Coase, Williamson, and 88 Years of Theory That Explain the AI Bill
Two economists explained why firms exist and how they organize. Their theories remain fully relevant in 2026, now applied to networks where humans and AI agents operate side-by-side.
The 90-Second Version
In 1937, a 26-year-old Brit named Ronald Coase answered a question nobody found interesting: why do firms exist at all, if the market already organizes everything? Because coordinating work in-house comes out cheaper than buying every piece on the open market. In 1985, an American named Oliver Williamson opened up the middle ground and showed that a whole spectrum of arrangements sits between the pure market and the closed firm. Both took home the Nobel. In 2026, the ordinary company has a new tenant living inside the hierarchy, right next to its people: the AI agent. It is a fourth economic drawer that neither Coase nor Williamson saw coming, yet their ruler measures it without complaint. The theory has been ready for 88 years. Putting a dollar figure on what that fourth drawer consumes is the job nobody has sat down to do.
There is one equation that refuses to balance, and it is bothering boardrooms everywhere in 2026. The AI pitch promises to speed up each person's execution by 30% to 50%, and the company buys the promise: a tool on every desk, a generative model behind every workflow. The individual metric does climb. Some processes do move faster. Then the CFO opens the consolidated margin and it has not followed. In some cases it got worse. The gain everyone swears they have keeps vanishing somewhere on the way to the line that counts.
Where did it go? Why doesn't it show up in the number the CFO carries into the board meeting? The answer is not new. It is old enough to be embarrassing. It was written in 1937, in a short essay that changed the economics of organizations for good. It picked up reinforcement in 1985, in a book that became one of the most cited works in all of social science. And it holds in 2026 with one adjustment that matters but does not overturn the theory. The world brought in a new object. The old ruler still reads it.
1937: The Question Nobody Was Asking
Ronald Coase was 26 when he published "The Nature of the Firm" in 1937. The essay asked a question that sounded almost silly to the economics of its day, the kind that looks naive right up until someone notices nobody has an answer. If the market is efficient, if price allocates resources better than any planner could, why do firms exist? Why isn't everything bargained task by task, out on the open floor of the market?
His answer was short and weighed a ton. Buying on the market also costs you, and it costs you before you ever pay the price. It costs you to find a supplier who won't fleece you, to negotiate, to draft the contract, to check that what arrived matches what was agreed, to fight when it doesn't. Every one of those steps eats time, attention, and money. Coase gave the whole pile a name: transaction cost.
The rule that falls out of this is clean. When coordinating a task through the market costs more than coordinating it in-house, the task comes into the firm. When the market comes out cheaper, it stays outside. And the company grows up to exactly the task where it makes no difference, the point where doing it inside costs the same as buying it outside. That tie is the boundary of the firm. It is not the size of an ambition. It is the size of a bill.
The Nobel only arrived in 1991, 54 years after the essay. The delay was not contempt. It was that the 1937 argument took decades to truly digest. By the time it landed, it was already the skeleton of almost everything taught in the economics of institutions today.
1985: Williamson Extends the Skeleton
Oliver Williamson took Coase's argument and pressed it for detail. Fine, transaction cost exists. But how much, in which situation, and how do you actually choose between one way of organizing the thing and another? In 1985 he published "The Economic Institutions of Capitalism," which joined the list of the most cited books in all of social science.
What he brought has three pieces that matter for 2026. The first knocks down a false choice: it is not market or firm, all or nothing. Between the two ends sits a whole corridor of arrangements, long-term contracts, joint ventures, stable supplier networks, franchises. Each one is a point partway between buying everything outside and doing everything inside.
The second piece is the ruler that decides which point in that corridor fits. Three variables. How much of the investment is specific to that relationship and useless for anything else. How much uncertainty hangs over the environment. And how often the transaction repeats. Very specific investment plus high repetition, the firm wins, it pulls the work inside. Both of them low, the market wins, it leaves the work outside. In the middle of the middle, the hybrid arrangement can be the cheapest of all.
The third piece is where the focus lives: the way you govern the relationship, not the product and not the service. How you arrange the economic relationship between the parties so it costs less and returns more. Williamson named the approach New Institutional Economics. He took the Nobel in 2009, sharing it with Elinor Ostrom.
| Arrangement | When it wins the bill | Where the cost concentrates | The usual example |
|---|---|---|---|
| Market | Low asset specificity, low repetition, low uncertainty | Finding a supplier and negotiating every time | Commodity purchase, single hardware supplier |
| Hybrid (long-term contract) | Everything in the middle: specificity, repetition, and uncertainty all medium | Watching the delivery and holding the risk that the other side pulls its own way | Long supply contract, joint venture, franchise |
| Firm (hierarchy) | Very specific investment, high repetition, high uncertainty | Running it all in-house and giving up the push of the market | Core function kept in-house, product team, finance |
88 Years: The Skeleton Held Through Every Wave
The first wave to put the Coase and Williamson ruler to work in practice was the outsourcing of the 1990s. Companies stopped to ask, one activity at a time, which of them was generic enough to push out to the market. Out went basic IT, out went basic bookkeeping, out went operational HR. What stayed inside was the core nobody else did the same way.
The second wave, in the 2000s, sent work out of the country, what the market called offshoring. Same ruler, applied to the map. Wherever the task could be done far away without losing what made it specific, it crossed the border. The theory stayed the same. The boundary just changed address.
The third wave was the platforms of the 2010s. Airbnb, Uber, and the food-delivery apps proved that a hybrid arrangement built on a network could beat the classic firm on specific ground. The contract shrank to micro, turned dynamic, and started being stitched together by algorithm. The middle corridor Williamson had drawn on paper was already there. The technology only made it cheap enough to run at scale.
Across 88 years, the skeleton held through all three. Each wave moved where the boundary sat best. None of them touched the logic of why a boundary exists at all. That stubbornness is what makes the ruler trustworthy, and what makes it worth picking up again now.
2026: The Fourth Structure Nobody Predicted
Now comes the case neither of them could have predicted. The AI agent works inside the firm, but with an economic behavior that fits none of Williamson's three drawers. It is not an outside supplier, it is not on the market. It is not a person with an employment contract, no salary to defend, no career to build, no self-interest to pull. And it is not quite the hybrid in the middle either, because it never sits down to negotiate a single term with the company.
It is a fourth drawer. It lives inside the hierarchy, with extremely high asset specificity, trained on the company's own data, its instructions tuned to each context. Self-interest is zero, it does not stray off purpose because it has no purpose of its own. But output uncertainty is sky-high: it hallucinates, it drifts, it embarrasses you in the rare cases. And its operating frequency runs almost without a ceiling, because running it again costs next to nothing. Add the traits together and you have a partner with no ambition, no laziness, and no judgment, one that will happily repeat a task a million times and fail in a way nobody saw coming.
Williamson did not see it. Coase did not see it. But their ruler measures it all the same. The 2026 question is the 1937 question with one more drawer: now that there is a fourth coordination arrangement inside the firm, where does the optimal boundary fall, and what does it cost to govern all four sharing the same decision?
| Drawer | Who is in it | Where the cost concentrates | How to measure it |
|---|---|---|---|
| Market | One-off purchase, outside supplier | Searching, negotiating, watching the delivery | Price paid + the cost of buying |
| Relational hybrid | Long-term contract with a third party | Watching the deal + risk of the other side pulling its own way | Contract management + recurring audit |
| Classic firm | Your own people working in an organized structure | Coordinating in-house: meetings, decisions, escalation | Payroll + synchronous time spent in collective rituals |
| Agent inside the firm | Model with tools, autonomous agent, automated decision | Inference (cheap) + human repair when it fails (expensive) | Token spend + payroll of whoever calibrates and signs off on the machine |
The fourth drawer did not arrive to replace the other three. It cuts across all of them. In a mid-sized Brazilian B2B company of 500 people, one weighty decision now passes through all four in the same afternoon, with no warning. That is why the bill got hard to read: the spend splits across four arrangements at once, and the report shows it all summed into a single number.
The Firm's Optimal Boundary Moves When the Agentic Edge Gets Cheap
There is a practical consequence of this fourth drawer the AI pitch still hasn't faced head on. If coordinating through an agent keeps getting cheaper, and it is, the boundary of the firm leaves its spot. A task that lived inside the hierarchy can slip into the agent drawer. A task that had gone out to the market can come back inside, now run by an agent instead of a person. The line Coase drew does not disappear. It walks.
And it is walking right now, while you read this. The company that can measure what each of the four drawers costs decides where to put each task with the bill in hand. The one that doesn't measure will find out, two or three years from now, that a competitor arranged its coordination far more cheaply and has been quietly opening up margin. By the time that kind of edge shows up in the results, it is already consolidated and expensive to catch.
The theory delivers the why. Counting every day what each link in that network costs is the operational job still missing. The 4 edges of human-agent coordination in cash is how that work shows up concretely in your operation.
What Changes in the Measurement Problem
Measuring transaction cost has always been the theory's weak heel, and economics knows it. Coase himself wrote in 1937 that this cost is hard to put into a direct number. For decades you could live with the flaw, because the most expensive drawer, the firm, had a cost you could map reasonably well: it was in the payroll and the formal rituals, you could point at it.
The fourth drawer ended that comfort. It slid a slice into the total cost that keeps growing and that escapes the payroll entirely, at least the way we read payroll. The time spent refining prompts is chopped into tiny pieces scattered across every decision. The back-and-forth between people to fix what the AI spat out has no label. And the human repair, for when two agents can't understand each other, falls onto the senior engineering payroll without ever earning a line of its own. The money is going out. It just isn't written anywhere you can add up.
It is the same invisible spend that holds up the central argument of human-agent coordination cost as the invisible vector of AI governance. Coase explains why this cost sits at the center. Williamson explains why it has to be split by arrangement, not summed into one lump, so you can finally read the right order of magnitude for each part.
Regulatory Compliance Answers a Different Question
Some people mix up the economic conversation about governing coordination with the regulatory conversation about complying with AI law. They are two different things. Brazil's PL 2338 and the EU AI Act, in force on a staggered basis starting in August 2026, deal with rights, risk, and algorithmic transparency. Necessary, enforceable, with a deadline on the calendar. But not a word about what it costs, in cash, to coordinate human and machine inside the firm.
They are two different floors of the same building. You can be fully square with the law and still burn millions coordinating blind. What changes in your operation with PL 2338 and the EU AI Act in 2026 is one floor. Governing the cost of coordinating across the four drawers is the other. Tending to one and forgetting the other leaves the company armored on one side and bare on the other.
How to Apply the Theory Without Buying a New Suite
Putting the theory to work in the first quarter does not call for a new tool. Four moves are enough to bring Coase and Williamson into your next results review.
- See which drawer runs each thing today. Take five activities that weigh on the company and say, for each one, which arrangement is handling it. Market, the SaaS purchase, the one-off supplier. Hybrid, the long contract, the partnership. Firm, the team inside. Agent, the generative workflow, the automated decision.
- Pin a cost to each drawer with loaded payroll and tokens.You don't have to nail it on the first pass. What you want here is the order of magnitude, not the fifth decimal place. Rough on purpose is enough to have the real conversation.
- Compare against the company's historical boundary.Is there an activity in the firm today that's a candidate to slip into the agent drawer? Is there an activity that went to the market that would come back inside, now run by an agent? The theory asks for this question to be routine, not an isolated event.
- Reread it every quarter, alongside the forecast.Put the split of coordination cost by drawer on the same dashboard where recurring revenue and margin already live. Quarter by quarter, the boundary walks. Whoever measures, sees it. Whoever doesn't finds out late, and finding out late is expensive.
Frequently Asked Questions
Why go back to Coase in 2026 to talk about AI?
Because in 1937 he answered the question that is weighing on people again. If the market is so efficient, why do firms exist instead of every task being bought loose out there? His answer was blunt: buying outside costs you too. It costs you to find a supplier, negotiate, write the contract, watch the delivery. When coordinating the task inside comes out cheaper than buying it outside, it joins the firm. AI changed that price on both sides at once, inside and outside. So it changed where the optimal boundary of the company falls. The 1937 question still holds. The 2026 answer is a different one.
What did Williamson add to Coase?
Coase said coordinating costs you. Williamson, in 1985, asked how much, in which situation, and how to choose among the ways of organizing the thing. His contribution was to show that it is not just market on one side and firm on the other, with nothing in between. The middle is wide: long-term contract, joint venture, stable supplier network. And the choice among those arrangements depends on three things, how much of the investment is specific to that relationship, how much uncertainty surrounds the environment, and how often the transaction repeats. He won the Nobel for it in 2009, shared with Elinor Ostrom. What he did not see was a fourth arrangement being born inside the company: the agent.
Why does their theory support a fourth structure in 2026?
Because the AI agent fits none of the drawers that already existed. It is not an outside supplier, it is not on the market. It is not a person on the team with a contract, a salary, and a career. It works inside the firm with the face of a supplier and the badge of an employee at the same time. Extremely high asset specificity, because it was trained on the company's own data. Zero self-interest, because it does not negotiate a salary or pull its own way. And extremely high output uncertainty, because it hallucinates and drifts. Williamson did not draw this hybrid, but the ruler he built measures it just the same. The new question is only one: what does it cost to govern that fourth drawer, in cash?
Isn't this just transaction cost theory applied to outsourcing?
No, and the difference is the whole point. Outsourcing uses Coase and Williamson to settle the old make-or-buy: this I do inside, that I buy outside. The agent flips the switch somewhere else. It works inside, but behaves economically like something that is outside. You are not deciding whether to make or buy. You are figuring out how to govern a structure that did not exist in 1985, with the awkward detail that the instrument to measure what it costs is still being built while the bill runs.
Where does a 500-person company start applying this framework?
With a question that fits on a napkin. Take five decisions that weighed on the last quarter. In each one, how much of the total time went to coordinating, and how much went to actually executing? Within the coordinating, how much was human with human, and how much brought in an agent? That alone gives you the order of magnitude of the boundary your firm is on today. The next step is to look at whether the boundary is in the right place, or whether there is activity that would migrate to another drawer, more agent, more third party, more automated decision. The theory is 88 years old. The measurement is the part nobody has done.
The Bottom Line
Coase was 26 in 1937 when he wrote the essay that still explains why firms exist. Williamson was 53 in 1985 when he finished the book that showed how to arrange the economic governance between them. Both took the Nobel in different decades. The theory did not age a day. It only had to open a fourth drawer to fit the partner who showed up in 2026.
Good theory does not lose value when the world changes, it gains a new case to solve. The new case of 2026 is the agent working inside the firm, side by side with the people. The tax on coordinating that hybrid network is being paid right now, chopped into small pieces in every decision, too small for any one of them to draw attention and too large in total to ignore. Whoever measures captures the gain AI promised. Whoever does not will find out, in some board meeting down the line, that the margin never caught up with the productivity pitch everyone bought.
Coase and Williamson delivered the why. The instrument that counts, in cash, what the fourth drawer consumes is the part nobody has built yet. That leaves you a choice about timing: start now, with a napkin and loaded payroll in hand, or wait for the next wave, the regulatory one or the competitor's, to arrive asking for the answer on the day it is already too late.