The Thing You Never Built

01The things that never get built

Almost everyone is carrying one.

A business described so often at dinner that the family can finish the sentences. A program the town obviously needs, which someone would start if only they had a year off. A product sketched on the back of a meeting agenda in 2014. The tool a team has requested in three consecutive planning cycles. A book about the summer the mill closed. A scholarship in a parent's name. A better way to do the thing you do all day, which you can see clearly and have never once had time to build.

These things do not go unbuilt because nobody cared. They go unbuilt because caring was never the expensive part.

To make almost anything real, someone has to do the research, design the thing, write the words, build the software, file the paperwork, set up the payments, schedule the people, answer the email, keep the books and send the hundred small follow-ups that turn a decision into a fact. For most of history the honest answer to "could I build that?" was a calculation about those jobs: whether you could personally do them all, hire people to do them, or raise money to hire people to do them. For most people, most of the time, the answer was no.

This essay is about what happens when that calculation starts to change. It is not a prediction that everyone will become a founder, or that AI will build your dream while you sleep. Some of what is said about AI is true, a great deal of it is exaggerated, and the most important parts are still uncertain. The evidence, including the evidence against, is laid out below.

The argument is narrower and, I think, more consequential: the amount of personally supplied human capacity required to turn a conviction into a working thing is falling. Not to zero, and not evenly. But enough that a question millions of people answered years ago, and answered no, deserves to be asked again.

02The old economics of agency

In 1937 a young economist named Ronald Coase asked why firms exist at all. If markets are so efficient, why don't we buy every input, every hour of every specialist, on the open market? His answer: "The main reason why it is profitable to establish a firm would seem to be that there is a cost of using the price mechanism."¹ Finding the right people, agreeing prices, writing contracts and coordinating the work all cost something. Organizations exist because, past a certain point, it is cheaper to employ capability than to assemble it piece by piece.

That insight explains a great deal about who has historically been able to act on their intentions. Between an intention and an outcome sits a chain:

Intent→Expertise→Labor→Coordination→Capital→Execution→Institution

Every link introduced friction, and most of the friction was priced in human capacity. Capital mattered largely because capital bought capacity: it hired the programmer, retained the lawyer, paid the designer, staffed the office. Organizations aggregated capacity. Employment was, at its core, an exchange of personal capacity for money. Entrepreneurship was the act of assembling enough capacity, borrowed, hired or personally supplied, to get an idea across the gap.

Under those economics, agency was rationed by access to capacity. A person with a clear vision, deep domain knowledge and an unusually good idea, but without money, technical skills, a network or the stamina to do it all personally, mostly could not act. Not because the idea was wrong. Because the organization the idea required was out of reach.

This is the sense in which I mean agency: not intelligence, not opinion, not even skill, but the practical ability to cause something to exist in the world. Agency has always been more unequally distributed than intelligence.

03When capacity is the bottleneck

I should say why this question is not abstract for me.

I spent my career in banking, payments and transformation, in large organizations where the hard part was rarely the idea and almost always the execution. Then my own capacity changed. I live with a disability that limits how long I can sit, read, look at a screen or work anything like a conventional day. Some days that's a few productive hours; some days it's none. I won't describe the medical details, because they aren't the point. The point is economic: my available execution capacity became small and unpredictable, while my judgment, experience and convictions stayed where they were.

That is an unusual position from which to look at work, and a revealing one. Disability is an edge case that exposes something universal: human capacity is finite, variable and unequally distributed, and our economic institutions mostly assume it is none of those things. They assume eight to twelve reliable hours, a body that can commute, sit and look at screens, and a life without competing obligations.

The numbers show how much that assumption costs. In 2025, 38.1% of working-age Americans with a disability were employed, compared with 74.8% of those without.² More than one in four U.S. adults reports a disability.³ About 63 million Americans, nearly one adult in four, provide ongoing unpaid care for someone with a complex medical condition or disability, and six in ten of them also hold jobs.⁴

One detail stands out. Employed people with disabilities are more likely to be self-employed: 9.1% versus 5.9%.² That gap partly reflects age (half of people with a disability are 65 or older) and partly, it seems, barriers to conventional employment, not just preference. But it also says something about what people do when the standard arrangement doesn't fit: they try to build their own.

And the constraint is not confined to disability. Age, caregiving, parenthood, chronic illness, geography, a demanding job, a thin bank account: sooner or later nearly everyone runs into a version of it. The retired engineer with forty years of knowledge and limited energy. The parent with a clear idea and no free evenings. The manager who knows exactly what her company should build and has no budget, no team and no permission. I suspect the capacity constraint is among the most common reasons good ideas stay ideas. It is also, historically, among the least discussed.

I'll be careful here. There is not yet good causal evidence that AI improves employment or self-employment outcomes for disabled people. What exists is mostly self-reported. In one UK government pilot, for example, neurodivergent staff were significantly more satisfied with an AI assistant than their colleagues, and one said it had "levelled the playing field," while the same evaluation found no evidence of overall productivity gains.⁵ My own experience is one data point, not a study. What follows is an argument about mechanism, tested against the best evidence available, not a claim of proof.

04What is actually changing

Here is the mechanism. AI is beginning to separate human intent from human execution capacity. Portions of research, writing, analysis, design, software, coordination and administration (the jobs that used to require hiring, buying or personally performing) can increasingly be called on through language, combined around a goal, and checked by a person who knows what good looks like.

The evidence for this is real, although narrower than the marketing.

Individuals are starting to match small teams. In a pre-registered field experiment with several hundred Procter & Gamble professionals, individuals working with generative AI produced product ideas of similar quality to two-person cross-functional teams working without it, and spent about 16% less time than individuals working without AI.⁶ People using AI produced solutions that balanced technical and commercial considerations regardless of which function they came from. That is a small, specific result. It is also exactly the mechanism this essay is about: capability that once required a second person, available to one.

Experience is becoming partially transferable. When a large firm gave an AI assistant to more than 5,000 customer-support agents, productivity rose 15% on average and about 30% for the least experienced and least skilled; agents two months into the job performed like agents with more than six months' tenure.⁷ In a randomized experiment with 453 professionals, ChatGPT cut the time needed for mid-level writing tasks by 40% and raised graded quality by 18%.⁸

Small organizations are filling gaps without hiring. In an OECD survey of more than 5,000 small firms in seven countries, 39% of generative-AI users that had faced a skills gap said the technology helped compensate for it, and 14% said it reduced their reliance on outside contractors. Tellingly, 83% said it had not changed how many staff they needed.⁹ The capability moved inside the firm without replacing the people already there.

The shape of new ventures is shifting. After ChatGPT's release, firm formation rose more than 10% in industries where AI can help with founders' early tasks, apparently because it made experimentation cheaper.¹⁰ Among venture-track startups on Carta, the share with a single founder rose from about 24% in 2019 to 36% in the first half of 2025.¹¹ Stripe's economists report that the number of one-person businesses on its platform earning more than $1 million roughly doubled between 2023 and 2025, though their data are proprietary and the link to AI is their interpretation.¹² Americans filed about 5.7 million applications to start businesses in 2025, the most on record, and a shrinking share of them plan to hire anyone.¹³ That surge began in 2020, before generative AI, so it can't be credited to AI. But the composition is consistent with a world in which more people try to build things without first assembling a staff.

And the systems are moving from answers to actions. By METR's measurements, the length of software task that leading AI agents can complete about half the time has doubled roughly every seven months since 2019, and faster since 2024. By early 2026 it had reached many hours. Reliability is a different matter: the length of task they complete 80% of the time is several times shorter.¹⁴ Agents can now take multi-step actions: building, filing, sending, scheduling. They do it imperfectly, which is why a person has to stay in the loop.

Put together, the picture is not "AI does everything." It is that capability is becoming callable. Some of what once had to be owned (an employee, a firm on retainer, a skill acquired over years) can increasingly be summoned, directed and checked. Coase's costs of organizing are not disappearing. Some of them are falling fast. Economists studying AI agents have begun asking what happens to markets and firms when they do.¹⁵

05Build what matters

Scarce capacity forces a question that abundant capacity lets you avoid.

When I could no longer count on conventional working hours, the usual entrepreneurial question ("what would make the most money?") stopped being the right one. With only a little capacity, I could build only a few things. So the question became: what matters enough to spend scarce capacity making real?

I think that question is the right one for almost everyone, and that it becomes more important as execution gets cheaper, not less. When building was expensive, the cost itself filtered what got built. As the cost falls, the filter has to come from somewhere else: from purpose, judgment, and an honest answer to "why should this exist, and for whom?"

My answer was young athletes.

06Community Academy

For years I had held a set of convictions about youth sports. That they had become heavily commercialized, with families paying facility margins for access to development. That the knowledge and measurement tools behind serious player development shouldn't belong only to families who can pay the most. That people who have played at a high level can be extraordinary mentors, if the economics let them. That development should be measured, individual and whole-athlete, not a matter of a guru's opinion.

Those convictions became Community Academy, a nonprofit training facility in Jacksonville Beach, Florida. It was incorporated as a Florida nonprofit corporation in January 2025 and opened later that year.¹⁶ It has an indoor facility with cages, mounds and a strength and movement space; a network of coaches that includes former college players; an evaluation that places each athlete on a development track; Rapsodo and Blast Motion pitch and swing measurement and a re-test cycle; camps, clinics and a homeschool group; published, modest prices; and a mechanism for families to donate unused sessions to scholarship players. Its stated model is that session fees go to the coaches who deliver them.

None of that is an outcome yet, in the sense that matters most. The outcomes that matter (what happens to young athletes over years) belong to the athletes, their families and their coaches, and they take time. I won't publish counts of athletes or sessions until they're reconciled against records, and I won't claim the former college and independent-league players who have coached there were developed there. They weren't; their careers came first. The full, sourced case study is here.

What I will claim is narrower. Someone working with severely limited capacity, who would once have needed a staff to attempt this, turned a long-held conviction into a functioning institution, together with the coaches and families who make it work. The causal chain, stated precisely:

AI-assisted organizational capacity→helped make the organization possible and manageable→the organization created an environment→coaches, athletes and families create the outcomes

AI didn't coach a single athlete. It helped carry the organization around the coaching.

07The invisible organization

A training facility looks like cages and coaches. A clinic looks like doctors. A product looks like an app. But behind every institution sits an organization nobody sees: the dozens of capabilities required to make the visible thing exist and keep existing.

It is worth mapping that invisible organization honestly, because the map shows both what is changing and what isn't.

Capability	Traditionally	AI and software can now carry	Irreducibly human
Research	Analyst, consultant	Landscape scans, comparables, literature, prices	Deciding which facts matter
Writing	Copywriter, comms staff	First drafts of pages, policies, proposals, emails	Voice; what is true; what not to say
Software and web	Developer, agency	Sites, forms, databases, automations, prototypes	What it should do, and for whom; owning the data
Design	Designer	Layouts, materials, variations	Taste: what good looks like
Administration	Assistant, office manager	Scheduling, reminders, filing, follow-ups	Promises made to people
Analysis	Analyst, finance	Models, scenarios, summaries of data	Choosing the risk
Legal and financial	Attorney, accountant	Preparation, checklists, first drafts for review	Licensed judgment; signing; accountability
Fundraising and sales	Development officer, sales	Prospect research, materials, tracking	Asking; relationships; trust
Leadership	Founder, director	Preparation for decisions	Purpose, decisions, responsibility

Read down the right-hand column. It contains purpose, judgment, trust, taste, relationships, courage and accountability. For many of the people this essay is about, none of those were the bottleneck. The engineer with the coding-club idea had judgment. The mother with the recipes had taste. The manager with the internal tool had relationships. What they lacked was the middle column, the part priced in hired capacity. That is the part that is changing.

08AI is not agency

If the argument stopped there, it would be marketing. It doesn't stop there, because the evidence doesn't.

AI helps inside its frontier and harms outside it. Among 758 BCG consultants, those given GPT-4 completed 12.2% more tasks, 25.1% faster and at higher quality, but only on tasks within the AI's capabilities. On a task deliberately chosen to be just outside them, consultants using AI were about 19 percentage points less likely to reach the correct answer than those working without it.¹⁷ The danger isn't the obvious failure. It's the plausible one.

Experts can be slowed down, and not notice. In a randomized trial, experienced open-source developers working on their own mature projects were 19% slower with early-2025 AI tools, while believing afterwards that AI had made them about 20% faster.¹⁸ A 2026 follow-up couldn't settle the question, partly because many developers declined to submit tasks they didn't want to do without AI.

The gains can flow to those who already have judgment. This is the finding that matters most for my argument. In a randomized trial with Kenyan small-business owners, a GPT-4 business advisor had no average effect on profits; already-strong entrepreneurs may have gained more than 15%, while struggling ones did nearly 10% worse, depending on which advice they chose to act on.¹⁹ The study of firm formation cited above found no evidence that AI worked by filling founders' skill gaps; its benefits tilted toward high-ability entrepreneurs.¹⁰ A meta-analysis of 106 experiments found that human-AI combinations, on average, performed worse than the better of the human or the AI alone, especially on decisions, though better on creative and content tasks.²⁰

Delegation can cost learning. In a small randomized study, junior engineers who learned a new tool with AI help scored markedly lower on comprehension (50% versus 67%) without finishing significantly faster.²¹ Many workers now report receiving AI-generated "workslop," polished-looking output that creates more work for whoever receives it.²²

The big picture is still small. Daron Acemoglu estimates AI will add no more than about 0.5 to 0.7% to U.S. total factor productivity over a decade.²³ Linking adoption surveys to Danish administrative records, Humlum and Vestergaard find no measurable effect of AI chatbots on workers' earnings or hours two years after ChatGPT's launch.²⁴ Representative Census data show fewer than one in five U.S. firms used AI in a business function in late 2025.²⁵ And the same capability that lets one person do more appears to be shrinking entry-level hiring: employment of 22- to 25-year-olds in the most AI-exposed occupations is running well behind comparable workers.²⁶ Capacity that flows to some people can be taken from others.

So no: AI is not agency. A model can tell you how to start a nonprofit. That is intelligence, widely available and increasingly cheap. Agency is a nonprofit that exists on Tuesday, with a program, a way to sign up, a way to pay, a place on the internet and a person accountable for it. Between those two sits a long sequence of decisions, coordination and action that no answer, however good, performs on its own.

Ask→Answer

Intend→Understand→Decide→Coordinate→Act→Observe→Adapt↺

The counterevidence doesn't refute the thesis. It defines it. The gains go to people who bring judgment, domain knowledge and a clear sense of what good looks like, and who work in a structure that keeps a human at the decisions. That describes many of the people carrying unbuilt things, though not all: they have spent years understanding a problem, and what they have lacked is the capacity to act on it. For those without that judgment, the evidence says AI can make things worse, which is exactly why structure and a human check matter.

09The new scarcity

Every technological shift removes some scarcities and creates others. As certain kinds of execution become cheap, the things that cannot be made cheap become relatively more valuable: purpose, judgment, domain expertise, trust, taste, relationships, accountability, and the courage to ask someone for money, time or belief.

For decades we have talked about democratizing information, then intelligence. Those were real achievements. But information you can't act on doesn't change your life, and an answer is not an action.

Whether that happens is not determined by the technology. The same tools that let one person build a training facility let a large company do far more with far fewer people. Access is uneven, and benefits can concentrate. Democratizing agency, if it happens, will happen because people and institutions deliberately put new capacity in the hands of those who have long had conviction without the means to act on it.

10Yours probably isn't baseball

Community Academy happens to be baseball. Yours probably isn't.

Maybe it's the business you've described for a decade. A foundation. A product. A clinic intake that doesn't lose people. A trade school. A service you know your industry needs. A body of research. Something for your child. Something carrying your parent's name. An idea your employer will never build.

Historically, the next thought was: how could I possibly build that? That question had a real, often crushing answer, measured in hires, years and dollars.

Increasingly, the honest next thought is different: what would happen if I finally tried a small, real version? Not the whole institution. The first piece. A page with a way to sign up. A pilot with five people. A prototype shown to ten. Each costs a fraction of what it used to and teaches more than another year of wondering.

Many of those first versions will fail. That's fine. The economic meaning of an idea changes when testing it becomes cheap: it stops being a lifelong "someday" and becomes an experiment you can actually run.

11What your organization never built

Organizations carry unbuilt things too, and the same economics apply.

For an individual the question is: what have you always wanted to create? For an organization it is: what should we already be capable of doing? The client brief that should assemble itself before every meeting. The internal tool requested in three consecutive planning cycles. The service line nobody had capacity to launch. The knowledge that lives in one person's head and leaves when they do.

These rarely go unbuilt for lack of ideas. They go unbuilt because the organization's own capacity is consumed by running what already exists. Large organizations have more capacity than individuals, but they ration it through budgets, queues and priorities, and the things that never make the cut are often the ones everyone agrees should exist. Representative data suggest most firms are still at the beginning: those that use AI mostly use it in three or fewer business functions, and mostly to augment existing tasks rather than replace workers.²⁵

The method is the same for a company as for a person: define the capability precisely, map what it actually requires, decide what can be carried, what must be borrowed and what must stay human, build the smallest real version inside the tools people already use, and keep a person at every decision that matters.

12The question

I started Jubilee because I had lived this argument before I could articulate it. Most of the work behind this essay, the website you're reading it on, and much of Community Academy's organizational layer was produced with AI under my direction. The judgment, the decisions and the responsibility are mine. I can't take on many projects, and I don't want to pretend otherwise; Jubilee is built to put my limited hours exactly where a person matters, and to let the machines carry the rest.

What Jubilee does is simple to describe. You bring the thing you never built. We help define it, decompose it into the organization it actually needs, decide what can be carried by AI and software, what should be borrowed and what must stay yours, and then build the smallest real version of it, on accounts you own, with a person approving anything that matters. The first step is free and takes about six minutes. The second is one week. The details are here. The longer research argument behind this essay is in Jubilee's white paper.²⁷

But this essay isn't really about Jubilee. It's about a question most people stopped asking because the answer always seemed to be no.

For most of your life, the next question after "what should exist?" had to be whether you possessed the time, the money, the expertise or the people required to create it. Increasingly, you don't need all four.

That changes more than work. It changes who gets to act on conviction.