PDF describing it.
- I want to help you build the Claude Project and ChatGPT Project, skills, instructions, templates, research process, workflows, and supporting assets that enable the strategy to actually produce content.
If we redesign your offer, I do not want you to leave with a single pricing worksheet that you use once.
- I want to help you build a decision system you can use when creating your next offer.
If we solve an operating problem, I do not simply want to describe the new process.
- I want us to install the logic behind that process somewhere your business can use it again.

The objective is not for you to remember everything I told you.

The objective is to encode what we built.

That is what AI makes possible now.

And I think it fundamentally changes what good advisory work can look like.

What You Might Leave With

Every project will be different because every problem is different.

Depending on the engagement, you might leave with:

A redesigned offer and pricing architecture;
A complete positioning and messaging system;
A 90-day content strategy and editorial calendar;
A Claude or ChatGPT Project configured around your business;
Reusable AI skills, prompts, instructions, and workflows;
A business-model or financial decision framework;
An operating process your team can immediately begin using;
Research systems and templates for recurring work;
A diagnostic framework that helps you make the same class of decision in the future;
The actual documents, worksheets, models, and tools we create together.

The point is that there is something at the end.

Something you can open.

Something you can use.

Something that exists inside the business after the engagement is over.

The Four Sessions

The Operating Project runs through four moves:

1. Define

We diagnose the problem in plain language.
We determine what is actually causing it, decide what needs to change, and define the artifact or operating outcome we want to hold at the end.
We also establish the scope.
What are we solving?
What are we not solving?
What does done mean?

2. Design

Then we build the solution.

Depending on the project, that might mean pricing models, offer architecture, positioning frameworks, operating processes, content calendars, decision trees, or another custom framework.

This is not something I disappear and create in a black box.

We build it around the reality of your business.

3. Build

Then we turn the solution into a machine.

This is where we build the working system inside Claude, ChatGPT, or another appropriate part of your workflow.

We create the projects, instructions, skills, prompts, reusable assets, and processes required to make what we designed repeatable.

You own everything we build.

It lives in your account.

4. Install

Finally, we run the system on live work.

We test it.
We find where it breaks.
We refine it.
We make sure you know how to use it.

And we define what happens after the project closes.

Define. Design. Build. Install.

That is the Operating Project.

I Do Not Sell You Four Hours

The sessions are not timed.

Some may take 30 minutes.

Some will take longer.

We might work across more than four sessions.

The unit of measurement is not the clock.

It is progress.

We stop when we have completed the work required for that stage of the project.

That matters because I do not want you thinking about whether we have seven minutes left in the session.

I want us thinking about whether the problem moved.

A Plan Tells You What to Do. A Machine Does It for You.

There is a much bigger idea underneath this offer.

For most of my career, business advice had an unavoidable limitation.

Eventually the advisor had to leave.

The consultant delivered the deck.

The coach ended the call.

The operator went home.

The client was left with the process of turning all of that expertise back into action.

AI changes the economics of that handoff.

We can now begin encoding expertise into systems the client continues to own and use. This capability makes good thinking more durable.

That is the principle I want The Operating Project to embody.

You should not simply become better informed because we worked together.

Your business should become better equipped.

What People Are Saying

Who The Operating Project Is For

I think this engagement works best for founders, operators, creators, and business owners who have one meaningful problem worth solving now.

It is especially well suited to people who want to build rather than simply discuss.

You should be willing to make decisions, work through ambiguity, and use AI as part of the operating system we create.

You do not need to be an AI expert.

You do need to be willing to use the system once we have built it.

The Operating Project is probably not the right engagement if what you primarily want is general accountability, broad life coaching, indefinite access to an advisor, or someone to operate your company on your behalf.

This is finite by design.

One problem.
One project.
One outcome.

The Terms

The Operating Project is $999.

Your seat includes:

Four outcome-based working sessions.

All of the frameworks, worksheets, models, and operating assets we build together.

The Claude or ChatGPT system we create around the solution.

A full year of VIP access to my weekly Friday office hours.

Up to one year to schedule and complete your four project sessions.

There is no retainer.
There is no recurring consulting commitment.
There is no requirement that you start next week.

You can reserve the seat now and begin when the right problem is ready.

I would rather help you solve an important problem six months from now than manufacture an unimportant one because you bought today.

Bring Me One Problem Worth Solving

I am grateful for the more than 100 people who have trusted me with their businesses, careers, ideas, and ambitions this year.

Many of them have become friends and collaborators.

They have also helped me understand the kind of work I want to do more of.

I want to work on consequential problems with people who are serious about building.

I want the work to have a finish line.

And I want what we build together to keep producing value after I leave.

That is what The Operating Project is designed to do.

Bring me one problem worth solving.

We will define it, solve it, build the system around it, and put that system to work inside your business.

You will not leave with more homework.

You will leave with something that works.

— John

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After helping scale his family’s B2B industrial distribution company from $27.5 million to $75 million in revenue and selling the business in 2021, he has worked with business owners, founders, operators, and investors on company building, operating strategy, AI-enabled workflows, and growth. He is the founder of 6A East Partners, a research and advisory firm exploring the future of companies.

We Are All Becoming Companies.

John Brewton — Mon, 10 Aug 2026 11:38:06 GMT

Every company is racing toward the same destination. They are buying the same AI models, installing the same copilots, building the same agents and automating the same workflows. That will make companies more productive, but it will also make more of what they do easier for competitors to reproduce.

A study of 5,179 customer-support agents found that an AI assistant raised productivity most among novice and lower-skilled workers; other experiments with professional writing found faster, higher-quality output, while research into creative writing found that AI-assisted stories became better individually but more similar collectively. The floor rises, the range compresses, and competent work becomes more abundant.

At the same time, something is happening in the opposite direction. For more than a century, companies have manufactured humanity around an institutional container. Betty Crocker gave a flour company a woman customers could relate to; KFC turned its founder into an enduring character; corporate brands built voices and personalities because a legal entity has no childhood, scars, convictions or lived experience.

The company needed humanity because the actual human did not scale very well. Now the human scales. One founder can increasingly research, analyze, design, build, publish, distribute and operate with capabilities that once required teams.

Companies are becoming more machine-like while people are acquiring the capabilities of companies.

We are all becoming companies.

The Founder Becomes the Human Interface

A founder-led brand makes the company’s judgment, worldview and reason for existing more legible. Customers, employees, partners and investors are not only evaluating what a company sells; they are deciding whether they trust the people and ideas behind it.

Andreessen Horowitz has been building an explicit New Media strategy around a similar shift. Its recent New Media Playbook argues that individuals increasingly matter more than corporate brands, while its Lighthouse Playbook says founders should define their worldview before growth tactics and create preferential attachment: enough signal and trust that scarce customers, employees, partners and other resources are more likely to choose your company.

You can see extreme versions in Alex Karp, Palmer Luckey and Elon Musk. They are polarizing, but structurally useful examples: their names carry recognizable worldviews, and their companies are understood partly through the founder’s public decisions and personality. a16z has itself highlighted Musk and Karp as examples of technical thinkers newly advantaged by long-form and clip-based media, while giving founders like Palmer Luckey extensive space to communicate their ideas directly.

You do not need their fame or controversy to learn from the structure. If your name is going to become one of your company’s important interfaces with the world, define what that name means before you spend years amplifying it.

I Learned the Distribution Playbook First

I know this partly because I built my own audience in almost the opposite order. When I started creating content, I became obsessed with distribution: I committed to publishing every day for a year, learned LinkedIn deeply, invested in creator communities, and held more than 150 unpaid Zoom conversations with other creators.

It worked. I eventually documented the process in 0 to 55,000: The First 90 Days Playbook, and I still believe in much of it: show up, publish, build relationships, track what works, and get your reps.

One early breakthrough came from writing about Andy Grove’s High Output Management. Instead of summarizing the book, I wrote about mistakes I had made managing people myself, and the post took off. I came away with a formula I still believe in:

Credible source material + personal experience + external amplification = breakthrough.

But another lesson matters more now. For 12 to 15 months, I obsessed over follower growth and brand awareness while assuming the business model would eventually emerge. Viral posts sent large amounts of attention toward me before I had a sufficiently clear offer, owned audience or system for turning that attention into a business.

My conclusion at the time was: build the business while you build the audience. I still believe that, but today I would put another instruction before it: build the brand before you amplify the brand.

A lot of creator education starts downstream. It teaches hooks, carousels, social growth and distribution, all useful skills. But distribution is an accelerant; it cannot tell you what should be accelerated.

The creator playbook teaches us how to become known. The founder-led brand requires deciding what we want to become known for.

A Founder-Led Brand Is Not a Following

A founder-led brand is not your follower count, logo, headshot, or chosen family of fonts. It is not even your content; those are expressions and distribution layers targeted to your ICP. Your brand is the set of associations that forms when someone repeatedly encounters your name.

What do you believe?

What have you done?

What problem do you understand unusually well, what are you building because of it, and why should somebody trust you?

Eventually the question becomes more commercial: why should I attach myself to what you are building? The best founder-led brands do not simply create awareness; they make the right people more likely to move toward the company.

That is why I like a16z’s preferential attachment framing. Brand becomes more than communications.

Brand becomes a resource-allocation mechanism.

Seven Questions to Build Yours

I think every founder should answer seven questions before obsessing over distribution.

1. What is your journey?

Your résumé compresses your life into titles, companies and dates. Your founder brand should recover the experiences that explain how you became the person making today’s decisions: the terrible boss, the company you failed to build, the customer who changed your mind, or the decade you spent learning something that only recently became useful.

The journey is the moat. A model can acquire much of the information you learned, but it did not live the sequence that formed your judgment.

2. What do you believe?

This is your worldview. Not “integrity matters” or “AI is changing everything,” but the things you see differently enough to take a position on.

What does your industry get wrong? What advice has become outdated? What future are you trying to create? A memorable founder-led brand comes from conviction, not manufactured controversy.

3. What company does that worldview lead you to build?

The founder brand has to connect to an actual business. Otherwise, you risk becoming an interesting creator without becoming a stronger founder.

The bridge should be simple:

I believe X. Therefore I am building Y. For Z. To produce this outcome.

If you cannot complete that clearly, the brand and company are still too disconnected.

4. What should your name mean?

Force yourself to choose three associations, not seventeen. What do you want somebody to think after encountering your work 20 times?

Then choose what you do not want to be known for. Brands are shaped by exclusion as much as addition.

5. Who do you want to attach to you?

Not all attention is equal. Whose attention could actually change the trajectory of your company: customers, employees, investors, partners or collaborators?

For each group, ask: what must they believe about me before they choose to move toward me? That is a better question than simply asking how to get more followers.

6. What should you repeatedly give away?

Only now do we reach content. Teach, explain, document, show your work and make your thinking useful enough that people understand your worldview through what you give them.

AI should make this easier. Use it for research, organization, editing and derivative content, but keep the scarce part human: the ideas, judgment, story and final selection. As a16z’s Olivia Moore argues in Humans Are for Ideas, AI Is for Execution, the increasing abundance of execution raises the relative value of distinctive ideas, taste and creative direction.

7. How will the signal compound?

Now design distribution. Choose an owned surface, a primary discovery channel, a long-form format, and a cadence you can sustain.

Founder-led should not mean founder-only. a16z’s Lighthouse Playbook extends the idea to executives, engineers, customers and advisors whose own expertise can become additional human nodes of trust around the company.

The Workbook I Wish I Had When I Started

Download The Founder-Led Brand Builder

I still believe in publishing before you feel ready, but before telling somebody to publish every day, I would now spend a few hours answering more fundamental questions. Who are you? What shaped how you see the world? What are you building because of what you believe, and what should your name mean to the people who matter?

So I built the Founder-Led Brand Builder for exactly that. Upload it to ChatGPT or Claude and the model interviews you one question at a time, pushes past generic answers, looks for patterns in your experiences and beliefs, and turns the conversation into a Founder-Led Brand Blueprint: your origin story, worldview, positioning, desired associations, target audiences, signature stories, content territories and a 90-day distribution plan.

The workbook cannot manufacture your brand. The raw material is your life; it simply helps you see that material more clearly before you amplify it.

If building this kind of business and capability is especially appealing to you, and you want to learn everything you can in the year ahead, become an Operating Founder and join my 52-week-per-year Friday live Q&A sessions with The Operating Founder Community. We are a +100-strong group of Operators and Builders choosing a new vision for our futures.

Become an Operating Founder

The Human Interface to the Company

Companies are going to become extraordinarily capable machines. We should automate workflows, deploy agents, standardize what should be standardized and use AI everywhere it creates leverage.

But something else should happen at the same time: the humans should become more visible. The judgment, taste, history, conviction, and people willing to attach their names to decisions should become easier to see.

AI is giving more of us the operational capabilities of companies. That means more of us need to start thinking like companies about what our names stand for. The creator playbook taught us how to become known; the founder-led brand starts with a different question:

What do you want to become known for?

Define that first, then build the business and the audience around it. Because when everyone has access to increasingly similar machines, who is operating the machine matters more.

Companies are becoming machines.

We are all becoming companies.

- j -

Want to Build This With Me?

If the hard part is not understanding this framework but actually turning your brand, offer or content strategy into a working system, that is what I built The Operating Project to do.

Across four one-to-one working sessions, we take one important problem in your business from diagnosis to a finished result. We build the strategy, create the artifact, and then install the ChatGPT or Claude system that keeps doing the work after our sessions are over.

If your problem is your founder-led brand, we can turn your worldview, positioning, offers and content into a system you actually own.

Learn Everything About The Operating Project and grab your seat →

Start Working With John Today

Frequently Asked Questions

Is a founder-led brand the same thing as a personal brand?

Not quite. A personal brand can exist independently of a business; a founder-led brand is deliberately connected to the company you are building. Your worldview, reputation and identity create trust, but that trust should eventually make the company itself stronger.

Do I need to become an influencer?

No. The objective is not maximum visibility; it is useful visibility among the people whose decisions matter to your business. A founder with 5,000 highly relevant followers can have a far more economically valuable brand than someone with 500,000 people who have little connection to what they sell.

Do I need to become controversial?

No. Musk, Karp and Luckey are useful because their brands make the underlying structure obvious, not because founders should imitate their personalities. Distinctiveness requires conviction, not performative controversy.

Should AI write my content?

AI should absolutely help you research, organize, edit, repurpose and execute. I would be much more cautious about outsourcing your worldview, consequential opinions, personal experiences and final judgment; those are precisely the parts becoming more valuable as execution gets cheaper.

Does founder-led mean the company can never outgrow the founder?

No. In fact, the stronger model may eventually become a network of trusted humans around the company: founders, executives, technical leaders, employees, customers and advisors. The founder begins the signal; the organization eventually creates many credible sources of it.

What should I do first if I have no founder brand today?

Do not begin by choosing platforms or designing a content calendar. Start by answering three questions: What do I believe? What company does that belief cause me to build? What three things do I want my name to mean? Then use the Founder-Led Brand Builder to turn those answers into the rest of the system.

Appendix: Research and Sources

A. AI, Performance and the Compression of Competence

Erik Brynjolfsson, Danielle Li and Lindsey Raymond — “Generative AI at Work.”
NBER working paper studying 5,179 customer-support agents. AI assistance increased productivity by 14% on average, with substantially larger benefits for novice and lower-skilled workers. Read the research at NBER. (National Bureau of Economic Research)

Shakked Noy and Whitney Zhang — “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.”
Published in Science. Participants using ChatGPT completed professional writing tasks substantially faster while average output quality increased. Read the paper in Science. (Science)

Anil Doshi and Oliver Hauser — “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content.”
Published in Science Advances. AI-assisted stories were rated more creative and better written, while the collective set of outputs became more similar. Read the paper. (Science)

Andreessen Horowitz — “From Copilots to Agents: Rebuilding the Company Around AI.”
A current company-building example of AI moving beyond isolated assistance toward systems that perform increasingly large shares of business workflows. Listen/read at a16z. (Andreessen Horowitz)

B. Companies Manufacturing Human Identity

General Mills — “How Betty Crocker Got Its Start.”
The corporate history of the 1921 Gold Medal Flour promotion that led Washburn-Crosby to create Betty Crocker as a human persona for answering consumer questions. Read the history. (General Mills)

General Mills — “The Many Portraits of Betty Crocker.”
Additional history of how the fictional persona became a recognizable human face and trusted voice for the company. Read the history. (General Mills)

C. Andreessen Horowitz on New Media and Founder-Led Brands

Erik Torenberg, Ben Horowitz and Marc Andreessen — “a16z’s New Media Playbook.”
The February 2026 discussion in which a16z explicitly argues that individuals increasingly matter more than corporate brands and lays out its evolving approach to direct media. Listen at a16z. (Andreessen Horowitz)

David Booth — “The Lighthouse Playbook.”
The most directly relevant a16z source for this article. It argues that founders should define their worldview before growth tactics, develops the idea of preferential attachment, and describes how employees, customers and advisors can become trusted “lighthouses” around a company. Read the playbook. (Andreessen Horowitz)

Erik Torenberg, Alex Danco, Elena Burger, Henry Williams, Tom Hollands, Brent Liang and Gaby Goldberg — “New Media, One Year In.”
a16z’s June 2026 retrospective describes New Media as effectively “go-direct as a service” for startups and treats creative capabilities, owned distribution and founder communication as company-building infrastructure. Read the retrospective. (Andreessen Horowitz)

Olivia Moore — “Humans Are for Ideas, AI Is for Execution.”
A February 2026 argument that AI increasingly commoditizes execution while distinctive ideas, taste, identity and creative direction remain disproportionately human sources of value. Read the essay. (Andreessen Horowitz)

Alex Danco — “Why Nerds Are More Clippable.”
Explores how the shift from short, tightly controlled media appearances toward long-form source material and short-form clips creates new advantages for technical thinkers with distinctive personalities. The essay specifically cites Elon Musk and Alex Karp as examples. Read the essay. (Andreessen Horowitz)

Chris Dixon and Palmer Luckey — “Palmer Luckey on Hardware, Building, and the Next Frontiers of Innovation.”
A long-form founder conversation recorded at a16z’s Founders Summit and a useful example of the direct founder-media model discussed in the article. Watch/listen at a16z. (Andreessen Horowitz)

D. My Own Audience-Building Lessons

John Brewton — “0 to 55,000: The First 90 Days Playbook — What I’d Do Differently Knowing What I Know Now.”
My earlier breakdown of daily publishing, platform mastery, creator relationships, tracking and distribution—and, more importantly for this article, my realization that I spent too long optimizing follower growth before building the offers and business infrastructure beneath the audience. Read the full Operating article.

The same piece documents the early Andy Grove post that led me to the “credible source material + personal vulnerability + external amplification” formula and the 150+ unpaid creator conversations that became an important part of my education.

E. Work With Me

John Brewton — The Operating Project.
Four working sessions focused on one defined business problem and one finished outcome. The engagement moves from diagnosis to design, then builds a working system inside ChatGPT or Claude and installs it on a live piece of the client’s business. Learn about The Operating Project. (Operating by John Brewton)

A Conversation About The Importance of Building Real Human Relationships & Friendships in the Creator Economy w/ Special Guest and Great Friend and Entrepreneur Cory Blumenfeld

John Brewton — Fri, 07 Aug 2026 15:31:50 GMT

Available for iOS and Android

I Built Too Much. Here’s What I’m Cutting, and What I’m Scaling Instead

John Brewton — Fri, 07 Aug 2026 11:51:18 GMT

I wrote the full Do. Fail. Learn. Grow. Win. framework here.

This year, I ran a $99 offer: four coaching sessions for $99.

More than 100 people took me up on it in twelve days.

I ended up delivering more than 400 hours of coaching.

During those same months, this newsletter became a Substack bestseller. I launched two course cohorts. I opened a community on Skool. I started building my own software to help me run my business. And underneath all of it, I maintained a full-time commitment to my consulting work.

None of these endeavors failed.

That is what makes the next part important.

I simply built across too many surfaces:

Too many places where the business needed me to show up.
Too many products.
Too many platforms.
Too many workflows.
Too many pieces of software.
Too many things that were each perfectly reasonable on their own.

That is the dangerous thing about complexity.

You rarely wake up one morning and decide to make your business harder to run.

You add one good idea.

Then another.

Eventually, the work of operating all the good ideas becomes the bad idea.

So going into the next 12 months, I am concentrating this side of the business around three things:

Teach many people at once.
Build working tools that people can operate without me.
Publish useful work that compounds and people are willing to pay to receive.

Everything else has to justify why it deserves another piece of my time.

And when I stepped back from all of those decisions, I realized something else.

I had spent the entire year using one framework.

Do. Fail. Learn. Grow. Win.

The short version is:

Do: Put something real into the world.

Fail: Pay attention to where reality disagrees with your plan.

Learn: Figure out why.

Grow: Change the next attempt.

Win: Keep what works and use it to create the next opportunity.

Then go back to the start and Do More.

That is basically the story of my year.

And I think it explains much more about where this business is going than any five-year plan I could have written in January. It is also my sincere hope that my experience, the challenges I have faced, and the decisions I’ve made can help you in building your newsletter or solopreneur business.

Become an Operating Founder

1. Do

I did a lot this year.

Probably too much.

But I am increasingly convinced that the “too much” part was necessary to learn what I know now.

I launched the $99 coaching offer.

More than 100 people bought it in twelve days.

I spent more than 400 hours sitting directly with business owners, founders, creators, and operators.

I opened the Skool community.

I started the Friday Q&A calls.

I built the Compound Content Studio, where I taught people how to scale their content operations with Claude. The idea came from my now business partner, Wessal Khader, and man am I happy I said ‘Yes’ to her idea.

We launched the Oper(AI)te Summer of AI Fluency immediately afterward.

I started building internal software with Claude Code.

I experimented with AI agents.

I changed workflows.

I added tools.

I removed tools.

I tested different ways of teaching, consulting, coaching, building, and publishing.

There was no shortage of doing.

And that matters because almost everything I now understand about this business came after I put something in front of a customer.

That is very close to the argument Steve Blank made in his classic Harvard Business Review piece on the Lean Startup.

A new business model is not something you reason your way to perfectly in a conference room. It is a collection of assumptions.

So you put those assumptions into the world.

You watch what happens.

You change them.

Then you try again.

I could have spent six months designing the perfect coaching offer. I still would not have known what 100 actual clients were about to teach me.

I could have spent three months building the curriculum for Compound Content Studio. Instead, we designed it in three days, sold it for one week, and filled all 32 seats.

I could have spent months planning the perfect community. Instead, I built one and discovered that the part people really valued was the Friday call.

Reality is an excellent editor.

You just have to ship something before it can edit you.

There is also a much bigger economic change making this kind of experimentation possible for more people.

The U.S. Census Bureau counted more than 30 million American businesses with no paid employees in 2023. Together, those businesses generated nearly $1.8 trillion in revenue.

The creator economy has added another layer.

An IAB study led by Harvard Business School Professor Emeritus John Deighton estimates that the equivalent of more than 1.5 million Americans now work full-time as digital creators, nearly eight times the 2020 number.

And AI appears to be lowering the barrier to starting even further.

A 2026 study of more than 160,000 Product Hunt launches found a jump in entrepreneurial entry after ChatGPT arrived, disproportionately driven by solo founders.

The same study found an important limit:

Teams still dominated the highest-performing products.

That feels about right.

AI makes it much easier for one person to attempt things that once required a team.

It has not repealed the laws of time.

And time was the problem I eventually ran into.

I did not need more things I could do.

I needed fewer things I had to operate.

If you’re enjoying this article, you’ll probably love this article!

Quiet Quitting or Quiet Building? Our Changing Relationship With Professional Ambition

2. Fail

This is where I want to be careful with the word fail.

My community did not fail because nobody joined.

My coaching offer did not fail because nobody wanted it.

The Operating Project did not fail because nobody bought it.

In some ways, those are harder failures to notice.

Demand can hide a bad operating model.

You can have customers and still have the wrong business.

You can have engagement and still have the wrong product.

You can have revenue and still have terrible economics.

You can be growing and still be building something you do not actually want to operate.

That was the kind of failure I had this year.

The community worked. The container did not.

The Skool community is shutting down.

The Friday calls are staying.

I am moving those to Zoom.

Running the community cost me roughly three to four hours every week.

There was platform-specific content.
Engagement.
Moderation.
Administration.

And all the little things that come with maintaining another place where people expect you to show up.

When I looked at what members actually valued most, the answer was obvious.

The Friday calls.

Which, inconveniently, were the reason I built the community in the first place. Everything around them had become a second publication I had accidentally agreed to run.

Building a community is not a bad idea.

It just turned out not to be my good idea.

Bain has studied a much larger version of this problem for years.

In Killing Complexity Before Complexity Kills Growth, the firm describes complexity as a silent killer of growth.

Companies add products.

Processes.

Business units.

Layers.

Systems.

Each addition can make sense.

Eventually, more and more of the organization gets consumed by managing the organization itself.

I somehow recreated a tiny solopreneur version of the same problem.

I had a surface-area problem.

The lesson was not:Communities do not work.

The lesson was: The calls were the product. The community was the container.

So I am keeping the product.

And deleting the container.

Coaching worked. My calendar did not.

Coaching was probably the most enjoyable work I did this year.

More than 100 people.

More than 400 hours.

I learned more about how small companies actually operate than I would have learned from almost any dataset.

I am not walking away from coaching.

I am walking away from trying to build a large coaching practice.

One-to-one services scale almost perfectly with one thing I do not have more of:

My calendar.

That is an old professional-services problem.

Harvard Business Review has written about “putting products into services”: turning the repeatable parts of expertise into systems, technology, or products while keeping humans involved where judgment really matters.

That is increasingly what I want.

I do not want someone to need another hour with me every time they want to use something I know.

I want some of the knowledge to remain behind after the call ends.

That leads directly to the Operating Project.

We pick one problem in your business.

We develop the strategy.

Then we build the actual piece of technology you need to run it.

Usually that takes around four sessions, although some projects go longer.

The important difference is the output.

You do not leave with a plan.

You leave with a tool.

That distinction keeps getting more important to me.

A strategy document creates work.
A functioning system absorbs work.
A checklist helps someone remember what to do.
A tool can actually do some of it.

Career coach Julia Starr ran into a similar constraint after years of selling her time.

In a Business Insider account of how she used AI to extend her coaching methodology, she describes moving from coaching and prompts toward building an actual application around her framework.

That is where I think a lot of expertise businesses are going.

The expert does not disappear.

The expertise becomes more transferable.

3. Learn

This is the most important part of the framework.

Failure itself has almost no value.

People like to say:

“I learned from failure.”

Sometimes.

But plenty of people fail the same way for 15 years.

The value comes from identifying why something missed and changing what you do next.

That is Learn.

And by the end of this year, I had a few lessons I trust.

If building this kind of business and capability is especially appealing to you, and you want to learn everything you can in the year ahead, become an Operating Founder and join my 52-week-per-year Friday live Q&A sessions with The Operating Founder Community. We are a +200-strong group of Operators and Builders choosing a new vision for our futures.

Lesson 1: The thing customers value may not be the thing you built around it.

The Friday calls were the product.

Skool was the container.

That sounds incredibly obvious now.

It was not obvious when I was building it.

This is why testing behavior before building infrastructure matters.

I should have run four Friday Zoom calls first.

If people loved them, I could decide afterward whether they needed a community around them.

Instead, I built the container and then discovered which part people actually wanted.

Next time, I test the behavior first.

Lesson 2: Demand does not automatically mean scale.

I had plenty of coaching demand.

That was not the problem.

The problem was that every new client meant more of my calendar.

That is fine if your ambition is to build a great coaching practice.

Mine is not.

I want to teach.

I want to work directly with people.

I want to help solve real problems.

I just do not want every additional dollar of revenue to require another hour from me.

That is why the cohorts matter.

And it is why the newsletter matters even more now.

Lesson 3: I am better at one-to-many than one-to-one.

The first cohort was Compound Content Studio, which I built with a business partner I met on Substack. It was her idea. We taught people how to create social content with Claude.

We designed it in three days.

Sold it for one week.

Closed the cohort at 32 people.

It sold out.

Immediately afterward, we launched Oper(AI)te Summer of AI Fluency.

That program teaches people how to build teams of AI agents that perform real work inside their businesses:

A branding, visual, and voice team that creates proposals, decks, and social content
A research team that conducts market and company research at institutional quality
An operations team that absorbs administrative work
One custom agent team for each member, designed around the work their specific business needs done

The cohort model itself is not new.

Andreessen Horowitz wrote about the rise of cohort-based courses years ago: live instruction, deadlines, peers, accountability, and interaction without requiring the instructor to teach every student separately.

Creator businesses have arrived at similar models.

Former Big Tech program manager Jean Kang described to Business Insider how she moved from individual coaching into a six-week group cohort while building her independent education business.

Justin Welsh’s self-reported account of building a solo business past $10 million in cumulative revenue shows a business evolving across consulting, products, subscriptions, sponsorships, and community rather than relying forever on one-to-one work.

The point is not that cohorts are magic.

They are simply a much better match for how I want to work.

I want interaction.

I want questions.

I want to see people build.

I want to watch someone change what they are doing while we are working together.

I just do not need to explain the same thing 100 times in 100 separate Zoom windows.

Lesson 4: Publishing may be the highest-leverage teaching I do.

This one took me longer to see.

Operating already generates about $17,000 in annual recurring revenue from paid subscriptions.

I have tended to think about Substack in two ways:

As a publishing business.

And as the top of the funnel for everything else.

Someone reads an article.

Maybe they eventually join a cohort.

Hire me.

Buy an Operating Project.

Attend a workshop.

That will all continue.

But I think I have been underestimating the publication itself.

A newsletter is one-to-many teaching too.

Maybe the purest form of it.

I can learn something once.

Write it once.

Teach it to thousands of people.

And unlike a call, the work does not disappear when the hour ends.

It becomes part of an archive.

Someone can discover it six months later.

Or two years later.

A useful article can keep teaching after I have moved on to the next one.

That is exactly the kind of leverage I am trying to create everywhere else in the business.

So the newsletter cannot just promote the work.

The newsletter needs to become more of the work.

Lesson 5: Attempts compound.

This may be the biggest one.

The speed at which we designed Compound Content Studio was not luck.

It was the 400 coaching hours.

By then, I had heard the problems.

Again.

And again.

I knew what people struggled with.

I knew which explanations landed.

I knew where they got stuck.

I knew what they were willing to pay to learn.

That made the next thing faster.

This is where I think people sometimes misunderstand experimentation.

The failed attempt is not wasted if it makes the next attempt smarter.

Pieter Levels may be one of the clearest creator examples.

His public list of projects contains plenty of recognizable winners: Photo AI, Remote OK, Nomads.com, Interior AI.

It also contains a very large graveyard.

That is the important part.

He built a lot.

Most did not become huge businesses.

A few worked extremely well.

The failures were not an embarrassing detour around the process.

They were the process.

That is also why I like the Do. Fail. Learn. Grow. Win. framework.

It gives failure a job.

Failure creates information.

Learning turns the information into a change.

Growth proves that you actually made the change.

Then you get another chance to win.

Lesson 6: The work should leave something behind.

This is probably the most important lesson for how I think about client work now.

I want the work to end in something the other person can operate without me.

Not just advice.

Not just a deck.

Not just a call recording.

Not just “here is what you should do.”

Something that helps do the thing.

A workflow.

An agent team.

A research system.

A dashboard.

An application.

A tool.

That is where what I learned from coaching intersects with what is happening in AI.

Because for the first time, building the tool is becoming cheap and fast enough that it can be part of the engagement.

And that changes what a consultant can sell.

4. Grow

Growth is where you prove you learned something.

You change the next attempt.

That is what the next 12 months are about.

Not adding more.

Changing the shape of what is already working.

What I am closing

Skool goes away.

The Friday call stays.

Those three to four hours a week move toward making one thing better instead of keeping an entire platform alive.

What I am changing

Large-scale one-to-one coaching goes away.

The Operating Project stays.

It moves from $499 to $999.

The engagement includes more build time.

And the deliverable becomes much clearer:

The output is a working tool, not a plan.

What I am scaling: cohorts

The cohorts worked.

So the plan for the next 12 months is one to two new cohorts per quarter, plus workshops in between.

That becomes one of the centers of this side of the business.

But it is not the only one.

What I am scaling even harder: Operating

The other major bet is this newsletter.

Today, paid subscriptions to Operating generate about $17,000 in annual recurring revenue.

That is meaningful.

But I think it can become much more meaningful.

My goal over the next 12 months is to turn that into a $100,000-plus recurring-revenue business.

That is roughly a sixfold increase.

Which means I cannot get there by simply asking more people to pay for the same thing.

The product has to get better.

A lot better.

So I want to triple down on the thing I discovered I enjoy most this year:

Teaching:

My target is four to five articles every week.

Not four or five essays summarizing AI news.

There are already plenty of places to find AI news.

I want Operating to help people do something.

I want to show you what I am building.

How I built it.

Why I built it.

Where it failed.

What I changed.

What happened when a better model arrived.

What manual workflow became an agent.

What prompt became a Skill.

What piece of software I stopped buying because I built the part I actually needed.

What I learned from solving a real operating problem inside my consulting work.

And what you can take from all of it and use yourself.

I want Operating to become a running field guide for building a more capable business with AI.

There are three things I want to teach here.

1. Build with the models

AI is moving too quickly for this publication to become a static library of prompt tricks.

I want to teach against the frontier as it moves.

When Claude gets better at software development, I want to show you what suddenly becomes buildable.

When ChatGPT gets better at multi-step work, I want to test it against something real.

When agents become more reliable, I want to put them inside actual workflows and show you where they work and where they break.

When a model crosses a capability threshold, I want to go back to something that failed six months ago and try again.

The useful question is not:

What model launched this week?

It is:

What can we build now that we could not build before?

That is the question I want Operating to answer again and again.

2. Build a better business

The second source of material is my consulting work.

I get to sit inside companies and see how work actually happens.

The workaround someone created five years ago.

The spreadsheet everyone quietly depends on.

The process nobody has written down.

The piece of software everybody pays for but nobody enjoys using.

The meeting that creates six follow-up tasks nobody wants to do.

The handoff that only works because Susan remembers it every Thursday.

Increasingly, I can look at those problems and ask:

What would this look like if we rebuilt it for an AI-first company?

Some of those solutions will remain client-specific.

But many of the lessons will not.

I want to bring those lessons back here.

Not confidential client information.

The operating insight.

The pattern.

What failed.

What worked.

What somebody else can use.

That creates a loop I like a lot.

Consulting gives me real problems.

Building gives me real experiments.

Operating gives me a place to teach what I learn.

And the questions readers ask create the next problems worth solving.

3. Build something you own

The third subject may ultimately become the biggest.

I want to spend much more time studying how the best solopreneurs, newsletter writers, creators, and small internet businesses in the world actually operate.

Not just how they create content.

How they build businesses.

What they sell.

How they price.

Where recurring revenue comes from.

How they turn expertise into products.

How they move beyond one-to-one services.

How they build newsletters people pay to read.

How they create distribution they own.

How they use AI.

How small their teams are.

What they automate.

What they refuse to do.

What their technology stacks look like.

Where the margins come from.

And what we can learn from them.

Because I think there is a much bigger story happening underneath all of this.

For most of modern business history, if you were highly ambitious, one of the obvious paths toward creating financial security for yourself and your family was to build a career inside somebody else’s organization.

It looked like this:

Get promoted.
Manage more people.
Earn more.
Accumulate equity.

There is nothing wrong with that route.

It remains an excellent one.

But technology has opened another path.

A person with expertise, an audience, distribution, AI, and the ability to build products now has access to operating leverage that not long ago would have required employees, developers, designers, researchers, and meaningful capital.

That does not make building a one-person business easy.

It makes a different kind of business possible.

I want Operating to study that possibility seriously.

Who is actually making it work?
How are they doing it?
Which parts are real and which parts are internet mythology?
What do the best newsletter businesses understand about recurring revenue?
What do the best creators understand about distribution?
What do the best solopreneurs understand about focus?
And how do you turn those lessons into something exceptional that you own?

Something that creates income.

Something that gives you options.

Something that can create real financial security for you and your family.

That is a subject I can imagine exploring here for years.

The newsletter itself becomes part of the experiment

Appropriately, I am going to use my own newsletter to test the thesis.

Today:

About $17,000 in annual recurring subscription revenue.

The next milestone:

$100,000+.

That is not a small increase.

It means I need to make the paid product dramatically more valuable.

More useful articles.
More original research.
More builds.
More templates.
More frameworks.
More case studies.
More things readers can actually take and use.

But there is an important constraint.

I do not want to reach $100,000 by creating another giant surface-area problem.

I do not want five newsletters.

Seven communities.

Three membership tiers.

A Discord server.

A course vault with 400 videos nobody finishes.

The experiment is whether I can grow the revenue and value of one publication by making the publication itself dramatically better.

That is very consistent with everything this year taught me.

Do not automatically add another surface because you want more growth.

Make the best surface more valuable.

That is what I want to do with Operating.

The three engines

The shape of the business is becoming much clearer.

There are three engines.

1. Paid publishing

Operating.

Teach at scale.

Build an archive.

Own the relationship with the audience.

Create recurring revenue.

Turn the things I am already learning into something thousands of people can use.

2. One-to-many education

Cohorts.

Workshops.

Friday calls.

Teach people live.

Help them build.

Answer questions.

Watch where they get stuck.

Then feed those lessons back into the publication and the next cohort.

3. Tool-building

Operating Projects.

Consulting builds.

Internal applications.

Agent teams.

Take what I know and turn more of it into something another person can actually operate.

These three things feed one another.

The consulting work gives me problems worth solving.

The builds give me things worth teaching.

The cohorts tell me where people struggle.

The newsletter turns those lessons into reusable intellectual property.

The audience creates more questions.

Those questions create new experiments.

And the experiments give me more to write about.

That is the flywheel I want.

Not more surfaces.

More value moving through fewer surfaces.

The stack

The technology underneath all of this is changing too.

Here is what actually runs the business now.

Claude is the operating layer.

Writing, research, analysis, and most of the agent teams start there.

Claude Skills are how Claude learns the business.

Brand systems.

Voice profiles.

Meeting workflows.

Inbox workflows.

Research standards.

Each one encodes a way of working so I do not have to teach the model the business from zero every time.

Anthropic’s explanation of Agent Skills describes them as reusable instructions, resources, and code that give an agent specialized ways of performing recurring work.

That is exactly how I think about them.

I do not want to prompt better forever.

I want the standard built into the system.

Claude Cowork is where more of that work runs against my files and connected systems.

Anthropic’s Claude for Small Business points directly toward this model: agents, skills, connected information, and workflows built around functions such as operations, sales, marketing, finance, and customer service.

Claude Code is where I build.

Internal applications.

Workflows.

Small pieces of software.

Tools for my own company.

Tools inside client engagements.

ChatGPT Work is the second operating environment.

I want two frontier AI systems.

They have different strengths.

Different weaknesses.

Different failure modes.

Anything important gets a second opinion.

OpenAI’s current direction with ChatGPT Work points in the same direction as Claude: persistent projects, reference files, reusable templates, tools, and multi-step work rather than a blank chat box that starts from zero every time.

Perplexity Comet and Computer handles advanced web research.

Perplexity describes Comet as an AI-powered browser capable of researching, navigating, summarizing, and working across the web.

Zoom runs every call.

More interestingly, it is starting to become part of the connective tissue between systems.

Zoom’s MCP integration with Claude can expose meeting intelligence to Claude Cowork and Claude Code.

That matters to me much more than having another AI meeting summary.

A meeting should not disappear inside the meeting app.

It should become operating context.

Plaud captures meeting transcripts.

Attio is the CRM.

And increasingly, Attio is the spine of the operation.

It holds client context.

Transcripts.

Progress.

Relationships.

And through Attio’s MCP support, AI systems such as Claude and ChatGPT can work directly against that CRM information.

That is the direction I want the entire stack moving.

Fewer isolated apps.

More shared context.

Calendly handles scheduling and intake.

Superhuman on Gmail runs email, with Claude handling much of the triage and drafting against a written standard.

Google Drive holds documents and course material.

Notion holds documentation.

The tables are on their way out.

Substack is now one of the core products of the business.

It is the publishing platform.

The archive.

The distribution engine.

And increasingly, a recurring-revenue business of its own.

Today, paid subscriptions generate about $17,000 in annual recurring revenue.

The goal for the next 12 months is $100,000-plus.

That means the newsletter is more than top-of-funnel.

It is one of the things I am building.

Why I think this stack gets much more powerful over the next 12 months

This is where I am making a bet.

I think the stack gets substantially more capable without becoming substantially larger.

There are four reasons.

1. AI models are crossing usefulness thresholds very quickly

My experience this year has been strange.

There were things I tried to build in March that did not work.

I tried them again in June.

They worked.

I was not a materially better software developer in June.

The models were better.

That sounds anecdotal until you look at the broader trend.

The Stanford AI Index has documented enormous gains on software-engineering benchmarks, including rapid improvements on increasingly realistic coding tasks.

METR has tried to measure the same phenomenon differently.

Instead of asking only whether an AI gets a benchmark question right, it studies the length and complexity of software tasks frontier AI systems can complete at a given reliability level.

Its research on AI task-completion horizons found that historically, the length of software tasks frontier systems can reliably perform has been increasing very quickly.

METR is also careful about what that means.

Its researchers explain in their limitations discussion that a model’s measured task horizon does not mean you can simply replace a person doing an equivalent number of hours of real work.

Real work is messier.

People collaborate.

Requirements change.

Information is missing.

Judgment matters.

But the direction still matters.

If a task sits just outside the capability line today, I increasingly think it is worth trying again later.

“AI cannot do this” has a much shorter shelf life than it did two years ago.

2. Yesterday’s frontier capability keeps getting cheaper

Better models are only half of the story.

The other half is cost.

The 2025 Stanford AI Index documented a dramatic collapse in inference costs for a fixed level of capability.

That may be even more important for businesses than the benchmark race.

The interesting question is not:

Can the most expensive model in the world do this once?

The useful question is:

Can a sufficiently capable model do this cheaply enough that I can put it inside a workflow and run it 500 times?

That is the threshold that matters.

It is one thing to have AI write one email.

It is another thing to have a reliable system process every meeting, every transcript, every follow-up, every piece of research, and every client update.

As capability gets cheaper, more of those workflows become economically sensible.

3. The model is becoming the layer that operates the software

The first version of generative AI gave us a text box.

Ask a question.

Get an answer.

Copy it somewhere else.

That is not where the products are going.

Claude now has Skills, Cowork, Code, connectors, and MCP.

ChatGPT Work combines models with files, reusable systems, connected tools, and multi-step execution.

Zoom can expose meeting intelligence to an agent.

Attio can expose CRM information to the same agent.

The model is slowly becoming less like another application in the software stack and more like the layer that operates the stack.

That is a much bigger change than better chat.

Consider one client meeting:

The meeting happens in Zoom.
Plaud records the transcript.
The transcript goes into Attio.
An agent can compare the conversation with prior meetings.
Extract commitments.
Update the CRM.
Generate the follow-up.
Identify the next action.
Prepare the next meeting.
And carry what happened into the next piece of work.

Those used to be separate activities spread across separate applications.

Increasingly, they can become one workflow.

This is why protocols like MCP matter.

The protocol itself is not particularly interesting to most business owners.

Shared context is interesting.

The less often I have to manually move information from one place to another, the more useful the entire stack becomes.

4. The cost of creating narrow software is collapsing

This is probably the biggest bet I am making.

For most of my career, if I wanted custom software, I had two options.

Buy something close enough.

Or hire someone to build it.

There is now a third option.

Describe it.

Build it with an AI coding system.

Test it.

Change it.

Keep what works.

The Financial Times described vibe coding as a new kind of DIY because natural-language development tools are making it possible for nontraditional developers to create applications that would never have justified the cost of a conventional software project.

Pieter Levels argues that this produces another consequence.

When everyone can build apps, distribution becomes the scarce thing.

That is an important inversion.

Software creation gets cheaper.

Understanding the customer becomes more valuable.

Having an audience becomes more valuable.

Owning distribution becomes more valuable.

Knowing the workflow becomes more valuable.

Knowing what should be built becomes more valuable.

Those are all good developments for somebody who just spent more than 400 hours talking directly to customers—and who publishes to an audience every week.

This is another reason the newsletter matters.

If building gets cheaper while distribution becomes more valuable, then owning the relationship with the audience becomes an increasingly important asset.

What I am replacing

The larger technology project underneath the stack is subtraction.

Gamma is gone.

Canva is gone.

The Notion tables are going.

In each case, I have replaced or am replacing parts of the software with something I built using Claude Code, sized around the way my business actually works instead of around the median customer of a software company.

This is the part I did not see coming.

The models became good enough to build the tool.

That changed what I needed to buy.

And that is beginning to change how I think about software categories altogether.

A surprising amount of software is an interface wrapped around a workflow.

For years, the hard part was turning a description of that workflow into functioning software.

That part is getting cheaper.

But I want to be precise here.

I do not think SaaS is going away.

And I definitely do not think you should cancel every subscription and vibe-code your accounting system this weekend.

The Financial Times has made a useful counterargument to the idea that AI coding simply kills SaaS.

Mature software companies provide things that homemade applications do not automatically give you:

Reliability.
Security.
Collaboration.
Maintenance.
Support.
Integrations.
Compliance.
Years of accumulated domain knowledge.

There are also very real risks in AI-built software.

WIRED reported on research into thousands of publicly accessible vibe-coded applications that found serious authentication and data-exposure problems.

MIT researchers studying the roadblocks to autonomous software engineering make a similar point from the engineering side.

Writing code is only one part of software engineering.

There is also:

Architecture.
Testing.
Security.
Maintenance.
Requirements.
Debugging.
Reliability.
System design.

So my thesis is narrower than:

Build everything.

It is this:

For a small, internal, well-bounded workflow, the default question is changing from “Which software category should I buy?” to “Could I build the 20% of this software I actually use?”

That is a much more interesting question.

I am not rebuilding payroll.

I am not vibe-coding authentication for sensitive customer data.

I am not rebuilding accounting controls because QuickBooks annoys me.

I am looking around the edges.

The internal dashboard.

The research organizer.

The transcript system.

The intake processor.

The content workflow.

The little application that turns 17 steps into three buttons.

The software where the output is easy to check.

The software where failure is annoying rather than catastrophic.

The software that exists because my business works a particular way.

I think that category is about to get very interesting.

The productivity research points in the same direction

There is another reason I think these systems will absorb more operating work.

We already have strong evidence that generative AI can improve performance on bounded knowledge tasks.

In a controlled experiment published in Science, people using generative AI on professional writing tasks completed the work substantially faster while independent evaluators also rated the output higher.

A separate NBER study of customer-support workers found meaningful average productivity gains, with particularly large improvements among less-experienced workers.

Those studies do not prove that AI can run a company by itself.

That is not my argument.

They point toward something much more practical.

There are pieces of knowledge work where the cost of producing a useful first draft, analysis, recommendation, or piece of code has fallen dramatically.

A one-person company has a lot of those pieces.

And that is why this technology may matter disproportionately to very small businesses.

A large company can hire another analyst.

A solopreneur cannot.

But a solopreneur can increasingly acquire some of the capability of another analyst without acquiring another payroll line.

There are already field reports of people trying to push that much further.

One founder told Business Insider that AI agents handle work ranging from proposals and scheduling to invoicing.

Another described building a collection of specialized AI agents across legal, finance, HR, and operations.

Those are individual accounts.

Not controlled studies.

But that is why they are interesting.

They are field reports from people trying to figure out what the operating model looks like before anyone has written the textbook.

That is the experiment I am running too.

How small can the stack get while the capability of the business goes up?

5. Win

The wins are easier to see now.

This newsletter became a Substack bestseller.

More than 100 people trusted me enough to work with me.

I spent more than 400 hours learning directly from them.

Compound Content Studio sold out.

Oper8 showed us another way to teach AI fluency in a group.

The Operating Project became a much better product because I finally understood where the real value was.

Operating became a real recurring-revenue business.

I built software I could not have imagined building myself two years ago.

And maybe most importantly, I now have a much clearer picture of the business I actually want to operate.

But this is where the final part of the Do. Fail. Learn. Grow. Win. framework matters.

Win is not the end.

Win sends you back to Do.

I do not think I “figured out” my business this year.

That would be a dangerous conclusion.

I think I earned a better set of experiments for next year.

That is different.

And it is enough.

The next loop

The next 12 months are another experiment.

Can I teach more people without adding more calendar?
Can I turn roughly $17,000 of recurring newsletter revenue into $100,000-plus by making one publication dramatically more useful instead of creating five more products?
Can I publish four to five genuinely useful pieces every week without turning the publication itself into another surface-area problem?
Can the Operating Project turn more of my expertise into tools clients keep using without me?
Can I reduce the number of software products I buy while increasing what the business can actually do?
Can AI agents absorb more administrative work without creating an even bigger mess underneath them?
Can shared context remove more of the handoffs between systems?
Can I learn from the best solopreneurs and creator businesses in the world and turn those lessons into something useful for people building businesses of their own?
Can I operate fewer surfaces and produce more value?
Can I make this business simpler while making it more capable?

I do not know yet.

That is the point.

Do.

Fail.

Learn.

Grow.

Win.

Then go do it again.

- j -

If building this kind of business and capability is especially appealing to you, and you want to learn everything you can in the year ahead, become an Operating Founder and join my 52-week-per-year Friday live Q&A sessions with The Operating Founder Community. We are a +200-strong group of Operators and Builders choosing a new vision for our futures.

Become an Operating Founder

What I would do differently

What to do this week

Research Appendix

I have grouped the research and practitioner examples by topic so this can also serve as a reading list for anyone thinking through the same questions.

1. Do. Fail. Learn. Grow. Win.

Operating by John Brewton

Do. Fail. Learn. Grow. Win.: My 5-Step Framework

The framework behind this article.

The core idea is that failure only becomes useful when you diagnose it, change your next attempt, and convert the lesson into growth.

Why read it: It is the operating model I use for experiments, mistakes, decisions, and iteration.

2. The rise of tiny businesses, creators, and solo entrepreneurship

U.S. Census Bureau

Census Bureau Data Tell the Small Business Story

The Census data provides useful scale for the one-person-company conversation. Tens of millions of American businesses operate without paid employees and collectively produce enormous economic output.

Why read it: Solopreneurship is not an internet niche. It is a substantial part of the economy.

IAB / John Deighton

Measuring the Digital Economy 2025

Research into the scale and growth of digital creators in the United States.

Why read it: Useful context for understanding independent creators as an increasingly important business and labor category.

Kim, Kang & Song

Generative AI Fuels Solo Entrepreneurship, but Teams Still Lead at the Top

Research using more than 160,000 Product Hunt launches to examine changes in entrepreneurial entry after the arrival of generative AI.

Why read it: It is one of the most directly relevant academic attempts to measure whether AI is lowering barriers to entrepreneurship.

Important caveat: The researchers also find that teams continue to dominate the highest-performing products.

3. Complexity, focus, and surface area

Bain

Killing Complexity Before Complexity Kills Growth

Bain’s argument that successful organizations often accumulate products, processes, systems, and layers until managing the complexity starts eating the benefits of growth.

Why read it: This is the large-company version of the surface-area problem.

Bain

The Founder’s Mentality

Research and thinking around how companies lose speed and customer focus as growth introduces complexity.

Why read it: Particularly useful if your company is getting bigger while somehow becoming harder to operate.

4. Experimentation and iteration

Steve Blank / Harvard Business Review

Why the Lean Start-Up Changes Everything

The classic argument for treating a new business model as a set of hypotheses to test against customers rather than a plan to perfect in advance.

Why read it: It provides a research-backed business explanation for the “Do” portion of Do. Fail. Learn. Grow. Win.

5. Turning expertise into scalable products

Harvard Business Review

Putting Products into Services

An examination of how professional-services businesses can standardize, automate, and productize repeatable pieces of expert work.

Why read it: This is one of the closest established business concepts to what I am trying to do with the Operating Project.

MIT Sloan Management Review

How to Turn Professional Services Into Products

A more recent treatment of productizing expert services.

Why read it: Especially relevant for consultants, agencies, coaches, accountants, and other businesses whose core product has traditionally been expert time.

6. Cohorts and one-to-many education

Andreessen Horowitz

Investing in Maven

A useful early explanation of cohort-based courses and the value of combining live instruction, accountability, peers, and deadlines.

Why read it: Helpful for anyone choosing between prerecorded courses, cohorts, communities, and individual teaching.

Jean Kang / Business Insider

I Quit My Big Tech Job and Made $100,000 in Revenue Within Five Months

Kang describes moving from individual coaching toward a six-week cohort as part of building an independent education business.

Why read it: A practitioner example of changing the economics of teaching without removing live interaction.

7. Creators and solopreneurs worth studying

These examples are field reports, not academic proof. I find them useful because they show people actively testing these operating models.

Pieter Levels

List of All My Projects Ever

A public record of Levels’ many launches, including both major successes and many projects that went nowhere.

Why read it: The sheer number of attempts may be more instructive than any individual success.

Also useful:

Everyone Can Now Build Apps With AI, So Distribution Is the Real Challenge

Why read it: A useful argument for why cheaper software creation makes audience, customer insight, and distribution more valuable.

Justin Welsh

My Complete $10M Journey

Welsh’s self-reported retrospective on building a solo business across consulting, digital products, subscriptions, sponsorships, and community.

Why read it: A good example of a creator business evolving through repeated experiments rather than emerging with one perfect model.

Julia Starr / Business Insider

I Built an AI Tool to Mimic My Work

A coach’s account of turning part of her expertise into an AI-enabled application.

Why read it: Probably the practitioner example closest to the Operating Project thesis.

Justin Parnell / Business Insider

AI Agents Handle Everything From Invoices to Proposals

A firsthand account of putting AI agents into several operating functions of a small business.

Why read it: Useful for seeing what an AI operating layer looks like when someone actually attempts one.

Aaron Sneed / Business Insider

A Solo Founder Runs His Company With Specialized AI Agents

A founder’s account of building specialized AI roles across finance, HR, legal, and operations.

Why read it: The interesting idea is organizational. Instead of thinking only in prompts, start thinking in roles, standards, responsibilities, and handoffs.

8. What AI capability is actually doing

Stanford HAI

The 2026 AI Index Report

Stanford’s annual AI Index is one of the best broad sources for tracking model capabilities, adoption, costs, investment, research, and benchmarks.

Why read it: If you want one source for separating real AI progress from daily hype, this is a strong place to start.

METR

Measuring AI Ability to Complete Long Software Tasks

METR tries to measure AI progress by looking at the length and complexity of tasks agents can successfully complete.

Stanford HAI

2025 AI Index: Research and Development

Research into falling inference costs and improvements in model efficiency.

Why read it: Capability getting cheaper is just as important to business adoption as capability getting better.

9. The emerging AI operating layer

Anthropic

Equipping Agents for the Real World With Agent Skills

Anthropic’s explanation of reusable skills, instructions, resources, and code.

Why read it: This is one of the most important concepts in my stack. The model becomes dramatically more useful when your operating standards are encoded rather than repeatedly prompted.

Anthropic

Claude for Small Business

Anthropic’s direction for applying Claude, Skills, Cowork, and connected systems to functions across a small business.

Why read it: It points directly toward the kind of AI operating layer I am trying to build.

OpenAI

ChatGPT for Your Most Ambitious Work

OpenAI’s move toward persistent work environments built around files, projects, tools, reusable templates, and multi-step execution.

Why read it: Another sign that AI products are moving from question-answering toward completing bodies of work.

Perplexity

Comet

Perplexity’s AI-powered browser.

Why read it: An example of the browser itself becoming an AI operating surface for research and web-based work.

10. The connective layer: meetings, CRM, and shared context

Zoom

Zoom Meeting Intelligence in Claude

An example of Zoom exposing meeting intelligence to AI systems through MCP.

Why read it: The interesting idea is not another meeting summary. It is the meeting becoming usable context elsewhere in the business.

Attio

Attio MCP

Attio’s integration for allowing AI systems to work against CRM information.

Why read it: A strong example of why the CRM can become the spine of an AI-enabled operating system.

11. What the productivity research says

Noy & Zhang / Science

Experimental Evidence on the Productivity Effects of Generative AI

A controlled experiment examining generative AI’s impact on professional writing tasks.

Why read it: One of the clearest early randomized experiments showing meaningful improvements in speed and judged work quality.

Brynjolfsson, Li & Raymond / NBER

Generative AI at Work

A field study of customer-support workers using generative AI.

Why read it: Particularly relevant for understanding AI as a way of distributing expertise, standards, and best practices.

12. Build vs. buy and the rise of AI-built software

Financial Times

“Vibe Coding” Is the New DIY

A look at natural-language software development and what happens when people without traditional development backgrounds can build useful applications.

Why read it: Good context for the falling cost of creating narrow software.

Financial Times

Why Vibe Coding Isn’t Simply Going to Kill Business Software

The other side of the argument.

Established software vendors still provide security, reliability, support, domain expertise, maintenance, integrations, and infrastructure.

Why read it: Read this next to the piece above. The interesting future is probably somewhere between “buy everything” and “build everything.”

WIRED

Thousands of Vibe-Coded Apps Expose Corporate and Personal Data

Reporting on security and authentication problems found in AI-built applications.

Why read it: An important warning that easier software creation does not automatically produce safe software.

MIT

Can AI Really Code? Study Maps Roadblocks to Autonomous Software Engineering

MIT researchers examine the engineering work that remains difficult even as code generation improves.

Why read it: A useful antidote to both “AI cannot build software” and “software engineering has been solved.”

How I built a five-member Excel team in Claude and how you can too.

John Brewton — Wed, 05 Aug 2026 12:11:45 GMT

The best AI model in one financial-modeling benchmark edits 89.69 percent of the correct cells.

It gets only 34 percent of the final answers right.

That gap explains the problem with most AI spreadsheet workflows.

The model can make a lot of plausible-looking changes. It can move quickly. It can even appear productive. But none of that guarantees that the final answer is correct, traceable, or safe to use.

That is why I spent this morning building a 17-page spreadsheet operating system instead of writing another “better” prompt.

I did not build one all-purpose spreadsheet agent.

I built a five-role team.

Here is how I got there.

Watch my live loom recording of the process here.

First, Understand What the Workbook Is

Before I explain the research or introduce the five roles, it is worth explaining what I mean by a workbook.

This is not a workbook in the traditional sense.

It is not a document you print, complete by hand, and place in a binder. It is also not simply a collection of prompts.

The workbook is a system of instructions that the AI reads and then uses to interview you. You upload it into a new working session and ask the model to walk you through it. The model follows the rules inside the workbook, asks you a structured series of questions, and collects the information it needs to build a working AI team around the way you actually operate.

That distinction matters.

Most generic AI workflows begin with a finished prompt. Someone else decides how the work should be done, packages those assumptions into a template, and hands it to you.

The workbook begins one step earlier.

It helps the model learn how you need the work to be done before it starts building anything. For spreadsheet work, that means collecting details such as:

The types of analysis you perform
The workbooks and files you regularly use
The outputs you need to produce
The formulas and modeling standards you expect
The software and data sources involved
The assumptions that must always be stated
The controls required before work can be published
The recurring tasks that may eventually be scheduled

The result is not one enormous prompt that tries to anticipate every possible spreadsheet request.

The result is the infrastructure needed to handle many different requests consistently.

Why the Workbook Is Built Around Questions

The quality of an AI system depends heavily on the quality of the context behind it.

But most people do not naturally provide that context in one clean, complete prompt.

They leave out details because those details feel obvious to them. They forget to explain how a file is organized, which source should be treated as authoritative, what level of precision is acceptable, or what must happen before an answer can be shared.

Those omissions become failures later.

The workbook is designed to prevent that by turning system design into a guided conversation.

Instead of asking you to sit down and write a complete operating manual from scratch, the model asks one question at a time. Your answers become the raw material for the skills, role instructions, rules, checklists, and scheduled tasks that the team will use.

This makes the process more accessible, but it also makes it more rigorous.

The questions force decisions that a loose prompt allows you to avoid.

What is the team responsible for?
What is outside its authority?
Which files must it inspect?
Which software must it use?
What must it produce before another role can begin?
What conditions should stop the workflow?
How should the final work be checked?

By the end, the model is no longer guessing what “good spreadsheet work” means to you. You have defined it.

Why the Instructions Live in a Workbook

I use a workbook because the system needs to be portable, repeatable, and visible.

A good conversation with an AI can produce excellent work once. But conversations are easy to lose, difficult to audit, and hard to reproduce with another person or in another session.

The workbook gives the process a durable structure.

It contains the questions to ask, the order in which to ask them, the rules that must be followed, and the outputs that should be created from the answers.

That means I can upload the same workbook into a clean session and have the model run the process again without relying on the memory of a previous conversation.

It also means I can give the workbook to someone else.

They do not need to understand every decision I made while designing it. They can upload it, ask the model to follow it, answer the questions, and build a version of the system that reflects their own work.

The workbook is therefore doing two jobs.

First, it captures the operating logic behind the system.

Second, it teaches the user how to build that system for themselves.

Why the Workbook Comes Before the Team

The five roles I built are important, but the roles are not the starting point.

The workbook is.

Before the AI can build an Intake Clerk, Analyst, Modeler, Auditor, or Publisher, it needs to understand what those roles mean inside your environment.

Your Analyst may need to prioritize cash flow, while someone else’s Analyst focuses on inventory or customer acquisition.

Your Modeler may work inside Excel with files stored in SharePoint. Another company may use Google Sheets, a finance platform, and data exported from its ERP.

Your Auditor may need to reconcile every result to the general ledger. Someone else may need to compare it with a prior board report or an external data source.

The role names can remain the same while the actual instructions change substantially.

The workbook collects the information needed to make those roles specific.

It turns a general team design into your team design.

That is why I do not begin by telling the model, “Build me five spreadsheet agents.”

I begin by giving it a structured process for learning what those agents must know, what they are allowed to do, what they must never do, and how their work should move from one role to the next.

The Workbook Is the Control Layer

A useful way to think about the workbook is as the control layer between you and the AI team.

The roles perform the work.

The skills give the roles specialized methods.

The connected software, files, and data give them access to the materials they need.

The scheduled tasks determine when recurring work happens.

But the workbook defines how all of those pieces should be assembled.

It creates the instructions that govern the system before the system begins operating.

That is why I have started building workbooks for more and more types of work. They give people a repeatable way to create new AI capacity without needing me to personally design every team, skill, or workflow for them.

Once you understand how to build one, you can use the same approach for inbox management, research, financial analysis, reporting, project management, or almost any other structured knowledge-work process.

You are no longer collecting isolated prompts.

You are learning how to design the team that will use them.

Now for the process I documented for you today…

Phase 1: Research the failures

Time: 40 minutes

I began with one question:

What actually goes wrong when AI works inside a spreadsheet?

I limited the research to benchmarks and spreadsheet audits. No opinions. No productivity claims. Just documented failure modes.

Three findings shaped the entire system.

1. AI often fails because it does not inspect the workbook properly

SpreadsheetBench 2 evaluates AI agents on multi-sheet workbooks with real formulas, dependencies, and linked calculations.

The best model achieved 34.89 percent overall accuracy.

On debugging tasks, it scored just 12 percent.

The researchers identified two recurring problems:

Insufficient inspection
Incorrect target selection

In plain English, the models often failed before they started calculating.

They did not fully understand the workbook, or they changed the wrong cells.

That is not primarily a math problem. It is a looking problem.

2. AI can produce the right number in the wrong way

WorkstreamBench evaluates agents on end-to-end financial tasks.

The best agent scored 69.1 out of 100. Once the task required several chained calculations, the score fell to 53.4.

One of the most important documented failures was deceptively simple: agents sometimes pasted a correct number into a cell where a formula should have been.

The answer may be right today.

But the workbook is now broken.

Change an assumption next week, and the pasted number will not update. A reviewer cannot follow the calculation. The spreadsheet becomes a static answer rather than a working model.

Would you like to work with John?

Click Here

3. Review cannot be treated as an afterthought

Raymond Panko’s field audits found errors in 91 percent of operational spreadsheets.

A single reviewer typically catches about half of the errors.

A second reviewer, working independently and using a checklist, can raise detection to roughly 80 percent.

That matters because most AI spreadsheet workflows ask one agent to inspect the data, choose the method, build the model, calculate the result, and verify its own work.

That is the spreadsheet equivalent of asking an analyst to approve their own analysis.

Phase 2: Design the team

Time: 20 minutes

The research killed my original idea.

I had planned to build one spreadsheet employee, similar to the inbox employee I taught last month.

But the evidence pointed in another direction.

Inspection, analysis, modeling, and verification are not one job. They are separate jobs with different failure modes.

Combining them inside one large prompt does not create a capable employee. It creates an analyst checking their own work.

So I divided the workflow into five roles.

Each role has:

One primary responsibility
One artifact it must hand to the next role
One action it is not allowed to take

The Intake Clerk

The Intake Clerk profiles the workbook before any analysis begins.

It identifies the sheets, tables, formulas, inputs, outputs, date ranges, missing fields, and potential data-quality problems.

It is not allowed to calculate the answer.

Its job is to understand the environment before anyone starts changing it.

The Analyst

The Analyst decides how the question should be answered.

It defines the method, assumptions, required calculations, and level of precision before producing a result.

The rule is simple:

State the method before stating the number.

The Modeler

The Modeler turns the approved method into a live workbook.

It builds formulas, links assumptions to outputs, and creates a model that recalculates when the inputs change.

It is not allowed to hardcode a value into a computed column.

If a number is calculated, there must be a formula behind it.

The Auditor

The Auditor independently re-derives the headline number using a second method.

It checks the formulas, assumptions, units, ranges, and reconciliation gaps.

It also has the authority to stop the workflow.

If the analysis does not hold, the Auditor issues a hold.

The Publisher

The Publisher prepares the finished workbook and communicates the result.

But it cannot publish over an audit hold.

That rule is what turns the system into a team rather than five prompts pasted one after another.

Every role can reject the work it receives.

Phase 3: Build the operating system

Time: 45 minutes

The finished workbook is 17 pages.

It includes:

Nine questions that define your analysis standards
Eight prohibited spreadsheet habits
Four prompts to run in sequence
Six reusable prompts for future analyses

The prohibited habits may be the most important part.

They include:

Pasting values into cells that should contain formulas
Reporting numbers with no supporting cell reference
Applying assumptions without stating them
Presenting more precision than the underlying data supports
Editing a workbook before profiling its structure
Publishing an answer that has not been independently reconciled

Each prohibition is paired with a replacement rule.

That distinction matters.

Telling an AI what not to do is rarely enough. You also have to define what it should do instead.

And every rule in the workbook traces back to a documented failure mode. These are not formatting preferences. They are controls.

What this system actually gives you

Most people use AI in spreadsheets because they want speed.

But speed is not the main constraint.

A fast answer with no formula trail still has to be checked by hand. That manual review is the work you were trying to eliminate in the first place.

The real opportunity is not faster analysis.

It is additional analytical capacity.

This system gives an operator access to three functions that many small teams have never been able to staff properly.

A data profiler

Someone examines the workbook before the analysis begins.

They identify what is present, what is missing, what is linked, and what may be unreliable.

A modeler

Someone builds a workbook that continues working when the assumptions change.

The goal is not a spreadsheet that was correct once. It is a model that remains useful.

An auditor

Someone independently checks whether the headline number holds.

This is the role most small finance teams do not have. It is also the role the research suggests matters most.

A private company with a four-person finance team may never have had a second analyst independently re-derive a number before it reached the board.

Historically, that verification could cost as much as the original analysis.

It does not have to anymore.

The final output is not simply a number.

It is:

The number.
The method above it.
The formulas beneath it.
And an independent check beside it.

What to try this week

Choose one analysis you completed manually last month.

Pick something where you already know the correct answer.

Then run it through the five roles in order.

Have the Intake Clerk map the workbook before making any changes.

Have the Analyst explain the method before calculating the result.

Have the Modeler prove that the headline cell contains a formula rather than a pasted value.

Have the Auditor calculate the answer a second way and report the difference between the two results.

Every place where the AI’s work differs from your manual process should become a new rule in your operating profile.

Repeat that process for a week, and the system becomes more consistent.

Then use the Auditor on the spreadsheets your team produced last quarter.

Panko found errors in 91 percent of operational spreadsheets.

There is no reason to assume ours are the exception.

Become a paid subscriber to receive the working and build your Excel team this week!

Become An Operating Founder

And if building this kind of AI fluency is especially appealing to you, and you want to learn everything you can in the year ahead, become an Operating Founder and join my 52-week-per-year Friday live Q&A sessions with The Operating Founder Community. We are a +200-strong group of Operators and Builders choosing a new vision for our futures.

Download the workbook to build your Excel team below.

Quiet Quitting or Quiet Building? Our Changing Relationship With Professional Ambition

John Brewton — Tue, 04 Aug 2026 11:52:01 GMT

By John Brewton
Operating by John Brewton

At 5:07 AM, the house is quiet. The work laptop is still closed.

On a different computer, purchased with personal money and connected to personal accounts and applications, and filled with photos from last week’s Summer vacation to the beach, someone is preparing a proposal for a prospective customer to be sent before breakfast. At 8:30 AM, the computer will be put away, and at 9:00 AM, this person will begin the job that still pays the mortgage, provides health insurance, and finances the experiment.

From the company’s perspective, something may have changed.

This employee is no longer volunteering for every optional committee. They are not looking for another internal initiative to inherit. They are not treating every evening as overflow capacity for the organization. They still do the job. They still meet their commitments. They still act professionally.

But the surplus ambition has gone somewhere else.

Corporate America may call this quiet quitting.

I think something more interesting is happening.

For generations, the standard path for an ambitious person was reasonably clear. Join a good organization. Learn its systems. Take on increasingly important assignments. Earn promotions, bonuses and stock options. Accumulate responsibility and convert extraordinary effort into a better position inside the institution.

There is nothing wrong with that path.

A strong company can provide training, mentorship, capital, infrastructure, meaningful collegial relationships, distribution and opportunities that would be difficult to assemble independently. For many people, climbing the corporate ladder will remain the most intelligent way to build a career and provide for a family.

But it is no longer the only serious route available.

Cloud software, global payment systems, digital distribution, online marketplaces, independent contractors and generative AI have dramatically reduced the amount of infrastructure a person needs to begin operating a business. Work that once required a small department can increasingly be performed by one capable operator coordinating software, automation and outside specialists.

That does not mean building a company has become easy.

It means it has become possible for more people to try.

And that changes the destination and direction of our collective ambition.

Here’s everything you need to know about the business I have built so far.

Quiet quitting is the wrong phrase for a real shift

“Quiet quitting” became a popular term in 2022, but it has never had a stable definition.

For some people, it means refusing to work beyond the responsibilities and hours of the job. For others, it means becoming psychologically detached while continuing to collect a paycheck. For some managers, it describes any employee who stops providing an unlimited supply of discretionary effort.

Gallup’s original analysis associated quiet quitting with employees who were “not engaged.” These were people who might perform the minimum required but felt psychologically disconnected from the organization. Gallup estimated that this group represented roughly half of the U.S. workforce in 2022.

But Derek Thompson argued in The Atlantic that the supposed trend was neither statistically new nor especially revolutionary. What social media had renamed quiet quitting had previously been known as having a job without allowing it to consume your identity.

Both interpretations contain some truth.

The label may exaggerate the novelty of workplace disengagement. But the underlying tension is real. In the first half of 2026, Gallup reported that 31% of American employees were engaged and 18% were actively disengaged. Engagement remained well below its 2020 level.

At the same time, this is not a universal story of workers rejecting their careers. Pew Research Center has found that most employees, including most younger employees, remain at least somewhat satisfied with their jobs.

The phrase quiet quitting therefore collapses several different behaviors into one:

There is boundary-setting, in which an employee performs the agreed job but stops treating every optional request as an obligation.
There is withdrawal, in which an employee becomes indifferent to quality, customers, and colleagues.
And there is strategic reallocation, in which an employee continues to perform professionally but directs more of their nonworking time toward building an asset they own.

This article is about the third group.

It is not an argument for doing poor work, falsifying hours, missing deadlines, or transferring your workload to coworkers. As Anthony Klotz and Mark Bolino noted in Harvard Business Review, withdrawal can damage team relationships when colleagues must absorb the responsibilities someone else has abandoned.

That is not entrepreneurship.

That is simply failing to keep a promise.

The defensible version is much cleaner:

Do the job. Meet the standard. Keep your commitments. But stop assuming that every remaining ounce of ambition belongs to the company issuing your paycheck.

The corporation has lost its monopoly on leverage

The modern corporation became powerful partly because it could coordinate resources an individual could not afford alone.

It assembled people, information, machinery, technology, distribution and capital under one roof. If you wanted access to that leverage, you usually needed to enter the institution and rise through it.

That bargain still exists. But technology has weakened its exclusivity.

Research on digital entrepreneurship has documented how technologies including cloud computing, mobile platforms, analytics and social media have changed the boundaries of entrepreneurial activity. Founders can acquire capabilities as services, reach customers directly and reorganize ventures more rapidly than was previously possible.

Generative AI has accelerated that change, although its effects are uneven.

In a controlled experiment involving professional writing tasks, Shakked Noy and Whitney Zhang found that participants using ChatGPT completed their work about 40% faster while average output quality rose by 18%.

A large workplace study subsequently published in The Quarterly Journal of Economics found that access to an AI assistant increased productivity among customer-support agents by an average of approximately 15%, with the largest gains accruing to less experienced workers.

But AI is not a universal competence machine. Research involving 758 Boston Consulting Group professionals found a “jagged technological frontier.” AI substantially improved performance on some knowledge tasks while making people more likely to produce incorrect answers on tasks outside the technology’s capabilities.

That distinction matters.

AI can increase the reach of judgment. It cannot eliminate the need for judgment.

It can help one person research, write, code, analyze, respond to customers, and automate administration. It can also help that person make a polished mistake much faster.

The opportunity is real, but so is the obligation to verify the work.

The one-person company is no longer structurally absurd

The solo economy is already much larger than the venture-capital headlines suggest.

The U.S. Census Bureau counted approximately 30.4 million nonemployer businesses generating nearly $1.8 trillion in receipts in 2023. “Nonemployer” means a business has no paid employees; it does not necessarily mean that every business is a technology-enabled company operated entirely by one person. But the numbers demonstrate how much economic activity already occurs outside conventional employer organizations.

At the upper end, a July 2026 Wall Street Journal investigation reported that thousands of solo operators using Stripe were producing more than $1 million in annual revenue. According to Stripe’s analysis, the number crossing $1 million doubled between 2023 and 2025, while the number crossing $10 million nearly tripled.

Stripe’s customers are not a statistically representative sample of every American business. Seven- and eight-figure solo companies remain unusual. Survivorship bias is substantial, and the same tools that reduce the cost of launching a business also reduce the cost of copying one.

The point is not that everyone can build a $10 million company alone.

The point is that doing so is no longer operationally impossible.

The corporation has not become obsolete.
It has lost its monopoly on leverage.

That creates a new option for a certain kind of ambitious person.

Instead of asking, How do I become indispensable to this institution?, they can begin asking, How do I build something that can operate without the institution’s permission?

Instead of placing every discretionary hour into internal visibility, they can place some of those hours into customer relationships.

Instead of waiting for an equity grant, they can begin accumulating equity directly.

Instead of concentrating their income, status and future inside one organization, they can begin constructing an alternative.

This is not an anti-corporate argument.

It is a pro-optionality argument.

The hybrid phase is not hesitation

Researchers have a more precise name for starting a business while remaining employed: hybrid entrepreneurship.

In a foundational Management Science study, Timothy Folta, Frédéric Delmar and Karl Wennberg defined hybrid entrepreneurs as people who enter self-employment while retaining their primary wage-paying jobs. Their research found that hybrid founders represented a significant portion of entrepreneurial activity and could use the overlap period to learn before entering full-time self-employment.

Joseph Raffiee and Jie Feng later tracked thousands of American entrepreneurs over 15 years. Their Academy of Management Journal study found that people who transitioned through hybrid entrepreneurship had substantially better venture-survival outcomes than those who immediately moved from employment into full-time entrepreneurship. The result is often summarized as a roughly 33% lower likelihood of failure.

Keeping the job is not evidence that you lack conviction. It may mean you are purchasing information before purchasing risk. The salary pays household expenses while you, the founder, learn:

Will anyone buy this?
Can I reach customers consistently?
Can I deliver the work profitably?
Will customers renew?
Is the opportunity large enough to justify leaving?
Do I enjoy operating this business once the novelty disappears?

The day job also prevents desperation from becoming the company’s first investor.

A founder who needs immediate income may accept the wrong customer, underprice the work, build for anyone willing to pay, or keep an idea alive because admitting failure would threaten next month’s rent.

A founder with a salary has more ability to say no:

A paycheck can become venture capital you provide to yourself, one deposit at a time.

This is the real opportunity inside quiet quitting.

Not doing nothing.

Building patiently.

Seven priorities while the day job still pays

The following are some operating rules, not universal scientific laws, that I’ve constructed based on my own experience making this drum in my career over a period of years.

Research supports the logic of experimentation, staged entry, clear boundaries, and financial caution. It does not establish that every founder needs exactly 18 months of cash, exactly three contracts or exactly three profitable months.

Those are decision thresholds.

Their purpose is to help solo founders switch out their feelings for actual evidence.

1. Bank eighteen months of runway

Runway is not the amount of money in a savings account.

It is the number of months your household and business can continue operating without requiring optimistic assumptions.

Start with the real monthly number:

Household fixed costs

business fixed costs
replacement benefits
taxes
contingency allowance

Then multiply it by 18.

Health insurance matters. Software subscriptions matter. Equipment matters. The tax reserve matters. The expenses your employer currently absorbs matter.

Eighteen months is intentionally conservative. It is not a magical figure. Some founders with highly predictable contracts may require less. Others with dependents, debt, long development cycles or volatile markets may need more.

The reason for caution is straightforward: new businesses remain uncertain. According to the SBA Office of Advocacy’s 2026 small-business data, only about 49% of new employer establishments survive five years. That statistic covers many kinds of businesses and does not predict the fate of any one company, but it is a useful correction to launch-day optimism.

Runway gives you time to survive delayed contracts, customer churn, pricing mistakes, and slower sales cycles.

More importantly, it improves the quality of your decisions.

2. Sell before you build

Three signed contracts beat a finished product.

This does not mean accepting money for something you cannot responsibly deliver. It means testing the commercial assumption before investing months in infrastructure.

Entrepreneurship is fundamentally a process of experimentation. William Kerr, Ramana Nanda and Matthew Rhodes-Kropf argue in the Journal of Economic Perspectives that entrepreneurial success probabilities are often low, highly skewed and unknowable until resources are committed. The answer is not to eliminate uncertainty. It is to structure smaller experiments that reveal information before making larger commitments.

Research on a scientific approach to entrepreneurial decision-making similarly suggests that entrepreneurs benefit when they state assumptions clearly, test them and allow evidence to kill weak ideas rather than searching only for confirmation.

The progression should usually look something like this:

Most founders reverse that sequence.

They build the platform, design the brand, automate the workflow and prepare for scale before discovering whether someone has a sufficiently painful problem and a sufficiently large budget.

A working product proves that you can produce supply.

A signed customer proves that demand exists.

Three customers do not prove you have discovered a large market. But they prove that the idea has survived contact with someone else’s bank account.

3. Audit Your Current Employment Paper

Before building anything, read every document governing your employment.

Not the summary.

Not the onboarding slide.

The entire handbook and everything you signed.

Review the provisions involving:

Intellectual-property assignment.
Confidentiality and trade secrets.
Outside employment or moonlighting.
Conflicts of interest.
Non-solicitation.
Noncompetition.
Customer and employee relationships.
Required disclosure or approval.
Notice periods.
Ownership of inventions created during employment.

Then have an employment lawyer in your jurisdiction review the provisions that matter.

Do not assume that noncompetes are universally unenforceable. The Federal Trade Commission’s proposed nationwide rule is not in effect and is not enforceable. After a federal court blocked it, the FTC dismissed its appeals in 2025. State law and individual contracts therefore remain central to enforceability.

A founder may also have duties while still employed that extend beyond the words “noncompete.” Employment-law scholarship has examined how duties of loyalty can apply to competition, corporate opportunities and employer resources during an active employment relationship. The exact obligations vary by role and jurisdiction.

A few hours of legal review are cheaper than discovering that your employer claims ownership of the business after it becomes valuable.

Do not take legal advice from a founder whose contract, role and state are different from yours. Find a good lawyer and listen to them

4. Keep the separation clean

Use personal hardware.
Use personal accounts.
Use personal software subscriptions.
Use personal payment methods.
Work during personal time.
Do not transfer files, prompts, templates, customer lists, code, research, pricing information or internal documents from the employer’s environment into your own.
Do not recruit coworkers or approach company customers without specific legal clearance.
Do not build a competing product while sitting in a meeting, even when the meeting should have been an email.

Clean separation is not only an ethical principle. It creates an evidentiary record.

File metadata, device logs, account histories, and timestamps should tell the same story you would tell.

Your standard should not be, “Could I explain this casually?”

Your standard is:

Could I establish a clean chain of creation if every device, document and timestamp were examined?

Nothing built on company time, with company data or through company systems, is worth the future dispute.

5. Count customers, not followers

Audience is useful, and feels good to build, but having an audience does not equate to having revenue, actual money in the bank.

Ten thousand impressions may tell you that a sentence was interesting. Ten paying customers tell you that a problem is expensive enough for someone to solve, and that your solution has done just that for those ten people.

The progression of evidence is simple:

Each step requires more commitment from the other person. Each step therefore provides more useful information.

Followers can reduce future distribution costs. A respected audience can become a meaningful asset. But founders often confuse public encouragement with commercial demand because public encouragement is easier to obtain and more pleasant to measure.

A stranger commenting “This is brilliant” risks nothing.

A purchasing manager signing a contract risks budget, reputation and time.

Early in the business, optimize for learning rather than applause.

Concentrated revenue may eventually become a problem. Depending on one customer can create fragility. But zero revenue is a more immediate problem than concentrated revenue.

Win the first customer.

Then the second.

Then determine whether you can win the third without relying on luck.

6. Protect the first ninety minutes—and a defined portion of the weekend

The best time block is the one you can defend consistently.

For many employed founders, that means working before the day job begins.

The morning has an advantage that has little to do with motivational mythology. It has not yet been consumed by someone else’s emergencies. The calendar has not slipped. The difficult client has not called. The meeting has not run long. Your attention has not been divided among twelve unfinished tasks.

Research on attention residue shows that switching from one unfinished task to another can impair performance because part of our attention remains attached to the previous activity. This helps explain why fragmented evening work often feels slower than protected, uninterrupted creation. (ScienceDirect)

The side business should therefore have recurring appointments with its founder.

Protect the first ninety minutes on selected weekdays. Use one defined weekend block for deeper work, customer delivery or sales. Make the schedule visible to the people whose lives are affected by it.

But do not convert every morning, evening and weekend into work.

Research on side hustles is genuinely two-sided. An Academy of Management Journal study found that empowerment gained through outside work could enrich performance in a person’s primary job. More recent Journal of Management Studies research found both gains and costs: thriving in a side hustle can improve well-being, but attention residue and resource depletion can also damage primary-job performance. Clear segmentation between roles helps. (Academy of Management Journals)

Recovery is not laziness either. A major Annual Review of Organizational Psychology and Organizational Behavior review concluded that detachment and recovery during evenings, weekends and vacations are important for sustaining well-being, motivation and performance.

Do not finance the company with chronic sleep deprivation, family resentment or deteriorating performance at the job still funding it.

7. Set a trigger, not a date

“I will quit in June” is not a strategy.

June does not know whether customers will renew.

June does not know whether the pipeline is real.

June does not know whether your largest client will disappear in May.

A date creates emotional momentum. A threshold creates decision discipline.

My preferred core trigger is:

Leave when recurring business revenue covers household and business fixed costs for three consecutive months—with the original runway substantially intact.

That is the beginning of the analysis, not the end.

Before leaving, also examine:

Whether customers are renewing or purchasing again.
Whether the pipeline can be repeated.
Whether revenue depends excessively on one relationship.
Whether gross margins support the owner after taxes and benefits.
Whether the founder has clean ownership of the work.
Whether additional founder time is likely to create additional revenue.

The final question is especially important.

Do not resign merely because operating the business while employed has become uncomfortable.

Resign when employment has become the principal constraint on a model that has already demonstrated demand.

The difference is enormous.

In the first situation, the founder is hoping that more time will create a business.

In the second, the founder has evidence that more time can expand an existing business.

The moral contract

There is a version of this strategy that is disciplined and honorable.

There is also a version that is dishonest.

The line is not complicated.

Your employer is entitled to the professional performance it pays for. Your coworkers should not be required to subsidize your company by carrying abandoned responsibilities. Customers should not receive inferior service because you were thinking about your own sales pipeline. Confidential information is not founder capital. Paid working hours are not a private incubator.

Keep your promises.

Meet the standard.

Disclose conflicts when the agreement requires disclosure.

Leave employer information, relationships, and property alone.

But the employment relationship has a boundary, too.

An employer may purchase your labor during agreed hours. The contract and the law may impose additional obligations. None of that automatically means the institution owns every future idea, every weekend, every early morning or the permanent direction of your ambition.

The business must be built outside the job—not instead of the job.

That is why I prefer quiet building to quiet quitting.

Quiet quitting sounds like retreat.

Quiet building is deliberate preparation.

What quiet building does not guarantee

Not every disengaged employee is a future founder. There is no credible research establishing that quiet quitting has caused the growth of solo entrepreneurship.

That connection is an interpretation of two changes occurring at the same time:

Work has become less central to the identities of some employees, while technology has made independent ownership more feasible.

This strategy is not equally available to everyone. Spare time, financial reserves, predictable schedules, health, caregiving responsibilities, and access to technology are distributed unevenly. Some people need the stability of employment more than they need entrepreneurial optionality. Some people genuinely enjoy institutional leadership and can create more value inside a large organization than they would alone.

There is no moral superiority in becoming a founder.

There is also no guarantee that the business will work.

AI does not create demand. A large audience does not create willingness to pay. More hours do not repair a weak proposition. The side business may produce energy, or it may create exhaustion. A secure salary may support intelligent experimentation, or it may allow the founder to avoid making a decision indefinitely.

The point is not that everyone should leave.

The point is that ambitious people have another route available—and the first stage of that route may occur while they are still employed.

Build Quietly.

At 8:30 AM, the founder closes the personal computer.

At 9:00 AM, the employee begins the workday and keeps the promises that have been made.

The employer still receives competent, professional performance.

What the employer no longer receives is the unquestioned right to every discretionary hour and every future ambition.

The founder keeps learning.

Keeps selling.

Keeps separating the work.

Keeps accumulating cash.

Keeps watching the numbers.

The first act of entrepreneurship is rarely a job resignation.

It is the choice to learn to wait and quietly build.

Keep the salary.
Protect the dream.
Perform honestly.
Leave when the arithmetic gives you permission.

- j -

Do you want some help starting your Quiet Build?

Become an Operating Founder, receive a 60-minute review of your content, business, and strategy, and join my weekly Operating Founder Office Hours, where I answer any and all of your questions and share what I’m working on and learning from building my business.

Become a Paid Subscriber Today!

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of Operating by John Brewton and conducts advisory and consulting work for global companies, asking the question: What is the future of companies?

Research appendix

A. Quiet quitting, engagement and the counterargument

1. Gallup: Is Quiet Quitting Real?

Read the Gallup analysis

Gallup’s 2022 analysis defined many quiet quitters as “not engaged”: employees who perform minimum requirements while remaining psychologically detached. The article linked falling engagement to unclear expectations, fewer development opportunities and weak manager relationships.

Used for: The definition of quiet quitting as disengagement and the estimate that roughly half of U.S. employees were not engaged in 2022.

2. Gallup: Employee Engagement Remains Flat as AI Adoption Accelerates

Read the 2026 Gallup report

Gallup reported that 31% of U.S. employees were engaged during the first half of 2026 and 18% were actively disengaged. It also emphasized that access to AI alone did not improve engagement; clear implementation and manager support mattered.

Used for: Current U.S. engagement figures and the argument that managerial context remains important.

3. Derek Thompson, The Atlantic: Quiet Quitting Is a Fake Trend

Read the article

Thompson argued that quiet quitting was not statistically new and that modest employee disengagement had existed long before the phrase went viral. He interpreted the popularity of the term as evidence of cultural frustration rather than proof of a sudden labor-market transformation.

Used for: The counterargument that the label exaggerates the novelty of the phenomenon.

4. Anthony Klotz and Mark Bolino, Harvard Business Review: When Quiet Quitting Is Worse Than the Real Thing

Read the article

The authors distinguished legitimate boundary-setting from forms of withdrawal that can damage relationships, performance and coworkers.

Used for: The ethical distinction between declining optional work and abandoning agreed responsibilities.

5. Pew Research Center: Young Workers Express Lower Levels of Job Satisfaction Than Older Ones, but Most Are Content With Their Job

Read the Pew analysis

Pew found that younger workers reported lower levels of strong job satisfaction than older workers, while 85% of workers overall remained at least somewhat satisfied.

Used for: Balance against the claim that an entire generation has rejected conventional employment.

B. Hybrid entrepreneurship and staged entry

6. Timothy B. Folta, Frédéric Delmar and Karl Wennberg: Hybrid Entrepreneurship

Read the Management Science abstract and publication record

The researchers defined hybrid entrepreneurship as entering self-employment while retaining a primary wage job. Using Swedish employee-employer data, they found that hybrid founders represented a significant share of entrepreneurial activity and examined how hybrid entry can facilitate later movement into full-time self-employment.

Used for: The formal definition of hybrid entrepreneurship and the concept of incremental entry.

7. Joseph Raffiee and Jie Feng: Should I Quit My Day Job? A Hybrid Path to Entrepreneurship

Read the Academy of Management Journal study

The authors studied thousands of entrepreneurs over a 15-year period and found that founders who entered through a hybrid stage experienced significantly better survival outcomes after transitioning to full-time entrepreneurship.

Used for: The claim that keeping a job while validating a venture can reduce the risk of failure.

8. Harvard Business Review: Why Going All-In on Your Start-Up Might Not Be the Best Idea

Read the research summary

This summary describes the Raffiee and Feng result as a 33% reduction in the likelihood of failure for founders who left their employment in stages.

Used for: The accessible interpretation of the academic result.

9. William Kerr, Ramana Nanda and Matthew Rhodes-Kropf: Entrepreneurship as Experimentation

Read the Journal of Economic Perspectives paper

The authors describe entrepreneurship as experimentation under conditions in which success probabilities are low, skewed and often unknowable before investment.

Used for: The argument that founders should structure smaller learning investments before making larger commitments.

10. Arnaldo Camuffo, Alessandro Cordova, Alfonso Gambardella and Chiara Spina: A Scientific Approach to Entrepreneurial Decision Making

Read the Management Science study

In a randomized controlled trial involving Italian startups, the researchers tested whether teaching founders to formulate hypotheses and evaluate evidence changed their decisions. The study supports disciplined testing over purely intuitive or confirmatory decision-making.

Used for: “Sell before you build” and the need to design experiments capable of disproving weak assumptions.

11. U.S. Small Business Administration: Plan Your Business

Read the SBA guidance

The SBA emphasizes researching demand, market size, saturation, pricing and customer behavior before making major investments.

Used for: The practical importance of direct customer research and demand validation.

C. Side hustles, focus and recovery

12. Hudson Sessions, Jennifer Nahrgang, Manuel Vaulont and colleagues: Do the Hustle! Empowerment From Side-Hustles and Its Effects on Full-Time Work Performance

Read the Academy of Management Journal study

The study found that empowerment generated by a side hustle could produce positive affect and enrich an employee’s performance in their primary job.

Used for: The evidence that outside work does not automatically damage full-time performance.

13. Min Liu, Shengxian Yu, Xin Liu and Shanshi Liu: The Effect of Side-Hustle Thriving on Full-Time Work Performance

Read the Journal of Management Studies study

The authors identify two competing pathways. Side-hustle thriving can improve well-being and help employees detach from their full-time work, but it can also create attention residue and deplete resources. Boundary segmentation influences the balance.

Used for: The two-sided treatment of side hustles and the importance of separating roles.

14. Sophie Leroy: Why Is It So Hard to Do My Work? The Challenge of Attention Residue When Switching Between Work Tasks

Read the publication abstract

Leroy’s research shows that people can retain cognitive attention on an unfinished task after switching to another activity, impairing subsequent performance.

Used for: The recommendation to create protected blocks instead of repeatedly switching between employer and founder work.

15. Sabine Sonnentag, Charlotte Venz and Anne Casper: Advances in Recovery Research

Read the Annual Review article

The review synthesizes evidence on recovery during breaks, evenings, weekends and vacations, emphasizing the role of recovery in sustaining well-being, motivation and performance.

Used for: The argument that weekends cannot become a second uninterrupted workweek.

D. Technology, AI and the solo economy

16. Satish Nambisan: Digital Entrepreneurship: Toward a Digital Technology Perspective of Entrepreneurship

Read the Entrepreneurship Theory and Practice article

Nambisan examines how cloud computing, mobile technologies, digital platforms and other tools alter entrepreneurial processes, agency and uncertainty.

Used for: The argument that digital technology has reduced the institutional infrastructure required to launch a venture.

17. Shakked Noy and Whitney Zhang: Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence

Read the peer-reviewed Science article

In a preregistered experiment involving professional writing tasks, participants using ChatGPT completed their work approximately 40% faster and produced outputs rated approximately 18% higher in quality.

Used for: Evidence that generative AI can increase individual productivity in defined professional tasks.

18. Erik Brynjolfsson, Danielle Li and Lindsey Raymond: Generative AI at Work

Read the Quarterly Journal of Economics article

The researchers studied the introduction of an AI assistant among more than 5,000 customer-support agents. Productivity increased by approximately 15% on average, with larger benefits for less experienced employees.

Used for: Evidence that AI can transfer elements of organizational knowledge and improve certain workplace outputs.

19. Fabrizio Dell’Acqua and colleagues: Navigating the Jagged Technological Frontier

Read the Organization Science article

The field experiment involving 758 BCG consultants found that AI improved performance significantly on tasks inside its capabilities but reduced accuracy on some tasks outside that frontier.

Used for: The warning that AI leverage remains uneven and requires human verification.

20. U.S. Census Bureau: Census Bureau Data Tell the Small Business Story

Read the 2026 Census analysis

The Census Bureau reported 30,427,808 U.S. nonemployer establishments in 2023, producing nearly $1.8 trillion in receipts.

Used for: The scale of economic activity conducted by businesses with no paid employees.

21. Stripe: Solo Founding Is at an All-Time High

Read Stripe’s analysis

Stripe reported growth in the number of high-revenue solo founders on its platform, while also documenting a wide and increasing gap between median and top-performing solo businesses.

Used for: Evidence of the opportunity and extreme outcome dispersion within solo entrepreneurship.

22. Te-Ping Chen, The Wall Street Journal: The Rise of Million-Dollar Companies With Just One Employee

Read the article

The article reports Stripe data indicating that thousands of solo operators generated more than $1 million in annual revenue. It also presents both successful and unsuccessful founder cases and describes the cost, competition and replication risks created by AI.

Used for: Current evidence that high-revenue solo businesses are becoming more common, while remaining exceptional and risky.

E. Legal boundaries

23. Federal Trade Commission: Noncompete Rule

Read the official status page

The FTC states that its nationwide Noncompete Rule is not in effect and is not enforceable. A district court blocked the rule, and the agency dismissed its appeals in 2025.

Used for: The warning that founders cannot assume a federal blanket ban invalidates their agreements.

24. Federal Trade Commission: FTC Files to Accede to Vacatur of the Non-Compete Clause Rule

Read the official FTC release

The release documents the FTC’s decision to dismiss its appeals and accept the rule’s vacatur.

Used for: Confirmation of the current federal procedural position.

25. Deborah A. DeMott: The Restatement of Employment Law and the Duty of Loyalty

Read the Cornell Law Review essay

The essay examines when employees may owe duties of loyalty and the difficulties involved in defining those obligations.

Used for: The caution that current employees may face legal duties involving competition, employer opportunities and employer resources even apart from noncompete clauses.

F. Business survival and risk

26. SBA Office of Advocacy: Frequently Asked Questions About Small Business, 2026

Read the 2026 SBA report

The SBA reports a five-year survival rate of approximately 49.2%, a ten-year survival rate of approximately 33.9% and a fifteen-year survival rate of approximately 25.5% for new employer establishments in the periods studied.

Used for: Context on business uncertainty and the rationale for maintaining substantial financial runway.

Research and legal note: The numerical thresholds in this article—including 18 months of runway, three initial contracts and three consecutive months of cost-covering revenue—are operating heuristics, not findings established by the cited studies. Employment, intellectual-property, tax and restrictive-covenant laws vary by jurisdiction and circumstances. Readers should consult qualified legal and financial advisers before acting.

Why We Business Builders, Writers and Creators Should Want Substack To Become the Next Giant Social Media Platform

John Brewton — Sun, 02 Aug 2026 11:41:03 GMT

A growing number of Substack creators believe the platform is losing its way.

Paid subscriptions are getting harder to grow.

Notes rewards people who post constantly.

Followers do not feel as valuable as email subscribers.

Video, Lives, clips, recommendations, and algorithmic feeds are becoming more prominent, while also adding to the overwhelm some creators feel about the work required to grow and maintain their audiences across the platform.

Scott Carney recently argued that “the subscription model has maxed out.” Taylor Lorenz wrote that she has more than 200,000 followers while her paid subscriptions have “fallen off a cliff,” making Substack feel increasingly like another social media feed.

Their experiences should be taken seriously.

Their conclusion should not.

My response to the concern that Substack is becoming another giant social platform is simple:

Good. It should.

Substack’s transformation into a large social platform is not a betrayal of the original value proposition that the paid newsletter model offered, nor is it an inevitable path to declining quality of content, community, and creator revenues.

It is an opportunity for all involved to grow.

The real danger is not that Substack becomes larger, more algorithmic, more social, or less about writing and more about multimedia.

The danger is that it becomes the wrong kind of social platform.

Here’s everything you need to know about the business I have built so far.

Harder does not mean worse

Every valuable market becomes more competitive as more people discover its value.

Early Substack writers entered a much emptier marketplace.

There were fewer publications, fewer professional creators, fewer competing Notes, and fewer sophisticated subscription businesses. Category growth lifted many writers at once.

That was never going to last forever.

Companies, especially the venture-backed kind, have a requirement to grow at all costs.

People frequently describe a platform as “better” when what they really mean is that there were fewer people competing with them on it, so they personally felt more successful.

The loss of early-mover ease is not the same thing as platform decline.

Substack passed five million paid subscriptions in March 2025, after crossing one million in 2021 and three million in February 2024. By April 2026, Substack said nearly 100,000 publishers were earning money, including almost 30,000 outside the United States.

The latest official paid-subscription milestone I can verify is still the five-million figure announced in March 2025. We should not pretend we know that paid subscriptions are continuing to accelerate at the same rate.

But “growth may be slowing” and “the platform is collapsing” are radically different claims.

An individual writer can lose paid subscribers while the total market grows.

A category can become saturated while the broader platform expands.

A creator’s conversion rate can decline because the platform is bringing a much larger and colder audience into the top of the funnel.

That last distinction matters.

A follower who encounters one 100-word Note is not the same kind of prospect as a reader who intentionally visits a publication, reads three essays, and gives the writer their email address upon subscription sign-up.

When a platform adds a high-volume, low-intent discovery channel, the blended conversion percentage can fall even as the number of opportunities to convert readers rises.

That is not necessarily cannibalization.

It is a different funnel.

Email is not a discovery system

Email is one of the most valuable distribution tools a creator can possess.

It is direct.

It is portable.

It reaches people who have explicitly asked to hear from you.

I have grown a social following of +80,000 people across platforms over the last two and a half years, but the single most valuable assets I have after all of that work are the 2,250 people of the 6,000 total subscribers on my Substack that open every single article I write and every single offer I make.

Substack also allows creators to export their subscriber lists.

But email begins after discovery.

Before someone can subscribe, they must first encounter the creator. Then they must understand the creator’s value. Then they must develop enough interest or trust to invite that creator into their inbox.

A traditional newsletter platform helps you communicate with the people you already have. A media network helps you find the people you do not have yet.

That is the strategic difference between Substack and infrastructure-first alternatives such as Kit or Beehiiv.

Those platforms may offer better economics, customization, advertising products, or operational control for certain publishers. Substack’s differentiated value is the network surrounding the publishing software. The larger that aggregate market becomes, the more differentiated value is created for all Substack creators.

Substack says its network, anchored by recommendations and the app, now produces more than half of all subscriptions and 30% of paid subscriptions on the platform. In February 2024, it reported that the network was already driving 50% of new subscriptions and 25% of new paid subscriptions.

That reality has certainly held for me. While I arrived at Substack with a following of +25,000 on LinkedIn (now +40,000) and an old email list from that effort, I chose to start from zero and not upload any prior accumulation of email addresses into. I came to Substack because I wanted the network effects, and so I wanted to understand what it meant to grow organically. Here’s what that has looked like for me.

The feed is the customer-acquisition engine for the newsletter.

Without Notes, recommendations, follows, profiles, search, Lives, video, clips, and algorithmic distribution, Substack is just an email service with payment processing and fails to deliver the mechanisms that have created so much opportunity for its community of creators.

In order for our collective opportunity to grow, Substack needs a large geography to direct their machinery against. In the absence of more consuming eyeballs and creating platform members, Substack fails, our new businesses fail, and we move on to the next thing.

Now let’s talk about the difference between social media and interesting media…

TikTok’s real innovation was not short video

TikTok changed media distribution because it reduced the importance of the social graph. Older social platforms began by asking:

Who are your friends?
Whom do you follow?
Which accounts have you already chosen?

TikTok began with a different question:

What are you likely to care about?

Its For You system learns from behavior such as what a person watches, finishes, skips, likes, shares, searches for, or marks as uninteresting.

TikTok has said that follower count and a creator’s previous viral success are not direct ranking factors in its recommendation system. The objective is to match an individual item with people likely to value it, including people who have never encountered its creator before. TikTok’s explanation is here.

This is how they invented the Interest Graph that has become the underpinning for what all social algorithms are being engineered to achieve across all the social media platforms.

The old feed asked:

What did the people you follow publish?

The new feed asks:

Of everything published, what are you most likely to care about?

The rest of the industry has followed.

YouTube says recommendations generate more viewing than subscriptions or search.

Meta explicitly describes Facebook’s future as “social discovery” and uses AI to recommend content from accounts people do not follow.

LinkedIn’s researchers describe retrieving out-of-network posts based on a member’s topical interests, narrowing hundreds of millions of possible posts to roughly 2,000 candidates.

This is not an isolated product trend.

It is the migration from social media to interest media.

Substack must find the voice

Substack should not simply copy TikTok. TikTok is primarily trying to identify the next video someone will watch. Substack must answer a more valuable question:

Which creator, publication, perspective, or body of work might this reader value repeatedly?

TikTok finds a single video, a single, often disposable piece of content.

Substack must find the voices that their platform member would most likely want to consume multiple pieces of content from.

That requires an interest graph built for ideas and expertise.

It should understand what a creator writes and speaks about across an entire archive. It should understand the subjects readers identify in their profiles, the publications they return to, the Notes they save, the Lives they attend, the writers they follow, the subscriptions they purchase, and the subscriptions they retain.

It should then connect and direct those signals to drive a more valuable content consumption experience for the user, and it should provide the creator on the other side with as much clarity as possible for what pieces of content drove that new post engagement, follow choice, subscription click, and paid commitment.

A reader who repeatedly engages with organizational psychology, AI, leadership, and company operations should be able to discover a new writer working at that exact intersection, even if the writer has 300 followers rather than 300,000.

That is the promise of the interest graph.

More creators entering the platform does increase competition.

But it also gives the platform more ideas, more expertise, more behavior, and more potential matches from which to learn.

Growth creates crowding, yes, but it also creates richer discovery.

Following should be a signal, not an entitlement

The critics would be right if Substack’s future were simply a larger follower hierarchy. If the biggest accounts automatically receive the most distribution, early incumbents will own the feed.

Follower counts will become status markers.

Frequency will defeat substance.

Outrage will defeat expertise.

New creators will technically be allowed to publish but structurally prevented from being discovered.

That is not the social platform Substack should build.

Substack describes following as a lighter-weight way for readers to get to know a creator before committing to a subscription. That is exactly how following should work. A follow should help the platform understand that a particular reader may value a particular creator.

A creator’s total follower count should not grant permanent ownership of everyone else’s attention.

Following should be an input into relevance, not a deed to the feed.

Each Note, essay, podcast, clip, and Live should still have to earn distribution by being useful to the people receiving it.

I find that many of my most successful notes drive plenty of engagement and new followers, but not any new subscribers. Here’s an illustration:

In terms of engagement, this was my best-performing note from the last seven days. It did not drive a single new subscriber, but it did add quite new followers. I wouldn’t expect this note to have driven new subscribers. It was simply a funny picture, but it did introduce my work to 1,277 new Substackers via the feed. As the platform grows, that 1,277 will become 12,777, the number of followers a note like this drives will be greater, and ultimately the number of conversions from followers to free subscribers, and yes, to paid, will grow. That only happens if Substack continues to mature into a full-fledged social media platform. My biggest growth spurt of new free and paid subscribers (the period that made me a best seller) started with the outperforming success of two notes that led to 27 of my articles 5x-ing their total views over a period of several weeks.

Substack should normalize performance for creator size, test work with appropriately matched audiences, and reserve meaningful discovery capacity for smaller and unfamiliar publications. The point is not to penalize large creators. It is to prevent accumulated popularity from replacing present-tense relevance.

Optimize for relationships, not reactions

Substack also has an economic advantage that TikTok, Instagram, and X do not.

Its subscription-centered business model benefits when a reader establishes a durable financial relationship with a creator. (Substack)

Its recommendation system should therefore optimize beyond likes, comments, restacks, and time spent.

It should ask:

Did the reader visit the publication?
Did they follow?
Did they become a free subscriber?
Did they become a paid subscriber?
Did they remain subscribed after 30, 90, or 365 days?
Did they return to the archive or attend another Live?

A Note that receives 20,000 likes but produces almost no subscribers may be less valuable to the Substack ecosystem than one that receives 300 likes, generates 75 free subscribers, creates 12 paid subscribers, and retains ten of them a year later.

To create the media business all of us are talking about, first, we need platform activity, interest from a growing audience, and a real distribution mechanism underpinning our content.

Notes is customer-acquisition work when it introduces a creator to relevant readers and turns some of that attention into relationships the creator can retain.

It becomes uncompensated platform labor when it merely keeps the feed active, improves Substack’s engagement numbers, and leaves the creator with little more than another vanity metric inside the platform.

This is so critically important. If you expect the Substack platform to deliver new free subscribers, much less paid subscribers, to your publication each week, then all of those users need to have a feed they can’t wait to get back to as their reason for opening the application and potentially discovering your work.

Creators should demand better attribution, not less discovery.

They should be able to see which Notes generate profile visits, which follows become subscriptions, which Lives create paid conversions, which topics introduce them to new interest communities, and which acquisition channels produce durable readers rather than rapid churn.

The quality of the analytics is, in my opinion, maybe the biggest opportunity the Substack team has to increase value for their creators, keep them on their platform long term, and continue to attract the new and best creator names on the planet. Substack needs to show their creators how to better monetize their publications from every possible angle in their sales funnel.

Do not ask Substack to remain small

Substack was never going to remain a quiet neighborhood of early writers, unusually high conversion rates, and limited competition.

Nor should it.

Creators should not ask Substack to freeze the advantages of its earliest years.

They should ask it to become the best interest-based media network in the world.

That means a platform where an unknown writer can reach a reader because the work is relevant, not because the writer is already famous.

It means using Notes, Lives, recommendations, video, and following to create subscriptions rather than replace them.

It means treating follower count as a clue, not a governing currency.

And it means converting algorithmic discovery into direct, durable relationships that creators can actually build businesses around.

Substack becoming a giant social platform is not the threat.

Becoming the wrong kind of giant social platform is.

The feed is not the enemy.

The wrong algorithm is.

- j -

Do you want some help building your newsletter or creator business?

Become a Paid Subscriber Today!

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of Operating by John Brewton and conducts advisory and consulting work for global companies, asking the question: What is the future of companies?

Appendix: Sources and evidence

The current creator debate

Substack’s scale and discovery network

The shift toward interest-based distribution

Evidence limitations

Substack has not publicly disclosed comprehensive platform-wide churn data, detailed cohort conversion rates, or the full ranking logic governing Notes and recommended posts. The latest official paid-subscription milestone I could verify is the five-million announcement from March 2025. There is also no public evidence resolving whether Substack’s recommendation system systematically favors or disadvantages paywalled posts.

The Value of “We Don’t Know” - Part One

John Brewton — Thu, 30 Jul 2026 11:17:12 GMT

My good friend Alex Pawlowski has written a warning to leaders and executives that personifies this expectation, and I am thrilled to be publishing it in two parts, this week and next.

Encoding your operation is still the right move. Encode the answers without encoding the questions, and you have built a company that reaches conclusions faster than it ever could.

I’m sincerely excited about these two pieces and am privileged to be publishing them via Operating by John Brewton.

Hope this helps, friends.

- j -

Subscribe to Alex’s Substack, The Strategy Stack, here.

I. The Disappearing Space Between Question and Answer

Every organization, regardless of its industry or ambition, exists because no individual possesses a complete understanding of the environment in which it operates. Markets shift without warning, technologies evolve faster than planning cycles, customers behave inconsistently, competitors conceal their intentions, and the internal dynamics of even the best-run companies rarely resemble the tidy diagrams found in annual reports. Organizations were never designed to eliminate uncertainty. They emerged because uncertainty exceeded the capacity of any single person to manage it alone.

Seen from that perspective, knowledge has always been something of a collective achievement. The engineer understands the architecture but not the market. The marketer recognizes customer behavior but not the underlying code. Legal counsel anticipates regulatory implications that product teams overlook, while finance evaluates risks invisible to both. What eventually appears as organizational intelligence is assembled from dozens, sometimes thousands, of incomplete perspectives that gradually converge into a decision. The organization becomes intelligent not because everyone knows the same thing, but because different people know different things.

This also explains why the phrase “I don’t know” occupies a far more important place in healthy organizations than most management literature acknowledges. Contrary to popular intuition, uncertainty is not the opposite of expertise. It is often one of its most reliable indicators. Experience rarely removes ambiguity. Instead, it sharpens the ability to recognize where established knowledge ends and where additional evidence becomes necessary. Senior physicians request further tests not because they lack competence, but because they understand how easily complex systems produce convincing yet incomplete explanations. Seasoned investors hesitate before declaring certainty because markets have repeatedly demonstrated how fragile apparently robust assumptions can become. The same pattern appears in engineering, intelligence analysis, scientific research and strategy. Expertise does not eliminate doubt; it changes the reasons for doubting.

That observation has quietly shaped the institutions through which modern societies produce reliable knowledge. Universities, peer review, investment committees, architecture reviews, clinical conferences and scientific replication all emerged from the same underlying recognition: individual judgment improves when it encounters competing perspectives before hardening into conviction. None of these institutions were created because people lacked intelligence. They were created because intelligence, left entirely on its own, has a remarkable talent for mistaking coherence for accuracy.

For a surprisingly long time, the practical mechanics of acquiring knowledge reinforced that discipline almost automatically. A difficult question rarely produced an immediate answer. It initiated a process. People searched libraries, consulted colleagues, revisited old notes, designed experiments, waited for data, challenged assumptions and occasionally abandoned conclusions that had seemed persuasive only days earlier. The interval between curiosity and explanation was often frustrating, yet it performed valuable intellectual work. It slowed judgment just enough for alternative interpretations to remain alive, making it less likely that the first plausible explanation would become the final one.

Generative AI compresses that interval with astonishing effectiveness. Questions that previously initiated investigation now frequently produce polished explanations within seconds. The technological achievement is extraordinary, but its organizational consequences deserve equal attention. Every coherent response shortens the period during which uncertainty remains visible. The question no longer lingers long enough to invite additional perspectives before an explanation arrives, and because that explanation is fluent, internally consistent and delivered with confidence, it often satisfies the immediate psychological need that originally motivated the question. Whether it fully deserves that confidence becomes a separate matter, one that many users never reach because the cognitive tension has already dissolved.

An intriguing study published recently offers an early glimpse of what this shift may mean. Researchers observed that participants who relied on generative AI became less willing to acknowledge when they did not know something. Confidence increased even when underlying understanding did not improve accordingly. Read in isolation, the finding resembles another entry in the growing literature on overconfidence and hallucinations. Read more carefully, however, it points toward a different possibility. The study is less about the reliability of AI than about the changing role of uncertainty in human reasoning. It suggests that one of the oldest transitions in the production of knowledge—the movement from not knowing toward understanding—may itself be acquiring a fundamentally different shape.

That possibility reaches far beyond individual cognition. Organizations learn because uncertainty travels. Questions move between departments, assumptions encounter disagreement, partial explanations are revised through conversation, and decisions gradually improve as competing interpretations collide with one another. Remove uncertainty too early from that process and something subtle begins to change. Conversations become shorter. Alternative hypotheses receive less attention. Confidence appears earlier than evidence. The organization continues to generate answers, but it may become progressively less effective at recognizing which questions still deserve to remain open.

The challenge, therefore, is not to decide whether artificial intelligence produces better explanations than humans. In many domains it already does. The more consequential question concerns the role that uncertainty has quietly played inside organizations all along. If explanation becomes abundant, inexpensive and permanently available, what happens to the intellectual habits that were originally built around its scarcity? Before exploring how AI may transform work, management or strategy, it is worth examining something more fundamental: the organizational function that uncertainty has performed for far longer than most of us have ever noticed.

II. Organizations Do Not Manage Knowledge. They Manage Distributed Uncertainty.

Traditional management literature often describes organizations as repositories of knowledge. The language is understandable. We speak about knowledge workers, knowledge management, knowledge transfer and organizational learning as though the defining characteristic of an enterprise were the amount of information it possesses. Yet this perspective places the emphasis on the outcome rather than on the process that produces it.

The more closely one observes how organizations actually function, the less convincing this description becomes. Companies rarely fail because the relevant information exists nowhere inside the business. More often, the necessary pieces are already present, scattered across departments, embedded in individual experience or hidden within operational routines that have never been connected into a coherent picture. Product teams understand technical constraints that commercial teams rarely encounter. Legal departments anticipate regulatory implications long before they become strategic concerns. Customer support notices weak signals months before they appear in market research, while finance often recognizes structural patterns that remain invisible to both. Every function operates with a legitimate yet necessarily incomplete representation of reality.

The organization exists because none of those representations is sufficient on its own.

That observation changes the way we think about expertise. Expertise is frequently imagined as the progressive elimination of uncertainty through accumulated experience. Organizational life suggests something rather different. As people become more experienced, they certainly develop deeper knowledge within their own domains, but they also become increasingly aware of the boundaries surrounding that knowledge. The experienced engineer rarely assumes that technical elegance guarantees commercial success. The experienced strategist understands that market data cannot predict technological discontinuities with confidence. The experienced executive eventually learns that every important decision contains variables no dashboard is capable of measuring directly.

The consequence is that organizations spend remarkably little time managing certainty. Their daily work consists largely of coordinating uncertainty between specialists whose perspectives overlap only partially. Meetings, design reviews, investment committees, cross-functional workshops and executive discussions all serve the same underlying purpose. They are mechanisms through which incomplete models of reality encounter one another before irreversible decisions are made. The value of these conversations lies less in exchanging facts than in exposing assumptions that remained invisible within individual disciplines.

This also explains why disagreement often becomes a productive rather than destructive force inside high-performing organizations. Two departments rarely disagree because one possesses knowledge and the other lacks it. They disagree because each observes reality through a different set of constraints. Engineering optimizes feasibility. Sales optimizes customer opportunity. Finance optimizes capital allocation. Legal optimizes compliance. Strategy attempts to reconcile those competing perspectives into a coherent direction. What appears as conflict on the surface frequently represents something far more valuable beneath it: the organization discovering where its own understanding remains incomplete.

Much of organizational learning therefore occurs before anyone learns anything new. It begins when existing explanations encounter friction. A marketing assumption collides with operational data. A customer interview contradicts internal expectations. A prototype behaves differently than predicted. A financial model no longer reflects observed market behavior. Progress starts at precisely the moment when confidence becomes difficult to maintain. The objective is not to eliminate disagreement as quickly as possible but to understand what reality is attempting to reveal through it.

Viewed from this perspective, uncertainty begins to resemble an organizational resource rather than an organizational defect. It identifies where current models no longer describe the environment with sufficient accuracy. It highlights where additional observation, experimentation or conversation is likely to generate the greatest returns. Organizations capable of recognizing those moments early often appear unusually adaptive, not because they possess superior forecasting abilities, but because they remain willing to revise their understanding before circumstances force them to do so.

Generative AI enters this landscape in a fascinating and somewhat paradoxical way. On the one hand, it increases access to information on a scale that would have seemed extraordinary only a few years ago. Individuals can summarize research, compare legal frameworks, draft technical documentation or explore unfamiliar domains within minutes. In many respects, AI lowers the cost of acquiring explicit knowledge more dramatically than any previous technology.

At the same time, lowering the cost of explanation also changes the signals flowing through the organization. If every uncertainty immediately produces a coherent narrative, the distinction between possessing an explanation and understanding the problem becomes progressively more difficult to detect. Questions that previously travelled through multiple perspectives may now terminate after the first plausible answer. Cross-functional conversations become shorter because apparent certainty arrives earlier. Teams may still collaborate, yet they increasingly collaborate around explanations that have not undergone the same process of collective scrutiny that organizations historically relied upon to improve decision quality.

The strategic challenge therefore extends well beyond accuracy. Even if future models become dramatically more reliable than those available today, organizations will still face the question of how uncertainty should move through the enterprise before action is taken. That movement has always been the hidden architecture of organizational learning. Decisions improve because uncertainty circulates between different perspectives, gathering evidence, exposing contradictions and forcing assumptions to justify themselves against reality. If that circulation gradually slows because explanation becomes abundant, organizations may discover that they have optimized access to knowledge while quietly weakening one of the mechanisms through which knowledge becomes trustworthy in the first place.

The discussion surrounding artificial intelligence often assumes that better answers naturally produce better organizations. History offers a more nuanced picture. Enduring institutions have rarely distinguished themselves through the speed with which they arrived at conclusions. Their advantage came from designing environments in which conclusions had to survive repeated encounters with alternative perspectives before they were allowed to shape consequential decisions. What deserves protection, therefore, is not uncertainty for its own sake, but the organizational processes that transform uncertainty into judgment rather than replacing it prematurely with confidence.

III. “I Don’t Know” as Organizational Infrastructure

When people speak about organizational capabilities, they usually think of engineering excellence, operational efficiency, strong leadership or an innovative culture. Rarely does anyone mention uncertainty. Yet many of the practices that distinguish mature organizations exist for one surprisingly simple reason: they create environments where uncertainty can be expressed before decisions become irreversible.

Consider how many institutional mechanisms depend on someone being willing to acknowledge that the current explanation may not be complete.

These practices were not invented because organizations distrusted intelligence. They were invented because intelligent people repeatedly arrive at convincing conclusions that later prove incomplete. The stronger the expertise, the more sophisticated those conclusions often become, which makes them even harder to challenge.

The phrase “I don’t know” therefore performs a function that extends far beyond individual humility. It becomes an organizational signal. It tells everyone else that the current model of reality has reached its limits and that additional perspectives are now valuable.

Interestingly, the most experienced professionals often become increasingly comfortable with this signal. Not because they know less than their junior colleagues, but because experience changes their relationship with confidence. After enough product launches, acquisitions, clinical cases or market cycles, certainty begins to feel less like expertise and more like a hypothesis waiting to encounter reality.

This observation creates an intriguing inversion of how expertise is commonly perceived.

None of these characteristics imply indecision. Organizations cannot function if every question remains permanently open. Decisions eventually have to be made, resources committed and strategies executed. The distinction lies elsewhere. High-performing organizations separate the exploration of uncertainty from the execution of decisions. They encourage broad disagreement while understanding the problem, and disciplined alignment once a direction has been chosen.

That separation has quietly become one of the defining characteristics of modern organizational life. Aviation investigates every incident regardless of how minor it appears because small anomalies often reveal systemic weaknesses. Hospitals conduct morbidity and mortality conferences not to assign blame but to examine where existing understanding failed. Intelligence agencies invest heavily in structured analytic techniques precisely because history demonstrates how easily coherent narratives become accepted without sufficient evidence. Across these very different domains, one pattern repeats itself: uncertainty is treated as something to be examined before it becomes something to regret.

Perhaps this explains why organizations so often struggle after periods of exceptional success. Success has a remarkable ability to reduce the perceived need for uncertainty. Strategies that repeatedly work begin to feel universally applicable. Processes that solved yesterday’s problems gradually become assumptions about tomorrow. Confidence accumulates not because reality has become simpler, but because past experience begins to substitute for fresh observation. By the time external conditions finally challenge those assumptions, organizations frequently discover that the habit of questioning them has quietly disappeared.

This is where the conversation returns to generative AI, although from a different direction than most current discussions. The central concern is not whether a model occasionally hallucinates or whether a summary omits important context. Organizations have always possessed imperfect information. Their resilience depended less on avoiding mistakes than on maintaining mechanisms capable of discovering them before they spread. The more interesting question is whether systems designed to produce immediate explanations unintentionally weaken one of those mechanisms by making uncertainty less visible inside everyday work.

John and I have spent much of the past year exploring organizations from different vantage points. One of us has focused primarily on strategy, organizational capability and judgment; the other on agentic systems, orchestration and the emerging architecture of AI-native work. Those conversations repeatedly converged on the same observation: the future of enterprise AI may depend less on how well systems answer questions than on how well they preserve the conditions under which good questions continue to emerge.

That possibility leads naturally to a broader question. If uncertainty has always functioned as an invisible form of organizational infrastructure, what happens when explanation becomes effectively abundant? What changes inside the enterprise when every employee, every team and eventually every autonomous agent can generate convincing narratives on demand? And perhaps more importantly, what kinds of systems should organizations build if their objective is not merely to generate explanations, but to preserve the quality of organizational judgment itself?

Tune in next week for the concluding sections of the essay…

Subscribe to Alex’s Substack, The Strategy Stack, here.

The 5 Operating Habits That Built a 75,000-Person Audience Around My Passion For Operating Companies in 2.5 Years

John Brewton — Wed, 29 Jul 2026 11:53:55 GMT

75,000 followers. 6,000 Substack subscribers. 200 paid. A Substack bestseller badge. And a new +$200,000 business vertical.

Over the last 2.5 years, I built an online audience of 75,000, a Substack with 6,000 subscribers, a Substack bestseller badge, + 200 paid members. Five essential operating habits created these outcomes.

Here’s how I learned and adopted them.

Habit 1: Start, then never stop

2.5 years ago I committed to posting on LinkedIn every single day.

I have not stopped since.

Some days the content was not ready. I posted. Some days I woke up late. I posted. In some weeks, the prior 7 posts underperformed. I kept posting.

The commitment has 2 parts: Start, and never stop. Most people manage to begin step one, but struggle to realize step 2.

What you can do: Pick 1 platform. Post daily for 90 days before you judge any of it. Do not negotiate or mentally manipulate your way around the committment.

Habit 2: Audit every post and every week

Every post gets reviewed. Every week’s batch gets reviewed. Auditing all of your work is essential and falls across three categories within the feedback loop:

Topic: Check whether the subject is something your audience wants from you and whether it aligns with your profile. Note whether it rode a broader conversation or was specific to you. Note who engaged early and propelled reach inside the algorithm.
Packaging: Carousel, infographic, hero image, video, or photo. Check the design: the fonts, the readability, how the post rendered in a mobile feed and on desktop. Log which formats won and which lost in a Google sheet and keep pushing forward.
Copy: My hooks run 8 to 12 words across 2 sentences. Check whether the hook stopped the scroll. Check the CTA length. Check whether the CTA matched the body of the post and moved the reader to the action you wanted.

What you can do: Build a log. 1 row per post, 3 columns: topic, packaging, copy. Review it every Sunday. The log will tell you what to make next before your instincts do.

Habit 3: Stop caring what people think

Criticism arrived early, and it never stopped. Haters are going to hate. Laughers are going to laugh. The people not doing the work, who wish they were doing what you’re doing, are the loudest.

None of it touched the numbers.

In 2.5 years, no critic has cost me a follower, a subscriber, a single piece of my happiness.

What you can do: Publish anyway, every day. Read comments for signal about where the value is in your work, not for opinions about you.

Habit 4: Create for the audience, not the ego

Curiosity is a habit: Reading constantly, studying what the audience responds to (and doesn’t), and constantly asking what would help them next. The audience tells you what it wants in the comments, in DMs, in which posts get saved and shared.

Teach them what they’re asking for, based on their feedback on why they show up each day for your work, not on what your ego imagines makes you look the smartest or appear most professionally or intellectually attractive.

The filter on every piece is whether the audience will find it valuable, not whether it makes me sound smart or look cool.

Those 2 filters produce different content.

They produce content that actually grows a business.

What you can do: Keep a running list of audience questions. Turn each one into a post.

Habit 5: Upgrade the production line, then reinvest the time

When I started, a LinkedIn carousel took 2.5 to 3 hours to build in Canva. Today, with Claude handling the design, the same carousel takes under 20 minutes once the concept, idea, and copy are set. Editing takes longer now than it did then, and it matters more, but that has also become my favorite part.

Every time a new model or new tool ships, I test it against my production process. If it makes the work faster or better, it goes in the stack.

The saved time goes back into the work. More posts, richer resources, more value shipped to the audience. Not shorter workdays.

What you can do: audit your production process once a quarter. Test 1 new tool each cycle. Spend every saved hour on output.

The system, in 5 lines + 1

Post every day.
Audit every week.
Ignore the noise.
Serve the audience.
Upgrade the tools and reinvest the time.
Start today.

The Operating Habits Workbook

0 to 75,000: The Operating Habits Workbook expands the 5 habits into 38 pages: 6 parts, 24 lessons, and 21 exercises you can complete at your desk, in a notebook, or with your favorite LLM. The 6th lesson covers material outside the habits and focuses on building the business while you build the audience.

The lessons report covers what happened in my data across the 912-day run.
The exercises turn each lesson into an artifact you produce this week: a signed posting contract, a 90-box streak tracker, the 8-column data tracker, a Sunday review ritual, a critic ledger, a chat pipeline, a production map, an offer sheet, and a conversion path map.
It closes with the 90-day execution plan and the Day 90 benchmark tables, which show whether it is working.

Commit to one part per week for the next six weeks. By Week 6, you will have a running streak, tracked data on every post, a daily engagement block, a quarterly tool test, and an offer a stranger can buy. These are the same assets in tools that took me 2.5 years to build by trial and error. The workbook puts you at a profoundly better starting line than I was.

Become a paid subscriber to access below.

Become a Paid Subscriber Today!

Become an Operating Founder for $399, and receive a 60-minute review of your content, business, and strategy. Additionally, you will be invited to join my weekly Operating Founder Office Hours, where I answer any and all of your questions and share what I’m working on and learning from building my business.

Stop Arguing (or Caring) About AI Writing About Slop or Not Slop

John Brewton — Sun, 26 Jul 2026 16:20:58 GMT

Historically, every time a tool made ideas cheaper to share, the people who had mastered the older, slower way treated it as a threat to defend against rather than a shift to move with. The scribes who spent their lives copying manuscripts did not celebrate the printing press. They saw a craft and a livelihood going to zero. Plato warned that writing itself would ruin memory and hollow out real understanding.

That was around 2,400 years ago.

Civilization grew anyway.

We are running the same fear-based reflex again. Our social feeds and inboxes are full of people litigating whether AI writing counts, whether it is cheating, whether the craft is dying. The scribes and the press are at it again. And while this argument burns, the enabling machines that create distribution for our ideas, work, products, and services are moving faster than they have in 500 years.

Here’s our current reality: The cost of producing words just fell to zero, and the only things left worth owning are the idea you choose and the distance it travels.

Five thousand years of the same job

For 5,000 years, every writing tool did one thing successfully.

It made ideas cheaper to share.

Clay tablets held temple receipts, not poetry.

A medieval book took weeks of skilled labor and got chained to the desk because it was worth more than the room it sat in.

Then the printing press cut the price of a book by two-thirds within 50 years.

Each drop did the same work.

More people could read, and more people could write.

This cycle of payoffs compounded for centuries.

In 1820, 12 percent of the world could read. Today it is roughly 90 percent. Cheaper copies did not make people read less. This innovation turned reading from an elite privilege into a basic human skill.

Every one of those tools shared a blind spot. The personal computer and the laptop made producing and editing words faster, and none of it touched the oldest cost of all, the effort of composing the thought itself. You still had to sit down and produce the words. That is the cost ChatGPT and Claude just took to zero. These are the first tools that compose and reason back. They draft, they read, they research across sources in minutes instead of weeks.

What actually got scarce

When a cost falls to zero, value moves to whatever sits next to it and is still expensive.

Composition is now cheap. So look at what it sits between. Upstream is the choice of what is worth saying. Downstream is getting it to the right people and being believed once it lands.

Judgment and distribution.

Those are the survivors.

Everything the AI-writing debate is fighting over is the one output that now fails to be scarce. The question was never how fast you can write. Speed is solved. The new question we should be centering our collective, ambitious, entrepreneurial gaze on is what you choose to think about, and how far you can move it.

The gate is gone

Here is my take on the opportunity we should all be aligned behind, but very few actually are.

For most of the internet’s life, reach was a function of accumulation. You built a list, a following, a set of contacts, and your ideas reached roughly the people you had already collected. That is the social graph. If you were small or new, you stayed small. Distribution was gated by who you already knew.

That gate is gone. A recommendation model now reads the content itself, scores it against what each person actually engages with, and serves it to strangers who have never heard your name. Follower count stopped being the gate and the metric to chase. The quality of your idea and what you can use the new stack to produce, distribute and build with it is the new hurdle, the new gate.

This is not a forecast I’m making as some Substack author you’ve never heard of befroe. Meta told a federal court, in its own antitrust filing, that just 7 percent of the time people spend on Instagram, and 17 percent on Facebook, is spent on content from friends. Zuckerberg has said more than 50 percent of what people see on Instagram is now chosen by AI, not by who they follow. The man who built the social graph testified that Facebook is no longer the culture, and floated wiping everyone’s friend connections to reset the platform around interests. The social network is barely social. The algorithm runs the room.

Become a Paid Operating Subscriber

Social media is dead. AI-enabled, interest-based media is the replacement.

So yes, the interest graph removed the gate, but it did not remove the competition. Cheap distribution does not spread attention evenly. Anita Elberse at Harvard studied this directly. In one set of 8 million songs that each sold at least 1 copy, roughly a third sold exactly 1 copy. Attention still stacks on whatever earns it. A model serving your idea to strangers is a momentary shot, not a guarantee, by far.

So the constraint did not disappear. It has just moved to a new place in the supply chain. It used to sit on permission, on access, on the size of your list. Now it sits entirely on the quality and clarity of the idea and how well you shape it. That is a harder game to fake and a fairer one to win. For the first time, a small company with a sharper idea can out-reach an incumbent with a giant list, because the list no longer decides.

Voice is the accelerant

If the interest graph decides who gets heard, voice decides how fast you can feed it.

Voice is the most information-dense form of human communication, and for the first time it is programmable. Andreessen Horowitz calls it the first and primary way people will now interact with AI, and 22 percent of the most recent Y Combinator class were voice-agent companies. You speak, the machine drafts and researches, and the loop from raw thought to finished asset closes to near real time. The cost of running these models fell about 280 times in 18 months, so the constraint is no longer money or minutes. It is what you point them at.

The reading side moved the same way. Most Google searches now end with no click at all. People take the answer, not the source. The idea that the machine surfaces matters more than owning the page it came from.

What to do with this

The uncomfortable part is that none of this rewards the skill most knowledge workers spent 20 years building. Fast, clean production is now table stakes a machine clears for free. What it rewards instead is judgment, and judgment is harder to fake.

So we need to focus our operating energies on the part that pays.

First, spend your scarce hours choosing ideas, not producing words. Keep a running list of the few ideas you want to be known for. That choice is now the work that pays most.
Second, move at conversation speed. Talk your thinking out loud, let the model draft and research, then you cut and judge. Do not type what you can speak.
Third, publish into the interest graph on purpose. Turn each idea into something consumable and teachable, ship it on a schedule, watch what the algorithm carries, and make more of that.
Fourth, own a few ideas in public until your name is attached to them. Repetition around a tight set of ideas is how you get known for them. Do not scatter.
Fifth, convert reach into revenue. Distribution is not a vanity metric. The point is to become known, sell more product, and sell more service off the back of the ideas.

Let other people spend the decade arguing about whether how the idea was written or produced matters. Don’t want to give away the ending, but none of that is going to matter one bit. Pick 1 idea worth being known for and publish it every week for the next 52. The writing was never the asset. What you choose to think about, create from those ideas, and how far you can push those concepts into the world of products, serrvices, subscriptions, VALUE FOR YOUR AUDIENCE, is the only thing that matters.

Stop worrying about debates that don’t matter.

Stop letting irrelevant online voices shame you.

Get to work, Substack.

- j -

The Interest-Graph Operating System

I built the system that is the logical, operating approach for this set of ideas. And my paid subscribers get it. The Interest-Graph Operating System is a workbook you paste into your own model (anyone you like or prefer), which then turns into your facilitator and walks you through the four moves: choosing the ideas worth owning, composing them at conversation speed, publishing them into the interest graph, and turning reach into revenue.

You answer its questions in your own words. It compiles the answers into a system you can run this week: your idea ledger, your prompt pack, your publishing plan, and the four key numbers that tell you if it is working.

If you’re just starting to understand these ideas, this workbook is the place to begin building the AI-fluent machine to architect your new professional future and business ambitions.

DOWNLOAD BELOW AFTER THE PAYWALL, FOLLOWING THE APPENDIX AND FAQ SECTION.

Appendix: Sources

Opening and the history of writing tools

Plato, Phaedrus (Socrates on writing and the loss of memory), Perseus Digital Library, Tufts University. https://www.perseus.tufts.edu/hopper/text?doc=Perseus:text:1999.01.0174:text%3DPhaedrus:page%3D275
Jeremiah Dittmar, “Information Technology and Economic Change: The Impact of the Printing Press,” Quarterly Journal of Economics, 2011 (book prices fell by roughly two-thirds; cities that adopted the press grew about 60 percent faster). https://academic.oup.com/qje/article-abstract/126/3/1133/1855353
Jeremiah Dittmar, plain-language summary, CEPR / VoxEU. https://cepr.org/voxeu/columns/information-technology-and-economic-change-impact-printing-press
“Literacy,” Our World in Data (global literacy 12 percent in 1820 to roughly 90 percent today). https://ourworldindata.org/literacy

The distribution shift: social graph to interest graph

“Facebook is Dead; Long Live Meta,” Stratechery by Ben Thompson, 2025, citing Meta’s FTC antitrust filing (7 percent of Instagram time and 17 percent of Facebook time spent on content from friends). https://stratechery.com/2025/meta-earnings-meta-turns-the-dial-social-network-r-i-p/
Mark Zuckerberg, Meta Q1 2024 earnings call, reported by Benzinga (more than 50 percent of Instagram content is now AI-recommended). https://www.benzinga.com/news/24/04/38420928/meta-ceo-mark-zuckerberg-more-than-50-of-the-content-people-see-on-instagram-is-now-ai-recommended
“The AI behind unconnected content recommendations on Facebook and Instagram,” Meta AI, 2023 (recommendation of content from accounts users do not follow). https://ai.meta.com/blog/ai-unconnected-content-recommendations-facebook-instagram/
“Spinning off Instagram, the decline of friending, and other takeaways from Mark Zuckerberg at the FTC monopoly trial,” CNN Business, April 16, 2025. https://www.cnn.com/2025/04/16/tech/mark-zuckerberg-testimony-meta-ftc-trial/index.html
Anita Elberse, Blockbusters, Harvard Business School, via “The Way of the Blockbuster,” Harvard Magazine, 2013 (market concentration; a third of 8 million songs sold exactly 1 copy). https://www.harvardmagazine.com/2013/12/the-way-of-the-blockbuster

Voice and the reading side

“AI Voice Agents: 2025 Update,” Andreessen Horowitz (voice as the first and primary interface; 22 percent of the latest Y Combinator cohort). https://a16z.com/ai-voice-agents-2025-update/
“The 2025 AI Index Report,” Stanford Institute for Human-Centered AI (inference cost fell roughly 280x in about 18 months). https://hai.stanford.edu/ai-index/2025-ai-index-report
“Google zero-click searches study,” Search Engine Land, 2026 (most Google searches end without a click). https://searchengineland.com/google-zero-click-searches-2026-study-479717

Frequently Asked Questions

On “Stop Arguing About AI Writing” and the interest-graph shift

Isn’t most AI writing just slop? Some of it is. Slop is a production problem, and production is the part that got cheap. The point of the piece is that arguing about whether a machine drafts well is the wrong fight, because the value moved to the two things a machine does not do for you: choosing the idea and judging what comes back. Quality still matters. It just stopped being about who typed the words.

Does this mean writing skill no longer matters? Writing skill matters less as production and more as judgment. Fast, clean production is now table stakes a model clears for free. What pays is taste, the ability to choose what is worth saying and to cut what is not. That is a harder skill than typing, and it is the one worth building.

What is the interest graph, in plain terms? The social graph reaches the people you already follow or already know. The interest graph reaches strangers, because a recommendation model reads your content and serves it to whoever is likely to care. Follower count stops being the gate and the idea becomes the gate. Meta told a federal court that only 7 percent of time on Instagram, and 17 percent on Facebook, is now spent on content from friends.

I have a small audience. Can this actually work for me? Yes, and that is the opportunity. The interest graph does not ask how big you are. It asks how good the idea is. A sharper idea from an unknown account can out-reach an incumbent with a large list, because the list no longer decides who gets seen.

If everyone can produce, isn’t the competition infinite? The competition is real, and attention still concentrates on a few winners. Anita Elberse at Harvard showed that digital markets are more concentrated, not less. The gate moved. It did not vanish. Removing the permission barrier does not hand you the crowd. It gives you a shot you did not have before, and the idea is what earns it.

Isn’t building on someone else’s algorithm risky? It is a real risk, and it is the cost of the reach. Manage it two ways. Convert algorithmic reach into assets you own, an email list and paying customers, so a platform change cannot erase you. And do not depend on one platform. Publish where the idea travels and move with it.

Do I need to use voice and AI to do this? No, but they are the accelerant. Voice is the most information-dense way to get an idea out of your head, and the model turns it into a draft you judge and cut. The loop from raw thought to finished asset closes to minutes. You can do the work by typing. You will just be slower.

How do I make money from this, not just collect reach? Attach an offer to every idea you choose to own. Publish the argument free so it travels, and gate the system, the template, or the service that acts on it. Track four numbers every week: reach, email subscribers, paid subscribers, and revenue. If a week moves none of them, change the idea, not the effort.

How is this different from telling me to post more? Volume without a chosen idea trains the algorithm on nothing and trains the reader on nothing. This is the opposite. Own 3 to 5 ideas, repeat them in new forms until your name attaches to them, and let the schedule do the compounding. One flagship a week for 52 weeks beats 12 brilliant pieces shipped on mood.

How much time does the weekly system take? The cadence is one flagship and five short cuts a week. Monday you choose the idea, Tuesday you compose with the prompt pack, Wednesday you ship, and the rest of the week you distribute and log the numbers. The initial build takes about 40 minutes. The upkeep is a few focused hours a week.

What is the Interest-Graph Operating System, and how do I use it? It is a workbook you paste into your own model. The model becomes your facilitator, asks you one question at a time, and compiles your answers into a finished system: your idea ledger, your prompt pack, your publishing plan, and your scoreboard. You build it once, then you run it every week. Subscribers get it with this piece.

Do. Fail. Learn. Grow. Win. | My 5-Step (How-To) Framework To Succeed At Anything.

John Brewton — Tue, 21 Jul 2026 22:34:54 GMT

Work w/ John

TL →DR

I have 1 framework: Do. Fail. Learn. Grow. Win. It carried me through a PhD program, a family distribution business, a consulting firm, and a creator business built from an audience of 0 to over 50,000.
Research says learning from failure is not automatic. A 10-year study of 71 cardiac surgeons found they learned from their own successes and other surgeons’ failures, not their own failures.
Failure alone is nearly worthless. Venture-backed founders with a prior failure succeed 23 percent of the time. First-timers succeed 22 percent of the time. The value is the lesson, not the failure.
The extraction tool is 1 written sentence naming why the miss happened. In a Harvard field experiment, 15 minutes of written reflection beat additional working time by about 23 percent.
Run 1 loop per week: ship 1 thing by Friday, name the miss, change 1 behavior, ship again. Count wins at week 12, not week 1.

Who the heck is John Brewton, where did this framework come from, and why should you care?

16 years ago, I was working to complete a PhD in economics at the University of Chicago when a professor gave me the best advice of my career. Take a year off, he said. Go get some experience actually operating a business. My parents owned an industrial distribution company and were working toward selling it. He thought I should go help. I had grown up in that entrepreneurial household and talked about the company through my entire graduate education, so I went.

The 1 year became 12.

I learned to operate by operating. We sold the company in 2021, and starting a consulting firm on the back of those relationships was the next logical step. So that is what I did. For that entire period, I had no LinkedIn, no newsletter, no “brand.” I produced zero content. It never occurred to me. I was head down, building a business, thinking about nothing else in the world but how to use our operational efficiencies to grow margin and razor thin market.

Then, almost 4 years ago, my life changed at once. I met my wife. We fell in love, got married, and had our daughter. The weekend we found out we were pregnant, my mother passed away. I am an only child. I had spent 12 years building a business beside my parents and stayed very close through these, often difficult, years. The largest loss of my life arrived the same weekend as the news of my greatest happiness.

Life is extraordinarily not linear.

That was the year I stopped expecting it to be.

Or maybe it was just the moment I realized that it had never been.

I moved to Northern Virginia to become a husband, a stepfather, and a father. That November I decided the business had to change to match the life. I wanted to be available for all of it, every experience of being a father, of being a Girl Dad. So I took a page from Gary Vaynerchuk’s playbook and made 1 simple commitment: post on LinkedIn every day for a year, starting from an audience of 0. I completed the year and have posted every day since. Today the audience across LinkedIn and the newsletter is over 50,000. Operating by John Brewton became a Substack bestseller. And my total business operations now span across 6 revenue streams, consulting being primary, courses, a paid newsletter, coaching, advisory work, and some content design work for projects I find especially inspiring.

I had wanted a content business for years and never started one. I admired Gary Vee and Casey Neistat from a distance while my ego managed credentials, attaching names and institutions to my work so it would look smarter to the outside world than I sometimes felt on the inside. Becoming a father ended that negotiation. Meeting other people’s expectations was the last thing I cared about when that period of my life began.

The life I have today looks nothing like the one I planned. I work more than I ever have. I love the consulting. I learn so much from clients each and every week. Opportunities are plentiful. None of it matches the picture I once drew, and 1 thing explains how it all got built anyway.

The DNA of every win is found in some loss and some failure along the way. I have 1 framework. It is 5 words long, and it has run every chapter of this story.

Download the worksheet I’ve built to help you get started here:

What is the Do. Fail. Learn. Grow. Win. framework?

Do. Fail. Learn. Grow. Win.

When the cycle completes, it’s back to doing more.

That is the entire system.

No software, no certification, no 2-by-2 matrix.

One warning before the steps, because the research here is humbling. Learning from failure is not automatic. A 10-year study of 71 cardiac surgeons across 6,500 procedures found that surgeons improved after their own successes and after other surgeons’ failures, but not after their own failures. Their own misses got filed under bad luck. If surgeons with lives on the line do not learn from their own failures by default, neither do you and neither do I. It is imperative that the loop exists to force what does not happen on its own.

How do you run each step of the loop?

1. Do: Ship something real. A draft, a newsletter issue, a proposal, a post. The framework does not begin with planning. It begins with output. Most people invert this and spend the doing budget on preparing to do. Volume is the entry fee. My version was 1 LinkedIn post a day, every day, starting from an audience of 0.

2. Fail: Some of what you ship will not work. This is not a defect in the system. It is a component of it. The issue nobody opens. The proposal that dies in review. Do not route around this step. Treat each miss as data arriving on schedule.

3. Learn: Failure only pays if you collect. Name the mechanism of the miss in 1 sentence. Not “the post flopped” but “the post buried its one idea in paragraph 4.” A loss you cannot articulate is a loss you will repeat. This step is not a journaling habit. In a Harvard Business School field experiment, trainees who spent the last 15 minutes of the workday writing down what they learned performed about 23 percent better than trainees who spent those minutes doing more work. Reflection converts doing into learning. Reflective writing is an essential tool.

4. Grow: Growth is the lesson applied, not the lesson admired. Change the template. Rewrite the opening. Reprice the offer. If your next attempt is identical to your last, no learning happened.

5. Win: The win usually arrives smaller than expected and later than planned. Take it anyway. A win is not a finish line. It is funding for the next round of doing. Bank it, then return to step 1 with more ambition.

Does failure actually make you more successful?

No. This is the uncomfortable part, and the data is blunt. A Harvard study of venture-backed founders found that previously successful founders succeeded 34 percent of the time on their next company. First-timers succeeded 22 percent of the time. Previously failed founders, 23 percent. Failure by itself bought 1 point. The failing was never the value. The lesson you extract from it is, and most people skip the extraction.

That is why the framework has 5 steps and not 2. Do and Fail happen to everyone who tries. Learn, Grow, and Win only happen to the people who commit to running the full loop.

How do you run multiple loops at once?

The first version to master is to simply run 1 loop at a time. The magic happens when you’re running the loop against is 6 revenue streams, 6 competing client interests, all at 6 different stages of the cycle on any given day. When that many loops run at once, no single failure can take you out of the game.

The framework survives every context switch because it assumes nothing about the domain. It assumes only a sequence.

Action produces information.
Information produces adjustment.
Adjustment produces results.

That sequence held for me at Harvard, as a PhD candidate, in a messy distribution business, across all consulting engagements and right in the glow of an analytics dashboard for my newsletter business.

I will never know the outcome or the final destination. Neither will you. The job is to keep executing, put one foot in front of the other, and build the life from whatever luck arrives and whatever realities show up. That is the nature of the work for every consultant, creator, founder, and operator trying to figure this out.

How do you start the framework this week?

Start small and start now.

Pick 1 project. Ship 1 real unit of work by Friday: a post, a proposal, a prototype, whatever the project calls for. When part of it misses, and part of it will, write 1 sentence identifying the reason. Change 1 behavior before the next attempt. Then ship again next week.

Run this framework for 12 weeks and count the wins at the end.

The best state you can be in is a continuous state of learning from what has not worked. That is where the growing comes from.

That is where all your future wins live.

Do. Fail. Learn. Grow. Win.

That is my framework, and I am sticking to it.

You should try, too.

I so sincerely hope this helps.

- j -

Work w/ John

I run this operating system inside real businesses: consulting engagements, advisory work, and coaching for founders, operators, and creators. If you want to build your version of the loop for your work or company, message me and tell me what you are building.

What else do people ask about the framework?

Is “fail fast” good advice? Not by itself. The founder data shows failure alone improves your odds by 1 point. Fail small, on purpose, with a hypothesis, then write the lesson down. Extraction is the value, speed is not.

What if nothing failed this week? Something did. “Nothing missed” is almost always a collection failure, not a clean week. Look at what underperformed, what ran late, and what you avoided. Name 1 mechanism anyway. A dry week with a named lesson is progress.

How long before the loop produces wins? Run it for 12 weeks before judging it. Weeks 1 to 4 produce volume and misses. Weeks 5 to 8 produce lessons and behavior changes. Weeks 9 to 12 produce the wins the first 8 weeks paid for.

Does this work outside of business? Yes. The framework assumes nothing about the domain, only a sequence: action produces information, information produces adjustment, adjustment produces results. School, sports, music, parenting, work. Same loop.

Where do I get the free field guide? The Loop Field Guide is free for subscribers. It is 2 pages: the framework on 1 page and the 5-blank weekly sheet on the other. Fill the sheet once a week in 5 minutes. Subscribe and it lands in your inbox.

Where does the research behind the loop come from?

Appendix. Every study referenced in this article, organized by the step it supports.

Arrow, K.J., “The Economic Implications of Learning by Doing,” The Review of Economic Studies (1962). The founding economics of experience-driven improvement. https://academic.oup.com/restud/article-abstract/29/3/155/1539235
Thompson, P., “How Much Did the Liberty Shipbuilders Learn?” Journal of Political Economy (2001). Productivity rose with cumulative output, and gains decayed when the doing stopped. https://www.journals.uchicago.edu/doi/abs/10.1086/318605
Ericsson, Prietula, Cokely, “The Making of an Expert,” Harvard Business Review (2007). Repetition without feedback plateaus. Practice targeted at weaknesses improves. https://hbr.org/2007/07/the-making-of-an-expert

Fail

Kahneman and Tversky’s prospect theory, confirmed by a 2024 meta-analysis: losses weigh roughly 1.9x gains, which is why shipping feels risky and preparing feels safe. https://www.sciencedirect.com/science/article/pii/S0167487024000485
Golman, Hagmann, Loewenstein, “The Demand for, and Avoidance of, Information.” People dodge feedback that might sting, even when it is free and useful. https://www.cmu.edu/dietrich/sds/docs/golman/the-demand-for__and-avoidance-of__information.pdf
Edmondson, A., “Strategies for Learning from Failure,” Harvard Business Review (2011), and “It’s OK to Fail, but You Have to Do It Right” (2023). Only intelligent failure teaches: new territory, goal-driven, hypothesis-informed, small. https://hbr.org/2011/04/strategies-for-learning-from-failure · https://hbr.org/2023/07/its-ok-to-fail-but-you-have-to-do-it-right

Learn

KC, Staats, Gino, “Learning from My Success and from Others’ Failure,” Management Science (2013). 71 cardiac surgeons, 6,500 procedures, 10 years. No learning from own failures. https://pubsonline.informs.org/doi/10.1287/mnsc.2013.1720
Eskreis-Winkler and Fishbach, “Not Learning From Failure — the Greatest Failure of All,” Psychological Science (2019). 5 experiments, 1,600+ participants. Failure feedback taught less than success feedback because people tune out. They learned normally from others’ failures. https://journals.sagepub.com/doi/abs/10.1177/0956797619881133
Di Stefano, Gino, Pisano, Staats, “Learning by Thinking,” HBS working paper. 15 minutes of written end-of-day reflection beat additional working time by about 23 percent. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2414478
Finkelstein and Fishbach, “Tell Me What I Did Wrong,” Journal of Consumer Research (2012). Experts seek negative feedback. Novices avoid it. https://academic.oup.com/jcr/article-abstract/39/1/22/1823160

Grow and Win

Weick, K., “Small Wins: Redefining the Scale of Social Problems,” American Psychologist (1984). Small, winnable units are the reliable path to large change. https://psycnet.apa.org/record/1984-25682-001
Amabile and Kramer, “The Power of Small Wins,” Harvard Business Review (2011). 12,000 diary entries from 238 professionals. Progress in meaningful work is the top motivator. https://hbr.org/2011/05/the-power-of-small-wins
Edmondson, A., “Psychological Safety and Learning Behavior in Work Teams,” Administrative Science Quarterly (1999). Teams that surface errors outlearn teams that hide them. https://journals.sagepub.com/doi/10.2307/2666999

The market evidence

Gompers, Kovner, Lerner, Scharfstein, “Performance Persistence in Entrepreneurship,” Journal of Financial Economics (2010). Success rates: prior success 34 percent, first-timers 22, prior failure 23. https://www.nber.org/papers/w12592
Bessemer Venture Partners, the Anti-Portfolio. A top venture firm’s public log of the great companies it passed on, and why. The loop log at institutional scale. https://www.bvp.com/anti-portfolio

How to Produce World-Class Research with Perplexity Computer & Claude -> Fable 5

John Brewton — Sun, 19 Jul 2026 18:12:58 GMT

The full methodology and workflow behind a 22-page, 12-chart research report: Perplexity Computer for data collection and essential research, then Claude on Opus 4.8 and Fable 5 for verification and production, along with my elite research skill file that holds the whole thing to an institutional standard. The skill’s architecture is broken down below, and the workbooks for building your own are attached.

Too Long →Didn’t Read

Perplexity Computer produced a 22-page research report and a full data workbook in 33 minutes across roughly 140 agent tasks. It still contained 5 substantive errors.
The core doctrine: never publish research from inside the machine that produced it. Move the output to a second, different model and audit it against a written standard.
The audit runs on a skill file with 6 sections: a closed 3-tier source architecture, 4 research modes, search and paywall discipline, 3 output templates, voice rules, and 8 quality checks.
Charts get rebuilt from the raw data in your own visual identity. The final report is written in your own voice and read by you in full before anything ships.
Attached to this post: the full report this article walks through, plus the Your Research Employee workbooks in 4 editions (Claude, ChatGPT, Gemini, Perplexity) so you can build your own analyst in about an hour.

The report runs 22 pages.

It carries 12 original charts, 4 parts, 2 appendices, and a complete source register with a live URL behind every figure.

The machine research behind it took 33 minutes across roughly 140 tasks.

The verification pass that followed caught 5 errors before a single chart was drawn.

The complete process took my team of agents 65 minutes. The work would have taken me two weeks.

Anyone can get an AI research report in 2026. The tools are extraordinary and improving quarter over quarter. Which means the differentiation has moved. World-class research is no longer about access to information. It is about the discipline you wrap around the machines and agentic teams you deploy to make it happen.

This piece walks through the exact pipeline I used to produce “The Last Bubble Went Public. This One Is Staying Private,” my quantitative history comparing the dot-com bubble of 1995 to 2003 against the AI boom of 2022 to 2026. I recorded the whole build on Loom. This is the written version, plus the part the video could not show: the full anatomy of the elite research skill that runs quality control, and the workbooks for building a research employee of your own on Claude, ChatGPT, Gemini, or Perplexity.

Hope this helps you this week, friends!

- j -

If you want to become AI-fluent this summer and learn how to build a team of agents to do the work discussed in this article and so much more, you should seriously consider joining my Oper(AI)tors Summer of AI Fluency cohort.

Grab Your Seat Now!

Use code TAKE500 to receive a $500 discount on your seat.

What is the best AI stack for producing research?

Perplexity Computer collects. Claude verifies and produces the output in my primary stack. Skills encode the standards so neither step depends on my memory or my mood.

Three tools, three jobs, no overlap.

The research ran in Perplexity’s Comet browser (download today!), which is my daily browser on every machine I own. The verification, chart production, and final write-up ran in Claude Cowork sessions on Opus 4.8 and the Fable 5 model. Dictation ran through Wispr Flow, which is why I can commission a research project at the speed of speech.

Step 1: How do you brief an AI research agent like Perplexity Computer?

I opened Perplexity Computer and dictated a brief. Not a query. A brief. The distinction decides the quality of everything downstream.

Watch the Loom here.

The brief had 4 components:

The history: Deep research on the dot-com bubble and crash, from the late 1990s through the early 2000s. I want to understand the period, not skim it.
The comparison: Set that period against the current AI cycle, quantitatively and qualitatively. Startups then against startups now. Trajectories then against trajectories now.
The data demand: Come back with financial data and historical data sufficient to build charts, graphs, and visualizations. This clause is the one most people skip. If you do not ask for chartable series, you get prose, and prose cannot be re-plotted in your own brand.
The standard: Exhaustive, attributable, accurate. Every claim is sourced.

Then I left my computer for a bit, and Perplexity deployed a team of agents that worked for 33 minutes across roughly 140 discrete tasks. Watching the full run is instructive, partly because Perplexity shows which models it selects for which activities. But you do not need to watch. You need the 2 deliverables it returns: a complete quantitative and qualitative report, and an Excel workbook holding every underlying data series as chartable tables.

View the Perplexity Excel data file here.

That workbook is the asset. The report is a draft. Hold that hierarchy in your head, because the next step depends on it.

The finished report. 22 pages, 12 charts, 4 parts, every correction stated in Appendix A, the complete source register in Appendix B.

Step 2: How do you verify AI research for accuracy?

Here is the doctrine at the center of this whole system. Never publish research from inside the machine that produced it.

Perplexity Computer is a serious research engine. The output is attributable, the writing is clean, the coverage is broad. And it still shipped me 5 errors. Not typos. Substantive, number-level errors of the kind that end up in someone’s investor memo and then in someone’s screenshot.

So the first thing I do with any Perplexity research package is take it out of Perplexity. I downloaded both files, opened a Claude Cowork session, attached them, and ran my elite research skill against the full package with one instruction: check that everything here is accurate, and show me what is not.

The audit came back (read the full audit MD file here) with a list.

The 5 that mattered most

Cisco’s crash was understated by 22 percentage points: The figure that circulates for Cisco’s dot-com decline is 66 percent. That number measures only through December 2000, which is 22 months before the actual bottom. The true peak-to-trough is 88 percent, from March 2000 to October 2002, from $79 to $9.50. The distinction changes the lesson. Being right about the technology protected Cisco’s business. It did not protect the price.
Anthropic’s valuation ladder was mislabeled: The verified sequence reads $61.5 billion, $183 billion, roughly $350 billion, then $965 billion, with the $965 billion Series H dated May 28, 2026. Versions dating that figure to November 2025 conflate it with the Microsoft-Nvidia deal, valued at roughly $350 billion. A 15x repricing in 14 months deserves correct dates.
The AI venture funding series stopped a year early: The draft ended the series at 2024, at $114 billion. The full record runs through 2025 at $202.3 billion, close to half of all venture capital deployed worldwide that year. Comparisons ending at 2024 understate the cycle. Flattering the past is the most common failure mode in bubble analysis, and it usually arrives disguised as a data cutoff.
The marquee AI IPO count was stale: The correct count is 2. Cerebras was listed on the Nasdaq on May 14, 2026, joining CoreWeave’s March 2025 listing.
The NASDAQ year-end close needed pinning: The verified 2025 monthly series puts it at 23,241.99.

I read the list, approved all 5 corrections, and instructed that all work going forward be updated against them. Every downstream chart and paragraph automatically inherited the corrected numbers. The final report states each correction openly in its methodology appendix, which is what a reader deserves and almost no AI-assisted research bothers to do.

The general principle: collect everything in one place, then verify with a different model, ideally a stronger one than the one that did the collection. Two models will not share the same blind spots. One model grading its own homework will.

This is the first layer of checking, not the last. When the full report is assembled, I read the entire thing myself. The machine audit removes the errors a human skims past. The human reader catches what no rule can encode. You need both, in that order.

Step 3: How do you rebrand AI-generated charts in your own visual identity?

The research you publish should look like yours. Perplexity’s charts are competent and generic. So the next instruction to Claude: take my personal brand skill, take every data series from the workbook, and re-create every chart plus any additional visualizations the data supports.

Claude returned 12 charts. Each one is rendered in my visual identity, the Forest Ink field register, as an individual PNG, plus a single PDF holding all 12, plus an HTML index page for reviewing the full set in one scroll. NASDAQ’s 6.2x five-year multiple. The casualty table shows that the survivors still fell to 88 percent. The private-marks ladder no public index can show you. The 280-versus-2 IPO chart encapsulates the entire thesis in a single image.

View all twelve charts here.

The proportion and consistency across the set is the point. A reader cannot articulate why a report feels institutional. Uniform chart grammar is most of the answer.

Two practical notes. First, this only works because the data came back as chartable tables in Step 1. Second, keep the individual PNGs. You will reuse single charts in newsletters, decks, and posts for months. The report is one container for the charts, not the only one.

Step 4: How do you turn AI research into a publishable report?

Final instruction: produce the complete research report, including the charts, written in the Operating voice. Clearly written, professionally written, minimal emotion. Tell the story, walk the reader through the numbers, let the two periods of history do the arguing. Return it as a branded PDF.

It came back as a 22-page document: executive summary, 4 parts, a methodology appendix listing the corrections, and a full source register. Some charts embedded in the first PDF pass imperfectly. This does not bother me. I had Claude produce a Markdown version I could format and embed cleanly myself. Expect this. The last 5 percent of production is still manual, and pretending otherwise is how polished-looking errors ship.

Then the human read. The whole report, my eyes, before anything publishes. The machines compress 2 weeks of analyst work into an afternoon. They do not remove the afternoon.

What goes inside an elite research skill for Claude?

Everything above depends on one file. The elite research skill is a document I have built and tweaked over roughly 6 months, and it is the difference between asking Claude to “check this research” and running an institutional-grade audit with repeatable standards. A skill, mechanically, is just a structured instruction file the model loads when the task calls for it. The leverage is in what the file encodes.

Mine has 6 sections. Here is what each one does and why it exists.

1. What sources should an AI research skill be allowed to cite?

The skill’s foundation is a closed list of approved sources, organized in 3 tiers. Tier 1 is elite publications: the New York Times, the Financial Times, the Wall Street Journal, the Economist, Harvard Business Review, MIT Technology Review, and MIT Sloan Management Review, each listed with its domain and its specific strength. Tier 2 comprises academic research centers: Harvard, Stanford, Princeton, Yale, Chicago Booth, and LSE, each with named search targets, such as specific working paper series. Tier 3 is institutional investment research: Goldman Sachs, Morgan Stanley, Bridgewater, Blackstone, AQR, BlackRock, JP Morgan, plus the top VC firms, a16z, Sequoia, Benchmark, Bessemer, and their peers, with a rule to pull published essays and market maps rather than portfolio announcements.

The operative word is exclusively. The skill instructs the model to cite nothing outside the architecture unless I explicitly ask. This single constraint kills the SEO-farm citation problem before it starts. Most AI research quality problems are source-selection problems wearing a disguise.

Every finding then carries 1 of 4 credibility tags: [PEER-REVIEWED] for Tier 2 academic work, [INSTITUTIONAL] for Tier 3 research, [EDITORIAL] for Tier 1 reporting, and [OPINION] for op-eds and commentary, which must still come from approved sources. I can scan a brief and weigh every claim by the tag beside it.

2. How does the skill match research depth to the task?

The skill defines 4 modes, each with trigger phrases so the model selects the right one from how I naturally talk, each with its own numbered method.

Thesis Research activates on phrases like “I’m writing about” or “I want to argue that.” Its method: restate the thesis in 1 sentence, find supporting evidence, then deliberately hunt complicating evidence, then historical parallels, then 2 to 3 quotable data points, then any disagreements between sources. The output emphasis is narrative ammunition for an article.

Market Intelligence activates on client names and phrases like “competitive landscape.” It works through recent Tier 1 coverage, institutional research, relevant academic work, then risks and opportunities not yet in mainstream coverage. Output emphasis: decision-ready intelligence for advisory work.

Trend Scanning activates on “what are the smartest people saying about” and similar. It restricts to the last 30 to 90 days, then maps consensus, divergence, and the narrative that is forming but has not crystallized. Output emphasis: compressed, newsletter-ready signal.

Deep Dive activates on “comprehensive research” or “everything on.” Broad search, findings organized by sub-theme, a timeline where the topic has one, the key debates mapped, the gaps named, and a meta-narrative synthesized at the end. This is the 10-plus-source mode, and it is the one the dot-com project ran in.

The modes matter because “research this” is not one task. It is at least 4 tasks with different depths, different recency requirements, and different outputs. Naming them lets the file route the work.

3. How should AI research handle search queries and paywalled sources?

The skill specifies exact search patterns, site: operators per source, working-paper queries for the academic tier, and firm-name patterns for institutional research. It sets a floor of 5 searches per request, more for Deep Dive. It mandates recency within 12 months unless the topic is historical.

It also contains the section I consider non-negotiable: paywall discipline. Most Tier 1 sources are paywalled. The skill instructs the model to attempt the full fetch, and when blocked, to extract only what is actually visible, provide the direct URL for me to open through my subscriptions, and tag the finding [PAYWALLED — link provided]. The controlling rule is on one line: never assume content is behind a paywall. Only report what you can actually see. Fabricated paraphrases of paywalled articles are among the most common and least detectable AI research failures. This section is the immune response.

4. What output formats should AI research produce?

Three fully specified templates, chosen by destination. Format A, the Research Brief, feeds articles: thesis, key findings with tags and full citations, complicating evidence, historical parallels, quotable data points, source disagreements, narrative hooks, and a list of paywalled sources to review. Format B, the Advisory Intelligence Memo, feeds client work: executive summary, market context, a key-intelligence table where every row pairs a finding with its source, credibility tag, and a “so what” implication, then competitive landscape, risks and blind spots, and recommended discussion points for the client call. Format C, the Signal Pack, feeds the newsletter: the signal in 1 paragraph, top insights at 2 sentences each, where the smart money disagrees, the emerging narrative, and a prioritized reading list.

Templates sound bureaucratic until you run without them. Unstructured research output is a wall of prose you excavate. Structured output is a component you install.

5. Why should AI research output never hedge?

The skill enforces the Operating voice on the research output itself. Declarative sentences, no hedging, active voice only, findings stated as facts. “Goldman’s research shows X,” never “Goldman’s research seems to suggest X.” A banned-word list that includes “perhaps,” “arguably,” and “it seems.” The reasoning: a research brief is not a neutral academic document. It is intelligence, written with an operator’s eye for what matters. Hedged findings defer the judgment work to future me, and future me is busy.

6. How do you quality-check AI research before delivering it?

The skill closes with 8 verification items the model must clear before delivering anything: every finding cites a specific approved source, every finding carries a tag, nothing off-architecture slipped in, paywalled items are flagged with URLs, the format matches the destination, the voice complies, source disagreements are surfaced rather than smoothed over, and the output is actionable rather than merely informational.

That 7th check deserves a sentence. Models are consensus machines. Left alone, they average conflicting sources into a smooth, confident, wrong middle. Forcing disagreements to the surface is where the most interesting angles live. Goldman versus Bridgewater is an article. Their average is nothing.

The skill also declares its integrations, it feeds my meeting-intelligence pipeline, hands off to my brand system for visuals, and supplies the source architecture when my contrarian filter needs counter-evidence. Skills compound when they reference each other. A lone skill is a tool. A connected set is a staff.

How do you build your own AI research employee?

You do not need 6 months to get a working version of this. With my Oper(AI)te co-instructor, Wessal Khader, I have built the whole construction process into a set of workbooks called Your Research Employee, attached to this post in 4 editions:

Each workbook is a file that an AI can run. You upload it, paste 1 kickoff prompt, and the AI walks you through the build as a guided interview, 1 question at a time. No fields to fill in by hand. The build takes about an hour across 4 steps:

The Research Standard, about 20 minutes: Nine questions that force specificity: the decisions your research actually feeds, the output shape you need, when a fast answer suffices versus a deep dig, which sources you trust on sight, which are noise to you, your domain, and whether you want a recommendation or evidence laid out for your own call. The test of a good standard is that it names a real decision. A brief with no decision behind it is a Wikipedia page.
The Source Rules, about 15 minutes: The side nobody teaches. The workbook ships with a measured floor built from published research on how AI research fails, including the Columbia Tow Center finding that 8 AI search engines got the source wrong more than 60 percent of the time across 1,600 queries. On top of that floor sit 9 banned research habits, uncited claims, single-sourcing, stale data stated as current, marketing repeated as fact, and 5 more, each paired with the positive rule that replaces it.
Sources and Assembly, about 15 minutes: Your trusted-source list, your internal documents, and your question workflow get tested and assembled into 1 paste-ready analyst profile: standard, rules, sources, mode, output format.
The Source Audit, about 10 minutes: A repeatable protocol for scoring any brief against a fixed table. 0 to 1 violations are decision-grade. 2 to 4 means fix the claims and update your rules tonight. 5 or more is a guess, wearing a citation. Each workbook closes with a swipe file of 6 prompts for daily operation, including my favorite, the test drive: make the analyst research a question you already know the answer to, then grade it against reality.

The 4 editions exist because the 4 platforms do this job differently. Claude is the only one that wires your private data in as sources through connectors, and it thinks the longest before answering. ChatGPT produces the strongest long-form synthesis and is the most confident when wrong. Gemini brings the biggest context window and the tightest Google Docs handoff. Perplexity tested most accurately on citations and holds a standing persona most loosely. No edition is best at everything. Pick the edition for the machine you live in, and if you run several, the steps are the same on each.

These are the files we will be using in Oper(AI)te this Summer, starting this Tuesday, the 21st of July

The Claude Edition: The only edition that connects your private data as sources through connectors.

The ChatGPT Edition: Built around Deep Research and Memory.

The Gemini Edition: Built around Deep Research with Google grounding and Workspace.

The Perplexity Edition: Built around Research mode, Labs, and Spaces.

Grab Your Seat Before They're Gone!

What should you do with this workflow this week?

The pipeline, compressed: commission research with a real brief that demands chartable data. Move the output to a second, different model and audit it against a written standard before you trust a number. Rebuild the visuals in your own identity. Produce the final document in your own voice, then read every word yourself before it ships.

This week, run 1 piece of it. Take any research you are currently relying on, drop it into a different model from the one that produced it, and ask for an audit against the workbook’s source rules. Count the violations. The number will tell you whether you have research or a guess with a citation.

And if you want to build the full system live, this is exactly what we teach in Oper(AI)te, the Summer of AI Fluency cohort I am running with Wessal Khader. Six live builds: a research analyst, a voice employee, a meetings employee, a visual brand, wired into the tools you already use. We start Tuesday, July 21. Through the deadline, code TAKE500 takes $500 off, $1,999 instead of $2,499. Details at unfazedfounder.com/operaite.

Grab Your Seat Before They're Gone!

The tools will keep improving without your permission. The standards will not.

Frequently Asked Questions

Can you trust AI research from Perplexity without verifying it?

No. Perplexity Computer is the strongest collection engine I use; every claim arrives attributed, and it still shipped 5 substantive errors in this single report, including a 22-point understatement of Cisco’s crash and a mislabeled valuation ladder for Anthropic. Attribution is not accuracy. Audit everything that feeds a decision or a publication.

Why should you verify AI research with a different model than the one that produced it?

Two models do not share the same blind spots. A model auditing its own output tends to re-approve its own errors, because the same weights that produced the mistake evaluate it. Collect in one place, then verify in another, ideally with a stronger model than did the collection.

Which AI is best for research: Claude, ChatGPT, Gemini, or Perplexity?

They do different jobs. Claude is the only one that reads your private data as sources through connectors, and it thinks the longest before answering. ChatGPT produces the strongest long-form synthesis and is the most confident when wrong. Gemini brings the largest context window and the cleanest Google Workspace handoff. Perplexity tested most accurately on citations. The staff you want runs each job on the machine that does it best.

What is a Claude skill?

A skill is a structured instruction file that the model loads when a task calls for it. Mine encodes an approved source list, 4 research modes, search and paywall rules, 3 output templates, voice rules, and 8 quality checks. The file replaces my memory and my mood as the quality-control layer.

What are the most common errors in AI research?

Wrong citations delivered confidently. The Columbia Tow Center found 8 AI search engines got the source wrong more than 60 percent of the time across 1,600 queries. After that: fabricated links, stale data stated in present tense, single-sourced claims, and vendor marketing repeated as fact. Every one of these is testable, which is why the workbooks ship with a source audit.

How long does this workflow take?

The machine research took 33 minutes. The full pipeline, audit, charts, and write-up included, fit in an afternoon, plus the time it takes you to read the final report in full, which is not optional. Building your own research employee from the attached workbooks takes about 1 hour.

Do you need to know how to code to build this?

No. Every piece of this system is a document and a set of prompts. You upload a workbook, paste 1 prompt, and answer questions in plain language. The most technical act in the entire pipeline is downloading a file from one tool and attaching it to another.

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is a career consultant and business operator, asking the question: What is the future of companies?

Connect with John on LinkedIn

The Most Common Lie at Work in 2026

John Brewton — Wed, 15 Jul 2026 22:35:26 GMT

Operating by John Brewton

Here is the secret almost nobody in your office or amongst your friend group is saying out loud. We are not using AI nearly as much as we pretend to, and we are not using it nearly as well as we would like the work world to think.

And that’s okay, friends.

This is one of the most common truths in the workforce and across corporates right now. The models sprinted ahead at a pace no prior technology has matched. Most people who did not start 2 years ago are quietly walking behind, overspending on tokens, prompting vaguely, and hoping nobody checks. Directors are reviewing reports that were completely generated by the models and are doing so with direct reports who can barely speak to their specifics. And the employees creating those reports are relying on models to draw conclusions from data that inaccurately communicates the company’s reality.

No shame in any of it, it’s just the simple, common truth of our times. The gap is real, but it is also the shortest, most closable gap of your career.

The space between layoff or promotion, anxiety and peace, becoming irrelevant versus becoming profoundly relevant to our companies’ futures is much smaller than we understand. Our social feeds, friendly conversations, and even late-night dreams are inundated with fearful headlines, images, and stories that we need to cast aside as we make the simple but powerful choice to become AI-fluent.

Grab Your Free Masterclass Seat Here!

TL;DR

Most people are using AI less and less well than they let on. So are most companies. The evidence says this is the norm, not the exception.
95 percent of enterprise AI pilots return nothing. The cause is a fluency and data gap, not weak models. (MIT NANDA, 2025)
“Workslop” is what the bluff looks like from the outside. It is a prompting-clarity problem that makes colleagues rate the sender 42 percent less trustworthy. (HBR)
The cost of a fixed level of AI performance fell by about 10 times per year, a 1,000x drop across 3 years. When the machine is nearly free, the value moves to the operator, who can aim it. (a16z)
AI raises the floor for everyone by about 14 percent on routine tasks. An uplift everyone gets is worth nothing as an edge. The edge sits at the ceiling, a 56 percent wage premium for real fluency. (Brynjolfsson, PwC)
The distance from bluffing to fluency is now measured in weeks, not years. Fluency is built, not innately gifted.

OPER(AI)TE · THE STAKES

A line is forming, and it runs straight through your industry. On one side are the people who use AI, or say they do. On the other hand, there are the people who operate it. The first group is already commoditized. The second group is getting promoted, paid, and hired to run the transformations the first group cannot finish.

Oper(AI)te exists to move you across that line before the fall. It is a live build, not a course you watch. Six sessions, a cohort, and a working staff of AI systems that you finish and keep. You can stop bluffing and get on with growing.

Join The Cohort

Use code TAKE500 to receive a $500 discount on your seat.

Why do 95 percent of enterprise AI pilots fail?

They fail on fluency, not on technology. MIT’s NANDA initiative studied enterprise generative AI and found that roughly 95 percent of pilots deliver no measurable return. The barrier was not model quality, regulation, or compute. It was a learning gap. Generic tools do not adapt to a company’s workflow; the data feeding them is too dirty to trust, and the people deploying them never learned to direct them.

Sit with what that number means for you personally. If you have been privately worried that you are the one who is behind, the evidence says otherwise. Nearly everyone is behind, including enterprises spending millions to make it look like they are not. You are not late. You are early and quietly bluffing, like almost everyone else in the building.

The same report found that internal builds succeed only a third as often as builds from vendor-purchased tools. Read casually, that sounds like an argument against building your own. It is the opposite. Internal builds fail because fluency is scarce. The operators who know how to scope, direct, and finish a build are the missing input, and they are the ones who turn a dead pilot into a working system.

The 95 percent that return nothing, against the thin margin that breaks through. The models were not the problem.

Is using AI the same as being AI-fluent?

No. Using AI is an activity. Fluency is direction. Harvard Business Review studied 2,500 employees over 8 months inside a company that already had high AI adoption. The best users did not use AI more than everyone else. They used it differently. They set ambitious objectives, treated AI as a reasoning partner rather than a task executor, and delegated complex work with clear, direct, thorough instructions.

The people struggling with AI are not failing at technology. They are failing at clarity. They prompt vaguely, burn tokens on wandering conversations, and get wandering output back. The fix is not a certification or a computer science degree. The fix is learning to say exactly what you want, with the context the model needs, in as few tokens as possible. That is a writing skill and a thinking skill that operators already possess.

The most striking finding in the study concerns measurement. Leaders track how much AI their people use because it's easy to see. They do not track how well because that number is hard to see and is the one that matters. Usage is a vanity metric. Fluency is the real one. Which is also why the bluff has survived this long. Nobody was measuring the thing that would expose it.

What is “workslop” and why is it not a character flaw?

Workslop is AI output that looks like work and shifts the real effort onto whoever receives it. Harvard Business Review found it now accounts for an estimated 15.4 percent of workplace content. Each incident costs the receiver close to 2 hours. Modeled across a 10,000-person company, the bill runs past $9 million a year.

The reputational cost is an issue. Colleagues rated the senders of workslop 50 percent less capable, 42 percent less trustworthy, and 54 percent less creative. Un-fluency is not neutral. It leaves fingerprints, and everyone around you can see them.

Workslop is not a you problem. It is a clarity problem. It is the direct byproduct of vague prompting meeting dirty data, a person who has not yet learned to direct the model, working inside a company that has not yet cleaned what the model reads. Both halves are fixable, and neither is a verdict on anyone’s intelligence. The data mess is not incompetence. It is the plumbing that needs to be fixed. The vague prompt is not a lack of talent. It is an unpracticed skill; it takes roughly 6 weeks to become competent.

OPER(AI)TE · WHAT YOU BUILD

Fluency is not something you watch. You get fluent by building real tools, on your real work, until the systems run without you narrating them. The bluff ends the first time a system you built does your actual job while you watch.

In Oper(AI)te, you build a staff of AI systems across 6 live sessions. A voice model trained on your own writing. Research and synthesis that deliver results in hours rather than afternoons. Meeting prep and first-draft decks in 45 minutes. One problem you choose, built live with the cohort. This is the same work inside John’s $50,000 to $250,000 consulting engagements, rebuilt for an individual operator.

Join The Cohort

Use code TAKE500 to receive a $500 discount on your seat.

Why does the falling cost of AI make skilled operators more valuable?

Because the scarce input ceases to be the machine and becomes the person aiming it. Andreessen Horowitz documents what it calls LLMflation. The cost of a given level of AI performance decreases by about 10x every year. GPT-3-level capability cost $60 per million tokens in 2021 and $0.06 per million tokens by 2024, a 1,000x collapse in 3 years.

Falling costs changed what AI could do. This changed who could use it. Which changed where the value sits. Most people are still arguing about which model to buy. The value already left the model and moved to the operator who can point it at the right problem, in the right way, without waste. Token efficiency is not a technical footnote. It is judgment, priced. The person who gets the right answer in 1 precise prompt is doing fundamentally different work than the person who circles it for 40.

Cost per million tokens fell from $60.00 to $0.06 in 3 years. As the machine approached free, the value rose to the operator.

Does AI give a company a lasting competitive advantage?

MIT Sloan Management Review clearly makes the argument. AI cannot deliver a sustainable advantage because access to models, data, and hardware is commoditizing in the same way access to personal computers and the internet did. Every rival will run the same models within a year.

What remains is the human layer. “AI does not change anything about the fundamental nature of sustained competitive advantage when its use is pervasive,” the review concludes. The differentiation moves to creativity, judgment, and the capacity to build. Those are the assets that are not commoditized by a vendor’s release schedule. For a company, becoming AI-native is not a purchasing decision. It is a fluency installation and workflow-reimagining event, person by person.

Does AI fluency actually raise your salary?

PwC’s Global AI Jobs Barometer puts the wage premium for AI skills at 56 percent. Roles that require those skills show pay growing 42 percent faster than roles that do not, and senior-level skills are appearing in junior AI-exposed jobs at 7 times the normal rate. Fluency is a promotable and money-making event for us operators.

The anxiety running through the workforce right now is rational. Layoffs are real, the pressure to sound confident about a technology you privately do not understand is real, and the people who were laid off and are now building consulting practices overnight are carrying both at once. The fear is rational. It is also pointed at the wrong thing. The threat was never the technology. The threat is staying a bluffer while the premium for being real compounds. The same force producing the fear is producing the widest opening most operators will ever get, and it is open right now.

The Brynjolfsson, Li, and Raymond study of 5,000 support agents found a 14 percent average productivity gain from generative AI, concentrated among the least experienced workers. AI lifts the floor for everyone. An uplift everyone receives is worth nothing as a differentiator. The premium migrated to the ceiling, to the people who can do what cheap, universal AI cannot do on its own. The distance between those two ends is now measured in weeks of deliberate building, not years of study. That is the part nobody carrying the anxiety has been told.

AI raises the floor by about 14 percent for everyone. Fluency reaches the ceiling, a 56 percent wage premium that stays scarce.

Has a technology shift like this happened before?

Twice, at least, and both times the pattern held. In 1979, VisiCalc put a spreadsheet on a personal computer. Within a decade, the tool was everywhere and nearly free. The people who captured the careers and the equity were not the ones who owned the software. They were the analysts who learned to model with it. The tool was the floor. Fluency was the ceiling.

Economic historian Paul David studied why electricity produced almost no measurable productivity gains for 40 years after its arrival. The dynamo was installed. The factories had not yet reorganized the work around it. The gain came only when the work changed, not when the machine did. MIT’s learning gap is the same finding, a century later. There is one difference this time, and it favors you. The spreadsheet and dynamo generations had to wait decades for the reorganization. This skill installs in weeks, on tools that cost less per month than a business lunch.

How do you actually become AI-fluent?

You build. Reading about AI produces the same result as reading about swimming. Start this week with 1 task you repeat every day, build an AI system that handles it end-to-end, and measure the time it takes to return. That single loop, run repeatedly, is the entire difference between using AI and operating it.

You do not need to write code. You do not need a developer. You do not need to have been early. You need clear prompting, clean inputs, and the discipline to build instead of watch. The most efficient way to destroy your own relevance is to wait for your employer to train you. The operators who close the gap first will spend the next decade running the transformations everyone else is still piloting, and they will do it without the knot in their stomach, because the confidence will be real.

What are the Oper(AI)te workbooks, and why should you never print them?

They look like documents. They behave like software. Sound Like You, the voice workbook, is the one I hand people first. It runs 21 pages, 4 steps, and about 1 hour, and the first inside page says it plainly. This is not a worksheet. Do not print it.

You upload the whole file to Claude, ChatGPT, or Gemini. The model reads it, finds the instruction block written directly to it, and turns into your interviewer. It asks 9 questions about how you actually write, pushes back when an answer is vague enough to describe anybody, collects real sentences you have written, and assembles the output. A locked voice profile. A measured prohibition list of the words and phrases that mark a draft as machine-made, each ban backed by published frequency data. A scoring audit that catches those tells in any draft you produce from that day forward.

Look at what that object is. A program, delivered as a document, that runs on whatever model you already pay $20 a month for. The document is the software. The model is the computer. Your work is the data. An operator 30 years ago would have called it a disk, and that is how to think about the full set. There is one for each segment of the course. Voice. Research. Administrative and calendar management. Brand identity. Each installs a capability into your model, then helps you direct your growing staff of agents with it.

How a workbook installs. The document goes into the model, the model interviews you through 9 questions, and a working skill comes out the other side.

Notice what the voice workbook does in the language of this article. Its entire function is to strip the machine accent out of your writing until what is left is you. It is an un-bluffing machine. You hand it to the model, answer honestly for an hour, and one of the fingerprints that exposes the bluff is gone for good.

Why does watching the build matter more than the curriculum?

Because fluency transfers by observation, not by syllabus. In the course I do not present these workbooks as finished artifacts. I build them live, from scratch, with the cohort watching. You see how I speak to the models. How I prompt, how I direct, how I tweak when the output drifts, how I adjust when a platform update changes the behavior underneath me. That layer never survives the trip into a slide deck or a recording, and it is the layer that actually makes someone fluent.

For more than 2 and a half years, all of my work has run through these models and platforms. Every capability jump, every quiet regression, every workflow rebuilt mid-quarter because a release changed what was possible, absorbed step by step because the business depended on it. The cohort compresses those spans into 6 sessions of standing next to the work as it happens. This is the oldest training method there is. The apprentice does not learn the trade from the manual. The apprentice learns it beside someone who has already made the mistakes, and asks questions while the tool is still open on the bench.

The individual Oper(AI)te cohort is $2,499. Set that number next to a 56 percent wage premium and a market where 95 percent of companies are still waiting up to 3 years for a return they may never see.
You leave with a staff of AI employees that does your real work, not a certificate that says you watched. Six build sessions. Two weekly Operator Build Labs. One cohort. A workbook for every segment, built live in front of you and installed into your own model.
Everyone is bluffing.
You no longer have to. It’s time to start building your AI-Fluent future.
Join The Cohort
Use code TAKE500 to receive a $500 discount on your seat.

FAQ

What is AI fluency? AI fluency is the ability to direct AI systems toward real work precisely, cheaply, and end to end. It is distinct from AI usage, which is measured by activity. Fluency is measured by output, by token efficiency, and by the systems you can build and run.

Is everyone really exaggerating how much they use AI? The measurable evidence points that way. 95 percent of enterprise pilots return nothing, leaders track usage instead of skill, and 15.4 percent of workplace content is AI output that pushes work onto the receiver. The gap between claimed adoption and real fluency is the defining feature of this moment.

Why do most enterprise AI projects fail? Because of a learning and data gap, not model quality. MIT found 95 percent of pilots return nothing. Generic tools do not adapt to a company’s workflow, the underlying data is too messy to trust, and most teams never build the fluency to fix either.

Is it too late to learn AI in 2026? No. The wage premium for AI skills is 56 percent and the skill requirements are still moving faster than most workers, which means the gap is open and closing. The honest starting point is admitting where you are, and nearly everyone is starting from the same place.

Will AI replace my job? AI raises the baseline for routine tasks by roughly 14 percent, which compresses the bottom of the skill curve. The durable roles belong to operators who build and direct AI, not to those who only produce the same output AI now produces for free.

What is workslop? Workslop is AI-generated output that looks like work but pushes the real effort onto the receiver. It accounts for about 15.4 percent of workplace content and makes the sender look 42 percent less trustworthy to colleagues. It is a prompting-clarity and data-quality problem, not a character flaw, and it is fixable.

What is Oper(AI)te? Oper(AI)te is a live, 6-session build program by John Brewton that teaches operators to build and direct AI systems on their own real work. Participants leave with a working set of AI tools, not a video library. The individual cohort is $2,499.

Who is Oper(AI)te for? It is built for operators. Employed managers, directors, executives, founders, and owners running lean, including people rebuilding as consultants or coaches after a layoff. It is not built for anyone looking for passive consumption or a done-for-you service.

How is Oper(AI)te different from a normal AI course? A normal course is something you watch. Oper(AI)te is a build you finish. The tools are built live from scratch in front of the cohort, so you see the prompting, directing, and adjusting that recordings never carry. Fluency comes from that, not from lessons.

What are the Oper(AI)te workbooks? They are documents designed to be uploaded into Claude, ChatGPT, or Gemini rather than printed. The model reads the workbook, interviews you, and assembles a working skill from your answers, such as a voice profile, a research system, a calendar and administrative manager, or a brand identity. One ships with every segment of the course. Sound Like You, the voice workbook, takes about 1 hour and runs 4 steps.

Appendix · Sources

MIT NANDA initiative, on the 95 percent enterprise AI pilot failure rate and the learning gap, reported by Fortune, August 2025. Link
MIT Sloan Management Review, “Why AI Will Not Provide Sustainable Competitive Advantage.” Link
Harvard Business Review, “What the Best AI Users Do Differently,” March 2026 (study of 2,500 employees over 8 months with KPMG and UT Austin). Link
Harvard Business Review, “AI-Generated ‘Workslop’ Is Destroying Productivity,” September 2025. Link
Andreessen Horowitz, “Welcome to LLMflation,” on the 1,000x decline in inference cost. Link
Brynjolfsson, Li & Raymond, “Generative AI at Work,” NBER Working Paper 31161, published in the Quarterly Journal of Economics, 2025. Link
PwC, 2026 Global AI Jobs Barometer, on the 56 percent wage premium and 42 percent faster wage growth. Link
Deloitte Insights, “AI tokens: How to navigate AI’s new spend dynamics.” Link

Note on sourcing: findings 1, 7, and 8 come from MIT NANDA, PwC, and Deloitte, which sit just outside the core elite-source tiers used for the rest of this piece. They carry the most quotable figures and are flagged here for transparency.

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies

How To Stay Relevant In Our AI-First World w/ John Brewton & Wessal Khader

John Brewton — Wed, 15 Jul 2026 18:08:45 GMT

Available for iOS and Android

How to Actually Make AI Write Like You, Not Like a Robot, With the Accent of Your Favorite LLM.

John Brewton — Mon, 13 Jul 2026 14:03:21 GMT

TL → DR

Every AI model writes in a default style. Researchers can identify which model wrote a text with 99.88 percent precision, even when the model was told to write differently.
The proof has been very public. The word “delve” rose 14-fold in science papers. “Meticulous” ran at 34.7 times its normal rate in AI-edited text. At least 10 percent of 2024 medical abstracts show AI processing.
Describing your voice to the AI is not enough. Models drop instructions as you stack them, and every dropped instruction gets replaced with the default style.
The fix has 2 parts and a compendium of samples. Write a positive profile of what you do. Write a prohibition list that bans the AI’s defaults by name. Attach 3 samples of your real writing. Load it all before every task.
Each platform needs its own setup. Claude takes the full profile as a Project or saved Skill instructions. ChatGPT caps custom instructions at 1,500 characters per field, so the full profile goes in a Project. Gemini needs the samples placed right before the task.
The stakes are growing. The AI’s favorite words are now rising 25 to 50 percent a year in human speech, and 80 percent of people who wrote essays with AI help could not quote their own essays.

Want to learn how to make AI actually write in your true voice?

Register for my free masterclass, happening tomorrow.

In the first quarter of 2024, the word “delve” appeared in scientific papers at roughly 14 times its historical rate. For 2 decades it had idled at about half a mention per 1,000 papers. Then it hit 7.9. Nobody called a meeting about this. No style guide changed, and no editor sent a memo asking the world’s researchers to start “delving.” A machine developed a preference, and several hundred million people carried it into the permanent scientific record, where it sits now like a watermark.

That watermark is the subject of this essay. The machine you write with has a voice of its own, and it is overwriting yours. Until you name that voice and explicitly turn it off, your instructions are negotiating with a fingerprint that was trained long before you arrived.

Most advice about AI writing never touches this. It teaches you to describe what you want, in ever more detail, and assumes the model is a blank instrument waiting for direction. The evidence says otherwise. The model arrives with commitments. You do not have a prompting problem. You have a fingerprint problem, and the fix has 2 jobs where almost everyone does 1.

The live masterclass is tomorrow. Everything in this article gets built in the room, in 1 hour: the 2-sided voice profile, the prohibition list, and the 3-machine setup for Claude, ChatGPT, and Gemini. You watch me build mine, then you build yours. The session is free.

Grab Your Free Masterclass Seat Here!

Why does all AI writing sound the same?

A team led by Dmitry Kobak analyzed more than 14 million biomedical abstracts and found that at least 10 percent of 2024 abstracts had been processed with a language model. In some subfields, the estimate reached 30 percent. The giveaway sat in 329 “excess words” that spiked past their projected baselines, and the composition of that list is the telling part. 66 percent were verbs, 18 percent adjectives. Vocabulary of style rather than substance.

A separate study of conference peer reviews measured the same signature from a different angle. In the AI-modified text, the adjective “commendable” appeared at 9.8 times its expected rate. “Intricate” ran at 11.2 times. “Meticulous” ran at 34.7 times. At EMNLP 2023, an estimated 16.9 percent of review sentences had been substantially modified by a model. The reviewers judging the field’s science were, in 1 of 6 sentences, letting the machine speak for them.

“Meticulous” at 34.7 times baseline, “delve” at 14, “intricate” at 11.2, “commendable” at 9.8. At least 10 percent of 2024 biomedical abstracts show LLM processing.

The issue reaches well past academia. Stanford researchers tracked it across the wider economy through late 2024 and found it in 17.7 percent of consumer complaints, 23.8 percent of corporate press releases, 13.7 percent of United Nations press releases, and roughly 1 in 10 job postings.

A quarter of corporate communication now arrives pre-inflected. Somewhere in a training pipeline, a preference formed the way an accent forms, absorbed rather than chosen, and it is now the most widely spoken accent on earth.

How do researchers detect AI writing?

Catching an author by counting words predates the AI era by 60 years. In 1964, the statisticians Frederick Mosteller and David Wallace settled a 175-year dispute over who wrote 12 unclaimed Federalist Papers. They ignored the arguments and counted the filler. Madison wrote “whilst” where Hamilton wrote “while.” Hamilton reached for “upon” at rates Madison never approached. The invisible words convicted him. All 12 papers went to Madison, and forensic stylometry became a discipline.

The same method now works on machines, and it has become far more effective. A 2025 study trained classifiers on 200,000 texts from 4 model families and identified which model wrote a given passage with 0.9988 precision. The authors’ summary sentence deserves to be quoted in full: large language models “have distinct and consistent stylistic fingerprints, even when prompted to write in different writing styles.”

Models are trained toward human approval, and the training compresses them. Researchers have documented that reinforcement learning from human feedback reduces output diversity. The text comes out more similar, regardless of what goes in. A 2025 paper traced the cause to what it called "typicality bias." Human raters systematically reward conventional, expected phrasing, independent of quality, so the model’s probability mass collapses toward its most typical sentences. Anthropic describes the endpoint of this process plainly in its own research. Post-training, in the lab’s words, selects “one particular character from this enormous cast” and places it center stage. The character is stable. It generalizes across models. It is a house voice.

Why does AI ignore my writing instructions?

The standard response is to write better instructions. Describe the voice. Add detail. Be firm. The instruction-following literature explains why this keeps disappointing you.

On IFEval, the benchmark for objectively checkable instructions, a GPT-4 class model failed roughly 1 multi-instruction prompt in 4; these were instructions a script could verify, including word counts and required keywords. A 2025 study stacked constraints and watched compliance sag, with the best frontier models honoring about 68 percent at a density of 500 instructions. The Allen Institute for AI then ran the sharpest version of the test. Models scoring high on familiar constraints collapsed on unfamiliar ones. Claude 4 Sonnet went from 91.3 percent to 42.3. Gemini 2.5 Pro went from 65.4 to 52.3. OpenAI’s o3 went from 95.0 to 69.3.

The same 3 models on seen constraints (IFEval) and unseen constraints (IFBench). Claude 4 Sonnet drops from 91.3 to 42.3. The best frontier accuracy at 500 stacked instructions is 68 percent.

Your tone rules are exactly the kind of unfamiliar, unverifiable constraint that decays first. When a constraint decays, the model reverts to the trained default. So the more elaborate your description of your voice, the more constraints you have queued for decay, and every decayed constraint is refilled with house style.

There is a second finding that matters even more for the fix we are all seeking. Models exhibit what MIT researchers call affirmation bias. They under-process negation and instead attend to the named items, with performance in the underlying study dropping by nearly 25 percent when statements were negated. The study was conducted on vision-language models. But the direction of the finding aligns with every operator’s experience with text models. A default you have not named stays on. A default you have named weakly also stays on. Suppression has to be explicit. It has to be specific. And the machine needs to hear it more than once, because its reflex is to register the noun and miss the “never.”

What is a 2-sided voice profile?

The first side of a working voice profile is the positive profile, the description of what you actually do. Your sentence rhythm. Your openings and your closings. Whether you lead with the conclusion or build to it. Your punctuation habits. How your register shifts between a cold email and an internal memo. Most people who take AI writing seriously have built some version of this. It is necessary, but by itself it does almost nothing about the fingerprint because it never mentions the fingerprint.

The second side is the prohibition list, the explicit inventory of what the model must never default to. The negation reflex, “it’s not X, it’s Y,” which the machine deploys with the regularity of a metronome. The rule of three, triplets in every sentence, makes it feel complete. The vocabulary you now know by heart: “delve,” “robust,” “seamless,” “testament,” “pivotal,” “underscore.” The grand contextual opener that clears its throat for a paragraph before saying anything. Signposting. Paragraphs of eerily uniform length. Each of these is a measured, documented default, and each one stays on until it is named and switched off. The complete inventory, with the measured effect sizes, sits in the appendix at the end of this piece.

The 2-sided architecture. Positive profile, prohibition list, and 3 few-shot anchors, loaded together before the task.

You can spend an hour describing your voice in exquisite detail, and the model will still hand you “delve” and a triplet, because you described what you want and never prohibited what it does. Both sides need to load together, or the calibration fails.

Do writing samples work better than instructions?

The third component outperforms both descriptions.

Show the model your actual writing.

The finding is as old as the modern field. The 2020 GPT-3 paper demonstrated that models can perform new tasks from a handful of examples in the prompt, with no retraining required, and quantified the gap. On 1 benchmark, accuracy went from 64.3 percent with instructions alone to 71.2 percent with a few examples. Later work sharpened the reason in a way that matters for voice. When researchers deliberately corrupted the content of in-context examples, performance barely moved. What the model was absorbing was the shape. The distribution of the text, the format, and the rhythm of how input becomes output. Which is to say, examples convey exactly what you cannot fully put into words about your own prose.

Anthropic’s documentation calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure.” OpenAI’s guidance compresses it to 3 words: “show and tell.”

This should only take 10 to 15 minutes. Pick 3 of your best pieces of writing in different registers: 1 email, 1 memo or post, and 1 longer piece. These are your few-shot anchors, your voice in its native format.

How do I set up a voice profile on ChatGPT, Claude, and Gemini?

Here is where nearly all AI writing advice goes soft. It claims to work on any model, which is precisely why it works well on none of them. The machines differ, and the differences are documented. A classifier separates their prose at 0.9988 precision. The same voice profile, loaded identically, produces 3 different drifts.

Claude honors explicit prohibitions most faithfully, and its failure mode is over-compliance. Strip out enough defaults and it can go flat. The fix is to pair each major prohibition with 1 positive rule. Avoid signposting, but open with strong declaratives. Load the full profile, positive side first, prohibition list second, anchors third, as per Project instructions, so every conversation inherits it.

ChatGPT’s default is agreeable polish. It will sand your fragments and finish your abrupt endings unless you name them as features. Its custom-instructions fields cap at 1,500 characters each, so the play is a condensed rule set there and the full profile with anchors as the first message in a Project, where it persists.

Gemini over-formalizes, an elevation reflex more pronounced than the other 2, and it weights recent context heavily. Put the profile in a Gem’s instructions, place the examples immediately before the task rather than burying them early, and give the prohibition list an explicit register rule. Match the examples, not what reads as professional. The position advice has empirical teeth. Language models reliably privilege the beginning and end of their context window over the middle, and Anthropic’s own docs report that placing the query after long context improves response quality by up to 30 percent.

1 profile, 3 platforms. Claude over-complies, ChatGPT polishes, Gemini formalizes. Each drift has a 1-adjustment fix and a placement rule.

The differences are proof that voice calibration is a technical discipline rather than a mindset, and that platform-agnostic advice stays shallow for that reason.

Is AI changing how people write and speak?

The strangest finding in this literature concerns your speech rather than your drafts.

Researchers at the Max Planck Institute for Human Development analyzed 360,445 academic talks on YouTube and 771,591 podcast episodes, roughly 7.35 billion spoken words. In the 18 months after ChatGPT launched, the words the model favors- “delve,” “realm,” “meticulous,” “underscore,” rose in unscripted human conversation at annual rates of 25 to 50 percent. The patterns stored in the technology “seem to be transmitting back to the human mind.”

Machine-favored words are rising 25 to 50 percent yearly in 7.35 billion words of human speech. 80 percent of AI-assisted writers could not quote their own essays.

The written evidence points the same way, and it is peer-reviewed. A Science Advances study gave 300 writers AI-generated story ideas and had roughly 600 evaluators judge the results. Individual stories got better, especially from the weaker writers. The collection got worse. Stories written with AI help were measurably more similar to one another than stories written without it. Each writer improved into the same writer. And the pull has a direction. A Cornell study of American and Indian writers using AI autocomplete found the suggestions steadily moved Indian participants toward Western styles, which means the default is 1 register, exported, rather than an average of everyone.

The New Yorker’s Kyle Chayka compressed the phenomenon into 5 words. AI, in his phrase, is “a technology of averages.” The MIT study he reported carried the detail that should bother you most. Of the subjects who wrote essays with ChatGPT’s help, 80 percent could not quote from what they had putatively written.

The professional class spent the last 2 years paying monthly subscriptions to sound like each other. The tool was sold as amplification and adopted as ventriloquism, and most users never noticed the transfer because the prose kept arriving fluent and finished. The fluency we were purchasing was cheap. What belongs to the machine is “finished,” and increasingly, the memory of having written at all.

Can AI detectors prove a text was written by AI?

The word-level tells are aggregate signals, not individual proof. AI detectors built on them misclassified 61.22 percent of essays by non-native English speakers as machine-written in a 1 study at Stanford. A single “delve” convicts nobody. The mechanism behind the vocabulary is also unproven. The popular theory blaming annotation workers’ dialects was tested directly and found unsupported. And the fingerprints move. One 2025 analysis found that DeepSeek’s prose was classified as OpenAI’s 74.2 percent of the time, which suggests the house styles can converge as labs train on similar data. The prohibition list you write this year will need an audit next year.

How do I make AI write in my voice this week?

What to do this week…

Pull your last 5 AI-assisted drafts and audit them against the tells in Figure 1.

Count the negation structures, triplets, marker words, and signposts.

This count is your baseline.

Write the positive profile in 1 sitting. Rhythm, openings, escalation, punctuation, argument order, register shifts. Keep it to one page max.

Write the prohibition list from your audit. Name every default you found, specifically. Not “avoid AI-sounding language” but “never use ‘delve,’ never construct ‘it’s not X, it’s Y,’ never signpost.”

Choose your 3 anchors and load all of them together before the task, on whatever machine you use, configured the way that machine requires.

Then track your edits for a week. Every word you swap and every sentence you reorder is data on a rule the profile still lacks. The gap between what the model gave you and what you shipped is your voice, captured in real time.

In the live session, we build both sides in the room, run the same profile through Claude, ChatGPT, and Gemini side by side, and fix each machine’s drift as it appears. Bring the 3 anchors.

What words and patterns give away AI writing? The full prohibition list.

The difference between a writer’s reframe and the machine’s reflex is about choice, and frequency. Use the pattern once, on purpose. Without clear directoin, the machine uses it 6 times a page. The same goes for every entry below. Possession of a single marker creates little impact, and detectors built on these markers falsely flagged 61 percent of essays by non-native English speakers. The tell most always lives in the presence of the same move, arriving again and again.

The markers fall into 2 evidentiary classes:

The vocabulary is measured, with effect sizes from studies covering more than 14 million documents.
The structures are documented by detection researchers and linguists, who observe rather than quantify them.
Load all 4 sections into your AI as explicit prohibitions, then pair each structural prohibition with 1 positive rule from your own writing.

Become An Operating Founder

Which words does AI overuse?

The 15 measured word families. “Delves” grew 6,697 percent, 2020 to 2024.

The core cluster, with effect sizes. Growth figures measure frequency change in scientific writing, from 2020 to 2024 (Juzek and Ward). Fold-increase measures overrepresentation in AI-modified peer reviews (Liang et al.). Frequency ratios measure excess above baseline in 14 million PubMed abstracts (Kobak et al.).

delve / delves / delved / delving. “delves” grew 6,697 percent. The flagship marker.
underscore/underscores/underscoring. “underscores” grew 904 percent.
showcase/showcasing/showcases. “showcasing” grew 1,396 percent.
meticulous / meticulously. 34.7-fold increase in AI-modified reviews.
intricate / intricacies. 11.2-fold increase. “Intricacies” grew 773 percent.
commendable. 9.8-fold increase.
boasts. Grew 918 percent.
comprehending. Grew 899 percent.
surpass/surpasses/surpassing. Grew 667 percent.
garnered. Grew 437 percent.
emphasizing. Grew 397 percent.
realm. Grew 381 percent.
groundbreaking. Grew 330 percent.
advancements. Grew 278 percent.
aligns. Grew 267 percent.

The documented adjective set, from the same studies’ extended lists: notable, versatile, noteworthy, invaluable, pivotal, potent, ingenious, laudable, lucid, cogent, methodical, profound, refreshing, intriguing, admirable, exceptional, remarkable, seamless, holistic, insightful, comprehensive, compelling, unprecedented, cohesive, tangible, multifaceted.

The documented adverb set: notably, aptly, methodically, compellingly, impressively, undoubtedly, undeniably, seamlessly, effortlessly, robustly, thereby, hitherto, herein, additionally, subsequently, crucially, markedly, judiciously.

A composition note: Of the 329 excess words of 2024, 66 percent were verbs and 18 percent were adjectives. The fingerprint lives in how the machine says things. Content words like “potential,” “findings,” and “crucial” also rose and remain weak signals on their own since any writer covering research reaches for them.

Which phrases give away AI writing?

“It’s not just X, it’s Y” and every variant.
“In today’s fast-paced world.”
“In the ever-evolving landscape of.”
“Plays a pivotal role in.”
“Serves as a testament to.”
“Rich tapestry.” “
Navigating the complexities of.”
“Unlock the potential of.”
“At the intersection of.”
“In the realm of.”
“A delicate balance between.”
“It’s important to note that.”
“It’s worth noting that.”
“Let’s dive in.”
“Let’s delve into.”
“In conclusion” and “Ultimately” are closing moves.
“Whether you’re a beginner or an expert.”
“Underscores the importance of.”
“Sheds light on.”
“The key takeaway.”
Stacked hedges like “could potentially.”
And every trace of the eager assistant: “Great question.”

Which sentence structures does AI default to?

The antithesis construction: “This isn’t X. It’s Y.” Deployed several times per page.
The rule of three: Nouns, adjectives, and clauses arriving in threes, in every sentence that wants to feel complete.
The participial tail: A clause bolted on with an “-ing” verb: “...showcasing its versatility,” “...underscoring the need for further research.” 5 of the 21 fastest-growing marker words are exactly these participles.
The em dash cascade: Several per paragraph, doing work commas and periods should do.
Uniform sentence length: Human writing mixes 4-word sentences with 40-word sentences. Machine writing hums at a steady cadence.
The balanced concessive: “While challenges remain, the opportunities are significant.” Every tension is neutralized in a single sentence.
Elevated formality throughout: No contractions, no slang, no fragments, even where a human would relax.
The signposted transition: “Additionally,” “Furthermore,” “Moreover,” “Notably” at the head of paragraphs after paragraphs. Several of these adverbs appear on the lists of measured overuse.

Which document habits give AI away?

The throat-clearing open: A paragraph of context before anything is said.
Uniform paragraph length: Every paragraph has 3 to 4 sentences; the whole page is a solid gray.
The scaffolded essay: First, second, third, finally, whether or not the material has that shape.
Bolded lead-in bullets: Lists where every item opens with a bolded phrase and a colon, much like this section, but it’s how I like to construct sections like this, so I do it anyway.
The both-sides shuffle: Every claim balanced by its counterclaim, no position taken.
Over-explanation: Restating the question, summarizing what was just said, announcing what comes next.
The tidy moral: A closing paragraph that resolves everything and lands on an uplifting generality.
Completion compulsion: Nothing left open, every thread tied, even when the honest answer is unresolved.

Frequently asked questions

Why does my AI writing sound like AI?

Every model is trained toward the answers human raters prefer, and raters reward conventional phrasing. That training compresses the model into 1 default style, the same vocabulary and sentence shapes for every user. Your instructions sit on top of that default. They soften it. They do not remove it until you ban its parts by name.

What words does AI overuse?

The measured core: delve, underscore, showcase, meticulous, intricate, commendable, boasts, comprehending, surpass, garnered, emphasizing, realm, groundbreaking, advancements, and aligns, plus their inflections. “Delves” grew 6,697 percent in scientific writing between 2020 and 2024. The full list, with phrases and structures, is in the appendix above.

How do I stop ChatGPT, Claude, or Gemini from sounding like AI?

Build a voice profile with 2 sides. Side 1 describes what you do: your rhythm, your openings, your punctuation. Side 2 bans the machine’s defaults by name: the words, the “it’s not X, it’s Y” reflex, the rule of three, the signposting. Attach 3 samples of your real writing. Load all of it before the task, every time.

Do writing samples work better than written instructions?

Yes. The foundational GPT-3 research showed accuracy rising from 64.3 to 71.2 percent on the same task when examples were added. Later work found that examples transfer format and rhythm, the parts of a voice you cannot fully describe. Anthropic calls examples “one of the most reliable ways to steer” tone. OpenAI’s version is “show, and tell.”

Can AI detectors prove a text was written by AI?

No. The word-level tells are reliable across millions of documents and unreliable on any single one. In 1 Stanford study, detectors falsely flagged 61.22 percent of essays by non-native English speakers as AI-written. A single “delve” convicts nobody. The tell is the pattern repeating across a whole body of text.

Is the same voice profile enough for every AI platform?

The words are the same. The setup is different. Claude takes the full profile as Project instructions and needs each ban paired with a positive rule. ChatGPT caps custom instructions at 1,500 characters per field, so the condensed rules go there and the full profile goes in a Project. Gemini weights recent context heavily, so the samples go right before the task.

Is AI changing how people talk?

The early evidence says yes. Max Planck researchers analyzed 7.35 billion words of podcasts and talks and found the machine’s favorite words rising 25 to 50 percent a year in unscripted speech after ChatGPT launched. The study is a preprint, so treat it as a strong early signal rather than settled fact.

How long should a voice profile be?

Short enough to survive. Instruction-following decays as rules stack, so a bloated profile fails faster. Target 1 page for the positive side, a named prohibition list, and 3 writing samples. Fewer, harder rules beat many soft ones.

Build it live, tomorrow.

This article gives you the doctrine. The masterclass gives you the working setup.

Tomorrow, live, we will build every piece from this article in 1 hour. Your positive profile. Your prohibition list, drawn from the measured word lists above. Your 3 few-shot anchors. Then we load the same profile into Claude, ChatGPT, and Gemini side by side and fix each machine’s drift as it appears, so you leave with your voice running on whatever AI you already use.

You do not need a technical background.

You need 3 pieces of your own writing and 1 hour.

The session is free.

Research appendix: every source in this article.

The measured tells

Kobak, González-Márquez, Horvát and Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary,” Science Advances 2025. 14M+ PubMed abstracts, the 329 excess words, the 10 percent figure. arxiv.org/abs/2406.07016
Liang et al., “Monitoring AI-Modified Content at Scale,” ICML 2024. The peer-review fold increases: meticulous 34.7, intricate 11.2, commendable 9.8. arxiv.org/abs/2403.07183
Liang et al., “The Widespread Adoption of Large Language Model-Assisted Writing Across Society,” Stanford HAI / Patterns 2025. The economy-wide adoption rates. arxiv.org/abs/2502.09747
Juzek and Ward, “Why Does ChatGPT ‘Delve’ So Much?,” COLING 2025. The 21 focal words with growth rates, and the tested-and-rejected annotator theory. arxiv.org/abs/2412.11385
Shapira, “Delving into ‘delve,’” 2024. The 14-fold rise of the flagship word. pshapira.net/2024/03/31/delving-into-delve
Gray, “ChatGPT ‘contamination’: estimating the prevalence of LLMs in the scholarly literature,” 2024. The scholarly keyword growth figures. arxiv.org/abs/2403.16887

The fingerprint and its mechanism

Bitton, Bitton and Nisan, “Detecting Stylistic Fingerprints of Large Language Models,” 2025. The 0.9988 precision result and the DeepSeek convergence finding. arxiv.org/abs/2503.01659
Kirk et al., “Understanding the Effects of RLHF on LLM Generalisation and Diversity,” ICLR 2024. Post-training reduces output diversity. arxiv.org/abs/2310.06452
Zhang, Yu, Chong et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity,” 2025. Typicality bias as the cause. arxiv.org/abs/2510.01171
Anthropic, “The Assistant Axis: Situating and Stabilizing the Character of Large Language Models,” 2026. The default character, from the lab itself. anthropic.com/research/assistant-axis
Mosteller and Wallace, “Inference in an Authorship Problem,” Journal of the American Statistical Association, 1963. The Federalist Papers attribution. jstor.org/stable/2283270

Why instructions decay

Zhou et al., “Instruction-Following Evaluation for Large Language Models” (IFEval), 2023. The 1-in-4 prompt failure figure. arxiv.org/abs/2311.07911
Jaroslawicz et al., “How Many Instructions Can LLMs Follow at Once?,” 2025. 68 percent at 500 stacked instructions. arxiv.org/abs/2507.11538
Pyatkin et al., “Generalizing Verifiable Instruction Following” (IFBench), Allen Institute for AI, 2025. The seen-versus-unseen collapse: 91.3 to 42.3. arxiv.org/abs/2507.02833
Alhamoud et al., MIT, on affirmation bias and negation, 2025. Why prohibitions must be explicit. news.mit.edu

Why examples beat descriptions

Brown et al., “Language Models are Few-Shot Learners,” NeurIPS 2020. 64.3 to 71.2 percent with examples added. arxiv.org/abs/2005.14165
Min et al., “Rethinking the Role of Demonstrations,” EMNLP 2022. Examples transfer format and structure, corrupted labels barely matter. arxiv.org/abs/2202.12837
Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” TACL 2024. Position in the context window changes what the model uses. arxiv.org/abs/2307.03172

Platform documentation

Anthropic, multishot prompting guidance. Examples steer tone and structure. platform.claude.com
Anthropic, prompting best practices. The query-at-the-end 30 percent figure. platform.claude.com
OpenAI, prompt engineering best practices. “Show, and tell.” help.openai.com
OpenAI, custom instructions for ChatGPT. The 1,500-character limit. help.openai.com
Google, tips for creating custom Gems. support.google.com

The spread into human writing and speech

Yakura et al., Max Planck Institute for Human Development, on LLM-favored words in spoken language, 2025, preprint. 7.35B words, 25 to 50 percent yearly growth. arxiv.org/abs/2409.01754
Doshi and Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 2024. science.org
Agarwal, Naaman and Vashistha, on AI suggestions homogenizing writing toward Western styles, CHI 2025. arxiv.org/abs/2409.11360
Chayka, “A.I. Is Homogenizing Our Thoughts,” The New Yorker, 2025. “A technology of averages” and the MIT 80 percent figure. newyorker.com

The limits of detection

Liang et al., “GPT detectors are biased against non-native English writers,” Patterns 2023. The 61.22 percent false-positive figure. cell.com

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies

The Five Founder Problems I Discovered Coaching +100 Substack Founders in 2026

John Brewton — Fri, 10 Jul 2026 11:10:10 GMT

TL → DR

400 hours of coaching Substack founders in 2026 surfaced 5 recurring problems: positioning, packaging, pricing, distribution, and AI.
Underneath all 5 sits 1 cause. A decision the founder will not make. The writing is never the constraint.
Positioning: pick the 1 problem you own. Companies that define a category capture roughly 76 percent of its value.
Packaging: productize the expertise. Service businesses trade at around 1 to 2 multiples, product businesses at around 6 to 8.
Pricing: attention is not revenue. Price to the reader’s outcome and put an end date on free.
Distribution: the email list is the only audience you own. Grow on 1 channel and convert to the list.
AI: the machine makes you roughly 40 percent faster and measurably more alike. Decide task by task what you draft and what only you write.

What you get

One year in the Operating Founder Community: A working room of founders, operators, and writers building real businesses. You bring the problem. The room and I will help you decide.
The weekly Q&A Masterclass: 52 live sessions, 60 minutes each, one a week for a year. Bring one problem. Leave with the decision and the next action. That is 52 hours of live coaching over the course of the year.
Three golden tickets: $1,000 off any course you choose in the next 12 months. Three tickets, $3,000 in course credit, yours to spend on the work that moves you.

What did 400 hours of coaching Substack founders reveal?

I spent 400 hours this year coaching Substack founders. Different niches, different subscriber list sizes, different decades of experience. I expected 400 hours of variety. I uncovered 5 clearly repeating problems.

Underneath the 5 challenges, sat 1 clear decision the founder needed to make. Every founder I worked with had real expertise and 1 unmade decision. There were no talent gaps nor any unwillingness to put forth the requisite effort.

But there was always a clear choice that needed to be made.

Here are the 5 conundrums these problems present.

How do I pick a niche for my Substack when I have broad experience?

The most common problem is broad experience with no chosen niche. When you can write about anything, you tell the reader nothing, and the subscriber who cannot name the 1 thing you own will not remember you exist.

Broad experience and diversity of storytelling belong in the work. Narrowness belongs in the positioning. Companies that define a category capture roughly 76 percent of its value. The narrow choice is where the value lives. Pick the 1 problem you solve better than anyone in your reach, and stop writing around it or expanding into spaces that confuse your market. This is critical when you’re at the starting line.

Go deeper:

The Difference Between a First Mover and a Category Creator (Harvard Business Review, 2019). The category-king research behind the 76 percent figure. Category creators capture the dominant share of a category’s market value.
Why It Pays to Be a Category Creator (Harvard Business Review, 2013). Category creators in the study grew revenue and market cap far faster than their industries. The business case for owning 1 problem.

How do I turn my expertise into a product instead of selling hours?

A newsletter that only sells hours caps the business at the size of your calendar. A client book concentrated in a handful of names looks like success, but can break when one disengagement email arrives.

Service businesses trade at around 1-to-2 times multiples. Product businesses trade at around 6 to 8. Crossing that chasm is about designing how your business makes money, not talent. Package the expertise once and sell it many times. If you keep the premium with 1-to-1 work, raise the price, or choose to go 1-to-many via communities and cohorts.

Go deeper:

Putting Products into Services (Harvard Business Review, 2016). The definitive playbook for converting labor-based expertise into productized offerings, with named cases from professional services firms.
What Professional Service Firms Must Do to Thrive (Harvard Business Review, 2021). A framework for which client work to keep premium and which to systematize.

When should I start charging for my newsletter, and how do I price it?

Subscribers are a leading indicator. Revenue is a separate transaction that occurs only when you build the offer and set the price. My best advice is to begin monetizing the moment you are clear on the offer you are making. Market response, or lack thereof, is a terrific indicator against which to make forward-looking decisions

Free is defensible as a runway and indefensible as a permanent state. Price to the reader’s outcome, not your effort, and put a date on the day the free period ends.

Go deeper:

A Quick Guide to Value-Based Pricing (Harvard Business Review, 2016). The mechanics of pricing to the buyer’s outcome and next-best alternative instead of your effort.
The Creator Economy Needs a Middle Class (Li Jin, Harvard Business Review, 2020). The structural analysis of why audience size and creator income diverge, with platform-level data.

Should I post on every platform or focus on building my email list?

Founders posting across 4 platforms watch their reach vanish the week they step away, or never see it show up anywhere. Moreover, platform reach is borrowed and can be revoked without notice.

This is why Substack matters. The email list is the asset you own. The post is not the asset. The list is. Grow on 1-to-2 primary channels strictly with the aim of growing your email and subscriber lists, convert a subset to the list. That list of people is the market you sell to.

Go deeper:

Pipelines, Platforms, and the New Rules of Strategy (Harvard Business Review, 2016). The canonical explanation of who controls demand on a platform and who captures the margin.
1,000 True Fans? Try 100 (Li Jin, Andreessen Horowitz, 2020). The economic argument that a small directly-owned audience paying real prices beats a large rented one.

Grab Your Free Masterclass Seat

Will using AI make my writing sound like everyone else’s?

Founders adopt the machine and fear it flattens their voice. They are right about the default and wrong about the cause. Generative AI made writers roughly 40% faster and 18% better and made their work measurably more alike.

The uniformity comes from how you prompt, not from a law of the tool. Decide task by task what the machine drafts and what only you write.

Protect your judgment. Keep your voice.

Hand over the production speed to the models.

Grabe a free seat in my writing with AI masterclass, happening next Tuesday!

Go deeper:

Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence (Noy and Zhang, Science, 2023). The randomized MIT experiment behind the numbers: 40 percent faster, 18 percent higher quality.
Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content (Doshi and Hauser, Science Advances, 2024). The homogenization evidence: AI-assisted work rated more creative individually and measurably more alike collectively.

What is your one problem underneath all five?

Positioning, packaging, pricing, distribution, and the machine are not 5 skills to study. There are 5 decisions to make, and each one takes an afternoon, not a year.

Most of the founders I coached were not overwhelmed.

They were stalling and calling it overwhelm.

It’s time to move beyond overwhelm.

Pick 1 of the 5 and make the choice this week.

What’s your pick?

- j -

Every founder I coach has real expertise and one thing missing. The Operating Founder Community is the room where you make them with me every week for a year.

What you get

One year in the Operating Founder Community. A working room of founders, operators, and writers building real businesses. You bring the problem. The room and I help you decide.

The weekly Q&A Masterclass. 52 live sessions, 60 minutes each, one a week for a year. Bring one problem. Leave with the decision and the next action. That is 52 hours of live coaching over the year.

Three golden tickets. $1,000 off any course you choose in the next 12 months. Three tickets, $3,000 in course credit, yours to spend on the work that moves you.

The math

52 hours of live coaching. $3,000 in course credit. One year to make the five decisions that turn expertise into a business. Most founders spend a year circling those choices alone. You will spend it making them in a room built for it.

Who is this for?

Founders and operators with real expertise and no clear next move. Writers and creators building on Substack who have the audience and not the offer. Anyone tired of studying the problem and ready to move forward.

Here’s a worksheet I put together to help. Become a Founding member of Operating by John Brewton to join our weekly Masterclass Office Hours, where we unpack and solve problems across members’ businesses.

Become An Operating Founder