TLDR. One passage from the middle of this essay summarizes the whole thing well: “Producing a claim and warranting a claim are different economic objects, and the second one is the scarce one. AI just made the first one free. So scarcity did what scarcity always does (Christensen): it migrated - all of it - into the second. That is the Coasean inversion…The bottleneck of the agreement economy was never intelligence. It was trust. It has always been trust.”
Buy a house, and you will notice something strange.
The hard part is over quickly. You find the house. You agree on a price. At that moment there exists a willing buyer, a willing seller, and a standing house. In a frictionless world the transaction is complete. In this world, what follows is six weeks and several percent of the purchase price. A title company reads decades of county records to confirm the seller owns what he is selling. An appraiser confirms the house is worth what you have already agreed it is worth. An inspector confirms the house is what it appears to be. An escrow agent holds the money, because neither side trusts the other to move first. A notary confirms that you are you. A clerk records the deed. An insurer sells you a policy against the possibility that the title company read the records wrong.
Nothing is built in those six weeks. Not a single nail is driven. And yet every one of those people is paid, and paid well, and paid for exactly one thing: doubt. Not the house - the house was finished before you arrived. What you are buying, line item by line item, is the resolution of doubt about the house, the seller, the money, and yourself.
Here is the claim the rest of this essay is built on. Every exchange in an economy carries two costs: the cost of making the thing, and the cost of agreeing about it. We measure the first cost obsessively - it is what productivity statistics are - and we barely see the second, because it hides inside the first. The title company's fee enters GDP looking exactly like the carpenter's. But the two costs have entirely different natures, entirely different histories, and, I think, entirely different futures. This essay is about the second cost: what it is, how it quietly became half of everything, why fifty years of attempts to automate it have failed, and why the technology that everyone now expects to finally collapse it will not do so in the way they expect.
The second cost got its name from a twenty-one-year-old.
In 1931, Ronald Coase - then an undergraduate at the London School of Economics, and at the time a socialist - won a traveling scholarship to the United States and spent a year touring American factories and businesses.1 He went with a question that embarrassed the economics he had been taught. The textbooks said the price mechanism was a marvel of coordination: prices aggregate everyone's knowledge, resources flow to their best use, no planner required. Fine, said Coase. Then why do firms exist? Why is the economy full of these islands of central planning - organizations in which nobody haggles, nobody bids, and a manager simply tells people what to do - floating in the ocean of the market? If the market coordinates so beautifully, why does anyone opt out of it? An economist of the day, D. H. Robertson, had marveled at the same sight: islands of conscious power in an unconscious sea, "like lumps of butter coagulating in a pail of buttermilk."
Coase came back with the answer, delivered it in a lecture in Dundee in the autumn of 1932, and published it in 1937. The answer is that using the market is not free. Before any exchange happens, someone must discover who is selling and at what price. Then the parties must negotiate, draft, and agree. Then each must confirm that the other performed - that the goods are what was promised, that the payment cleared, that the promise sticks. These frictions - later named transaction costs - are the toll the market charges for its own use. A firm exists wherever doing the work by command is cheaper than doing it by contract, and the boundary of every firm on earth sits exactly where those two costs cross. It is one of the great one-move explanations in all of social science: the entire architecture of the economy - what is a company, what is a market, where one ends and the other begins - falls out of a single quantity, the cost of agreeing.
The move I want to make with Coase's concept requires taking it apart. Economists after him decomposed transaction costs into three parts: search costs, the cost of finding your counterparty; bargaining costs, the cost of reaching terms; and enforcement costs, the cost of verifying performance and making the promise stick.2 Three legs, one vastly more important than the others.
For fifty years transaction costs remained what Coase himself ruefully called his article: much cited and little used. A concept, not a quantity. Then, in 1986, two economists - John Wallis and Douglass North - did something almost nobody had thought to do. They counted.
Their method was blunt and honest: go through the economy and separate the people who make things from the people who arrange the making. The lawyer, the accountant, the broker, the banker, the insurance adjuster, the purchasing agent, the clerk, the inspector, the manager whose job is coordination rather than production - all the occupations whose output is agreement. Call that the transaction sector, and measure it across a century of American life.
In 1870, the transaction sector was roughly a quarter of GNP. By 1970, it was over 45 percent.3
Sit with that. Somewhere in the twentieth century, without any announcement, the American economy crossed a line: it began spending more on agreeing about things than on many of the things themselves. And the detail buried inside the number is better than the number. Roughly half of the growth did not come from law firms and banks. It came from transaction workers inside ordinary companies - the purchasing department, the billing office, compliance, accounts receivable - which grew from about 2 percent of GNP to about 10. Bureaucracy, it turns out, is not a disease that infects companies. Bureaucracy is the transaction sector wearing a company badge. The modern corporation is a thin layer of production wrapped in a thick and thickening layer of agreement.
Why did it grow? The economist William Baumol supplied the mechanism, with the most famous string quartet in economics.4 A Beethoven quartet took four musicians and forty minutes to perform in 1826, and takes four musicians and forty minutes today. Two centuries of staggering productivity growth everywhere else, and none here - so the relative cost of the quartet rises forever, not because it got worse but because everything around it got cheaper. Now apply that to agreement. Since 1870 we have automated making with a ferocity unmatched in human history: half the American workforce farmed then; around two percent does now, growing far more food. Steel, textiles, transport, computation - the making side of the economy has been mechanized a hundred times over. And the agreeing side? Reading a contract, judging a claim, checking an invoice, verifying a counterparty - in 1970 this was done the way it was done in 1870: by a person, reading, at person speed and person cost. When you automate one half of an economy and leave the other half handmade, the handmade half must eat the economy. It is arithmetic.
So the center of gravity of the developed world moved, over one century, from making to agreeing - and nobody held a parade or a funeral, because the entire migration was booked as production. We became the first civilization in history that mostly does not make things. We agree about them.
Now the question that decides everything downstream: of the three legs - search, bargaining, verification - which one is holding up the weight?
Douglass North, who spent his career on this, gave a flat answer: the core of transaction costs is measurement and enforcement. The cost of knowing what you are actually getting, and the cost of making the promise stick. Strip away the paperwork and the professions, and the transaction sector is one thing at industrial scale: it is the manufacture of trust between strangers.
History has run the experiment in both directions, and the results are not subtle.
Run it downward first: remove verification from a market and watch what happens. In 1970 George Akerlof published a short paper about used cars.5 The seller of a used car knows whether it is sound; the buyer cannot verify it. So the buyer discounts every car for the possibility that it is a lemon. But at that discounted price, the owners of good cars refuse to sell - so the average car on the market gets worse, so the discount deepens, so more good cars withdraw. The market does not become expensive. It unravels. Akerlof's point, which three journals rejected before a fourth printed it, was that verification is not a fee charged on top of a market. It is the floor the market stands on. When it fails, the market does not limp. It ceases to exist.
Now run it upward: make verification cheap and watch what happens. Chicago, 1856. Before that year, a sack of wheat was not a commodity; it was a story you had to check. Grain traveled in sacks under its farmer's name, and every sale was a bilateral negotiation over the contents of those particular sacks - this farm's reputation, this harvest's quality, this buyer's inspection. Then the Chicago Board of Trade began sorting wheat into named grades - spring, red winter, white winter; No. 1, No. 2 - certified by inspectors, and within a few years backed by state authority.6 The grade did something profound: it decoupled the wheat from the farmer. Once a bushel was certified No. 2 spring, it no longer mattered whose bushel it was. Wheat stopped moving in sacks and started moving as a golden stream through steam-powered elevators, and the elevator's receipt - a paper claim on "5,000 bushels No. 2 spring" - became as good as any other identical receipt, which is to say it became a kind of money. And then the astonishing thing happened. By 1865 the Board had formalized the futures contract: a market not in wheat but in promises of wheat, traded by men who would never see a kernel. Understand what that is. A market in promises cannot exist when promises cannot be verified - a future on unverifiable goods is worth nothing. The grade did not lower the cost of selling wheat. The grade called a new market into existence, a market that had been not expensive but impossible.
I have argued elsewhere that capital cannot flow to what it cannot price (see: How Capital Works). This is the layer beneath that one: nothing can be priced that cannot be verified. The grade came before the future. It always does.
So the answer is in, from both directions. Search and bargaining are riders. Verification is the load-bearing leg. Make it impossible and markets vanish; make it cheap and markets appear that could not previously exist. This point, I believe, will define the next hundred years of capital markets.
We have been trying to automate agreement for half a century. The attempts have a family resemblance, and a family graveyard.
The design was always the same. Since machines cannot read, and agreement lives in language, the plan was to remove the language: get every party to agree, in advance, on a rigid machine-readable format - a schema - and then let computers exchange it cheaply forever. This idea has been implemented with total sincerity for fifty years. EDI and the X12 standards in the 1970s, so that purchase orders and invoices could flow between mainframes. EDIFACT for world trade. HL7 so that hospital systems could talk to each other. XBRL so that financial statements could be parsed. Committees, versions, working groups, decades.
And it failed the same way every time. The standing joke among health-data engineers is that once you've seen one HL7 interface, you've seen one HL7 interface. EDI still frays into fax machines and PDFs at its edges. Every standard fragmented into dialects, because reality kept producing cases the schema had not anticipated, and every implementation patched it differently.
The diagnosis is worth stating precisely, because it is the pivot of this essay. Standardization is prepaid agreement. A schema is itself a negotiation - conducted among all parties, before any transaction, binding forever, amended by committee at the speed reality changes, which is faster than committees move. The scheme replaced a million small transaction costs with one gigantic ex-ante transaction cost and called it progress. It works in exactly one circumstance: when the world is simple enough to fit in the box. The shipping container is the triumphant case - forty feet of standardized steel that remade world trade, precisely because a steel box does not have edge cases.7 The wheat grade is another: a few names were enough to describe a bushel. But agreement in general does not fit in a box, because agreement in general lives in language - in contracts, policies, statutes, exceptions, amendments, cover letters - and language is exactly the thing the schema was built to eliminate.
The cryptographer Nick Szabo gave the two sides their proper names: wet code and dry code.8 Law, contracts, and judgment run on wet code - flexible, interpretive, executed by expensive humans. Machines run dry code - rigid, exact, executed for free. The tragedy of the modern economy in one sentence: everything valuable in commerce was written wet, and for fifty years the only way to automate it was to force it into dry molds in advance - and drafting the mold cost more than the friction it saved. Even the boldest recent attempt, the smart contract, only sharpened the problem: dry code that enforces itself flawlessly, but blindly - every term must still be translated in advance, and some trusted party must still attest to what actually happened out in the wet world. The famous "oracle problem" of blockchains is nothing but the verification cost, undead, standing at the door of the vending machine.
And then - this is the part of the story that should discipline every prediction currently being made - we ran the full-scale experiment. The internet arrived, and serious people announced the collapse of transaction costs. Clay Shirky made the argument at book length in 2008: Coase's logic implies a floor, below which lie activities valuable to someone but too expensive to organize; collapse coordination costs and the floor breaks, and organizing-without-organizations erupts.9 And he was right! The floor broke. What poured through was Wikipedia, open source, Flickr pools, flash mobs - coordination without stakes. Nobody needs escrow to pool photos of a parade. But above the floor, in the contract economy - the half of GDP - nothing moved. Legal, compliance, audit, and administration grew straight through the internet decades. The reason is now visible: the internet collapsed the cost of communication, and communication is not verification. Email made talking free, and talking free is not trusting free - in fact it manufactured an unprecedented volume of claims to be verified by the same old handmade machinery. Round one of the Coasean collapse was run twenty years ago. The sector it was supposed to collapse ended the round larger than it began.
Which brings us to the machines that read, the thing everyone and their mother can’t stop talking about - LLMs.
Strip the marketing from the current moment in AI and one economically primitive fact remains: for the first time, wet code is machine-tractable. A language model can ingest the contract as it actually arrives - the scanned PDF, the emailed amendment, the fifty-page policy in whatever format some adjuster saved it in 2014 - and act on its contents. Nobody had to agree on a schema first. The fifty-year-old problem has inverted: instead of every party agreeing on the format before the transaction, ex ante, a machine interprets whatever shows up, ex post. Standardization has stopped being the price of admission to automation.
The fashionable conclusion follows within seconds, and you have already heard it, because it is currently being repeated everywhere: transaction costs now go to zero. Software agents will search every counterparty, negotiate every term, draft every contract in milliseconds; the firm - which exists, remember, only because contracts were expensive - dissolves into a swarm of market transactions; Coase's boundary gets redrawn overnight. The Coasean collapse. Every step sounds inevitable.
It is wrong. Not wrong at the edges - wrong at the joint, and wrong in the exact way this entire history predicts.
Run the decomposition one more time, because the three legs do not move together.
Search: collapsed. A model can canvass more counterparties in a minute than a procurement department canvassed in a year. Bargaining: collapsed. Terms, drafts, redlines, in seconds, at any volume.
Verification: look closely at what just happened to it. A language model is a probabilistic engine. Its output is fluent, plausible, and unwarranted - it produces the most likely continuation, not the certified one, and it does not know the difference. Which means that every artifact it generates arrives in the economy as a claim. And claims must be checked. We have just made the generation of agreement-shaped text free, while the checking of it costs exactly what it cost before - actually more, for two reasons. The volume of claims has exploded, because generation is free. And the signal that checkers relied on for triage has been destroyed, because fluency used to be evidence of effort and now is evidence of nothing. This is Akerlof's market, rebuilt at planetary scale: a fluent unverifiable claim is a lemon, and we have constructed the lemon firehose. An economy of cheap, confident, uncheckable claims does not have lower transaction costs than an economy of expensive, trusted ones. It has higher ones. You can watch the inversion in any office that has adopted these tools: the phrase "human in the loop" is spoken as reassurance, but listen to what it actually says - the loop is the transaction cost, relocated, and the human is standing in it.
Here is what makes this more than an engineering complaint. The asymmetry that just seized the economy was already known - discovered independently, in four fields that never spoke to each other.
The complexity theorists found it in the structure of computation itself: the P versus NP question, the deepest open problem in the field, is precisely the question of the gap between finding a solution and checking one. The entire class NP is defined by a single property - that a solution can be verified cheaply provided it comes with a certificate, a witness, a proof. Finding is hard; checking is easy - if and only if the answer carries its evidence.10 The economists found it in the structure of exchange: Coase's toll, decomposed by North into measurement and enforcement - the cost of knowing what you are getting. Akerlof found it in the structure of information: no verification, no market - the checking is the floor. And Szabo found it in the structure of law: wet code is cheap to write and expensive to trust; dry code is the reverse. Four fields, four vocabularies, no contact - one law. Producing a claim and warranting a claim are different economic objects, and the second one is the scarce one.
AI just made the first one free. So scarcity did what scarcity always does (Christensen): it migrated - all of it - into the second. That is the Coasean inversion. The technology that was supposed to collapse the transaction sector has instead concentrated the entire transaction cost into its verification leg. The bottleneck of the agreement economy was never intelligence. It was trust. It has always been trust.
So how does verification actually get cheap? Not by making the generator smarter - a smarter generator produces better claims, which are still claims. The computer scientists, characteristically, solved this decades ago, for exactly this problem shape: an untrusted generator producing artifacts that a recipient must nonetheless trust.
In the late 1990s, George Necula and Peter Lee introduced an idea called proof-carrying code.11 The problem: how can a computer safely run a program written by someone it does not trust? The standard answers were to vet the author (expensive, fallible) or to watch the program's every move at runtime (expensive, forever). Necula and Lee's answer was neither. Make the untrusted author do the hard work: require every program to arrive carrying a mathematical proof of its own safety, which the recipient checks mechanically, in a moment. The generator can be arbitrarily untrustworthy - it does not matter, because trust never attaches to the author at all. It attaches to the certificate. This is the P-versus-NP asymmetry picked up and turned from an obstacle into an architecture: let generation be hard, opaque, even hostile - as long as the artifact carries its witness, verification is cheap.
Now translate that into the agreement economy, because the translation is exact. The role of the model is not to be the judgment - probabilistic every time, checked every time, trusted never. The role of the model is translation. Point it at the wet code - the statute, the fee schedule, the sixty pages of terms - and have it compile that language, once, into dry code: explicit, deterministic logic that a human expert can audit, once, against the source text. From then on the dry code executes - exactly, identically, forever - and produces alongside every answer its complete derivation: these inputs, these rules, these steps, this result. The certificate. Checking a derivation against a rulebook is clerical; any auditor, or any machine, can do it. The model, in other words, is the compiler, not the runtime. And with that one move the economics of trust flip: verification stops being a cost paid on every transaction and becomes a fixed cost, amortized across every transaction the compiled rule ever touches. That is the entire trick. That is what it actually looks like when a transaction cost collapses - not when the claims get smarter, but when the claims arrive with proofs.
And once you see the trick, you see that civilization has been performing it - slowly, expensively, in the only material available - for five hundred years.
Double-entry bookkeeping, codified by Luca Pacioli in 1494: record every transaction twice, as debit and credit, so that the books verify themselves - error announces itself as imbalance. A self-checking ledger. Goethe has a merchant in one of his novels call it among the finest inventions of the human mind, and the sociologist Werner Sombart went further, arguing that capitalism itself was unthinkable without it.12 The wheat grade: a certificate stapled to a commodity. The audit: verification rented by the hour, sold with a signature that bears liability. And the clearinghouse - which is the case worth telling, because it shows what happens when the old machinery hits its limit.
In 1968, Wall Street was killed by paperwork. Not metaphorically. Trading volume had roughly tripled in a few years, to over twelve million shares a day, and every share still settled the way it had settled since the Dutch: a physical certificate, carried by a messenger, verified by a clerk. The back offices drowned. At the peak, billions of dollars of transactions sat unsettled; certificates were misdelivered, lost, and stolen outright; and the New York Stock Exchange - the beating heart of American capitalism - closed every Wednesday for the second half of 1968 so the clerks could catch up. It was not enough. By 1970, roughly a hundred brokerage houses had failed or been absorbed.13 And the fix, when it came in 1973, was not faster messengers or smarter clerks. It was the Depository Trust Company: put the certificates in one vault and stop moving them at all; replace the movement of paper with entries on a single ledger; replace thousands of bilateral verifications between counterparties who did not trust each other with one institution whose entire function is to be trusted. A clearinghouse is not a convenience. A clearinghouse is verification turned into an institution.
That is what every great agreement technology in history actually is - the ledger, the grade, the audit, the clearinghouse: verification machines, built out of paper, procedure, and people. The transaction sector - the half of GDP - is simply what those machines cost to operate. We did not stop building them because the idea stopped working. We stopped because people were the only material the machines could be made of, and people do not amortize. Every verification a person performs is performed again, by a person, at full price, forever. Proof, for the first time, does not have that property.
What happens next follows from running Coase in reverse.
His boundary sits where the cost of contracting crosses the cost of command - which means every institution in the agreement economy is sitting on a floor set by the price of trust, and the price of trust is about to change by orders of magnitude for every task where the rules are written down. The title search, the audit, the escrow, the rating, the claim adjustment, the compliance review: each of these is, at bottom, a computation currently performed by a profession. The transaction sector's wages - the accumulated salary bill of half the economy - are not going to vanish; they are going to reprice, the way handweaving repriced when the loom arrived, the way the scribes repriced when print did. That number, measured in 1986 and hiding in plain sight since, is why this transition is larger than the people forecasting "AI productivity gains" imagine. They are measuring the making half. The event is in the agreeing half.
But the cost side is the smaller half of the prediction, and every case in this essay says so. The grade did not make wheat cheaper to sell; it summoned the futures market out of nothing. The clearinghouse did not make settlement cheaper; it made modern volume possible - billions of shares a day where twelve million had broken the system. Cheap verification never merely discounts existing trade. It admits trade that was priced out of existence: exchanges that never happened because checking cost more than the deal was worth; promises too small, too fast, or too numerous to trust; counterparties who could never afford to prove themselves. The largest markets of 2050 will be markets that do not exist today - not because the goods don't exist, but because the promises can't yet be checked. There was no market for a promise of wheat until there was a grade.
The longest view is this. Trust has had two regimes in human history. The first was personal: you traded with kin, neighbors, and men whose reputations you could reach - and the world stayed poor, because the circle of trust stayed small. The second was institutional: law, audit, brand, clearinghouse - a vast apparatus that let strangers trade at continental scale, and whose operating cost is precisely the transaction sector, the standing overhead civilization pays so that strangers can cooperate. Douglass North spent his Nobel lecture on the claim that the West grew rich exactly when it built the institutions that made impersonal exchange possible. The third regime is now visible: computational trust - trust that attaches to proofs rather than to persons or institutions, verification as a property of the artifact instead of a service performed on it. Each regime did not replace the previous one's trades; it unlocked trades the previous one could not carry.
So here is the lens, since every essay should leave one. For the next twenty years, look at any cost in your world - any fee, any department, any six-week delay - and ask one question: is this paying for judgment, or is it paying for doubt? Judgment is real discretion under real uncertainty, and it will remain scarce and become more valuable. Doubt is a rule that is already written down, checked slowly by hand because no one has yet compiled it - and doubt is about to become an engineering problem. Most of what your eye lands on is doubt wearing judgment's clothes. The couple at the closing table were never paying for the house in those six weeks; the house was finished before they arrived. They were paying the standing tax that strangers charge each other for the impossibility of proof. That tax has been collected, in one form or another, since the first two people who didn't know each other tried to trade - and it is, at last, the most interesting number in the world to watch fall.