Coase and Claude: Transaction Costs in the Age of the Transformer

Tom Brown is Senior Counsel at Paul Hastings, specializing in FinTech law with deep expertise in payment systems, digital assets, and financial services regulation. He previously worked at Visa during its corporate restructuring and IPO process.

Two Crossings

In the midst of the Great Depression, a young scholar crossed the Atlantic to study why firms, not markets, were the locus of so much economic activity in industrial America. Ronald Coase spent the 1931–32 academic year in the United States on a traveling scholarship touring Ford and General Motors. He came to solve a puzzle for which classical economics had no answer. If markets were as superior to central planning for coordination of production as the textbooks claimed, why did so much production happen inside the walls of a single firm, coordinated by what looked suspiciously like central planning, rather than being negotiated transaction by transaction among independent contractors bidding against each other in the open market? 

Coase answered that question with an essay published in 1937, “The Nature of the Firm,” explaining that using a market is not free. Finding a counterparty, negotiating a price, verifying quality, and enforcing a bargain all cost something. When those costs are high enough, it becomes cheaper to pull the activity inside a firm and coordinate it by command instead of by contract. The firm, in Coase's telling, is not the natural unit of economic life. It is what remains once markets become too expensive to use.

In 1994 Jeff Bezos abandoned a promising career in finance to sell books over the internet. Bezos's insight ran the logic in the opposite direction. If firms exist because markets are costly to use, then anything that drives those costs down should shrink the firm and expand the market. The internet was exactly that kind of technology for one narrow category of cost: the cost of physically locating and distributing a generic, catalogable good. By Christmas 1998, the trajectory built on that insight was already legible. Amazon had been public for a year and a half and had just added music and video to a catalog that started with nothing but books. Any business whose entire value proposition amounted to “we are conveniently located and we have inventory” was, whether it knew it yet or not, already being disassembled. 

Coase conceived of transaction costs to explain why firms exist. Bezos demonstrated how little you need a firm as transaction costs decline. The agentic age requires us to answer a different but related question: if foundation models have driven the marginal cost of evaluating information close to zero, why do we rely on market intermediaries — brands, retailers, brokers, and marketplaces of all kinds — rather than going direct to the source? Coase’s logic also explains these institutions, once we recognize information costs as a distinct category of transaction costs. The strongest evidence is that the revolution has already happened in a place the theory would predict — just not in a place most people think to call “commerce.”

Proxies as Firms in Miniature

Coase's transaction costs came in several flavors: the cost of finding a counterparty, the cost of negotiating terms, the cost of enforcing a bargain, and the cost of simply knowing enough to trust the other side of the trade. That last category — call it information cost — has quietly organized an enormous share of commercial life, in institutions that have nothing to do with the boundary of a firm.

Building trust around individual human judgment is expensive. So commerce has always relied on proxies that let a buyer make a decision without directly verifying the thing they actually care about. Think a FICO score. Two of the most important are brand and identity, and each is best understood as a small, purpose-built solution to exactly the problem Coase described.

Brand is a proxy for quality, provenance, and consistency. When the underlying attributes of a product are costly to verify directly — is this handbag really lambskin, was this coffee really grown at altitude, will this appliance actually last ten years — a recognizable brand lets a buyer economize on that verification. For some categories, brand carries independent value even after quality is verified — a Rolex or a Hermès Birkin bag communicates something about the owner that has nothing to do with utility. But for a great many categories — dental floss, wheat, a phone charger — brand is doing almost pure information work.

Identity is a proxy for trustworthiness or its inverse, licit purpose. Know-your-customer regimes, credit scores, professional credentials, even a merchant's business license, exist because directly observing someone else's intentions is hard, and a stable identity lets an institution infer intent from history instead. The entire apparatus of consumer financial regulation is, in a sense, a machine for converting “who is this person” into “can this transaction be trusted.”

Artificial intelligence attacks this structure at its foundation. It is a general-purpose technology for driving the marginal cost of evaluating information toward zero. An agentic tool can read an ingredient list, a certification, a supply-chain audit, and ten thousand verified reviews as easily as it can recognize a logo. It can evaluate a counterparty's actual track record rather than inferring trust from a name. This is not automation of a decision that a human was already making efficiently. It is the elimination of the specific cost that made a proxy worth paying for in the first place.

This does not mean every proxy collapses at once, or for the same reason, or that collapse is uniformly good for consumers. Coase's framework is instructive here, too: negotiation costs and enforcement costs don't disappear just because information costs fall, which is part of why fraud remains a live problem even in a world of cheap verification. But it does mean the categories most exposed to agentic disruption are the ones where the proxy was doing pure information work with nothing else attached — and the categories most protected are the ones where the proxy is not substituting for hidden information but is the product.

Where the Robots Already Live

If AI is a technology for collapsing information costs, the place to look for its earliest and most complete effects is wherever those costs were already lowest to begin with — markets where the “product” had already been reduced to a number, with no brand, no identity, and no location left to verify. That description fits the currencies, stocks, bonds, and derivatives markets, where one firm demonstrated for almost 40 years that well-trained machines can read numbers better than any human.

Renaissance Technologies launched the Medallion Fund in 1988. Jim Simons, the firm's founder, was a mathematician and a Cold War codebreaker at the Institute for Defense Analyses before he ever traded a security. He built his research team not out of Wall Street veterans but out of mathematicians and scientists recruited from his own circles at the IDA and from Stony Brook University, where he had chaired the mathematics department. In 1993 the firm made two hires that, in hindsight, look prophetic: Peter Brown and Robert Mercer, two speech-recognition scientists poached from IBM, where they had spent years treating language not as a set of grammatical rules but as a probability problem — modeling the likelihood of the next word in a sequence given everything that came before. That is, for all practical purposes, the same intellectual move that underlies the large language models running today's agentic tools. Renaissance was doing next-token prediction on markets before anyone was doing it on language.

The results were astounding. From 1988 through 2018, Medallion averaged roughly 66% in annual gross returns, with only a single losing year across three decades — a record that defies claims about the efficiency of capital markets. What Renaissance actually built was a machine for identifying, with obsessive statistical rigor, the moments when the current price of an already-financialized instrument did not reflect the composite of its likely future states, and then executing on that gap, relentlessly, at a scale and speed no human trading desk could match. The product Renaissance was buying and selling had already been stripped of everything except a number and a probability. There was no brand to a particular block of currency or later equities. There was no neighborhood grocer's intuition standing between the algorithm and the asset. The verification costs Coase would have recognized — the cost of knowing who you were dealing with, or whether the thing itself was what it claimed to be — had already been stripped away by the structure of the market itself, decades before anyone called the result “agentic commerce.” What was left wasn’t a costless market. It was a market where the only remaining cost was predicting the next number, which is exactly the cost Renaissance built a business on exploiting.

Most agentic-commerce commentary misses this point: the first, largest, and most successful deployment of autonomous purchasing and selling decisions in human history has generated over one hundred billion dollars in trading profits while hiding in plain sight because nobody thinks of a hedge fund's execution engine as “commerce.” It is. It is simply commerce in the one category where the transaction costs were low enough to identify and trade on information asymmetries. 

Finding the Edge

The Renaissance template did not stay confined to public securities. Over the past decade it has migrated into markets for collectibles and other goods with a meaningful financial component — sneaker resale, for instance, now runs through exchanges with genuine bid-ask spreads, order books, and authentication infrastructure that would be recognizable to anyone who has traded a security. Automated buying and selling has also spread into event tickets, restaurant and hotel reservations, and airline fares. There are several distinct mechanisms at work here, and they reward different kinds of edge — which is another way of saying they exploit different transaction costs.

The first is genuine prediction arbitrage, closest to the Renaissance mechanism. A model forecasts a future state of a price that is allowed to float, and profits from being right before the market catches up. This exploits what might be called a cost of time: the cost of not knowing, today, what tomorrow's price will be. Airline fare-prediction tools are the clearest consumer-facing example. Sophisticated models ingest fare histories, seasonality, and route-level demand signals in order to forecast whether a given fare is likely to rise or fall before departure.

The second mechanism is speed arbitrage against artificially fixed supply. This exploits a cost of allocation rather than a cost of time. Ticket and restaurant-reservation bots are not forecasting a future price; they are racing human beings to a resource that has been mispriced — a concert ticket at face value, a reservation that costs nothing to book, a tee time at a public golf course — and reselling into a secondary market that already knows the “true” price. The edge here is not insight into the future. It is raw reaction time in the present, applied against a rule that a human institution chose to impose.

Just as buyers have used machine intelligence to identify and exploit these arbitrage advantages, sellers have deployed it to close the same gaps from the other side, exploiting a third cost: the cost of setting a price correctly in real time rather than in advance. Airlines pioneered the practice following price deregulation in the 1970s. It proliferated to other hospitality categories through the 1980s and 90s. The San Francisco Giants became the first professional sports team to test dynamic ticket pricing in 2009. By 2010, they had rolled it out to all single-game tickets. By 2012, they applied it to all tickets to all games, including the playoffs. By 2015, it was a universal phenomenon across Major League Baseball. Today, every fixed-stock commodity or service industry, from ride hailing to property management to nursing homes, relies on machine intelligence to set price. Where a buyer's agent profits from finding a gap between price and value, a seller's agent profits from closing that gap before anyone else can find it. Both are answers to the same underlying question Coase posed about the firm: given a specific cost of coordinating exchange, who can absorb it more cheaply, and who gets paid for doing so.

What Hasn't Been Agentified

Agentic commerce is not repeating the story of e-commerce so much as completing a different one that started at least a decade earlier. Agents are rendering a different set of proxies — brand, identity, the comfortable habit of a known counterparty — irrelevant, in exactly the categories where those proxies were pure information-cost solutions and nothing more. Coase mapped the boundary between the market and the firm focusing, principally, on the cost of contracting. AI moves a different boundary: the one between transactions that require informational intermediaries and those that do not. 

The categories most exposed to agentic commerce are not the ones that feel loosely “financialized.” They are the ones where the core product has already been reduced to a priced, comparable instrument, where the proxy layer — brand, identity, location, human relationship — solved an information problem and is now decorating a transaction that a sufficiently good model can execute better and faster on its own. The categories that are safest are not the ones furthest from technology. They are the ones where the proxy was never standing in for hidden information at all, but is an irreducible part of what's being purchased — where the human-preference layer, not the underlying commodity, is the actual product, and no amount of cheaper information changes that.

It is worth being explicit about how far this cuts, because the instinct is to stop at brand. Amazon, eBay, Uber, DoorDash, Google Shopping, Booking, and Expedia are not brands in the sense a Rolex is a brand. They are marketplaces. They exist to aggregate and present information that would be too expensive for an individual buyer to gather independently: which sellers exist, what they charge, whether they can be trusted, how their offering compares to the next one over. That is a purely informational service, which means it is subject to exactly the same test as brand, identity, and location. A marketplace does not become obsolete when a single supplier gets better information. It becomes obsolete when the buyer's own agent can do the aggregation, comparison, and trust-verification the marketplace was built to sell, at a marginal cost near zero, without needing the marketplace's interface at all. The theory here is not that brands lose pricing power. It is that the layer of commerce built to solve buyers' information problems — including the platforms that made their name doing exactly that — is most exposed.

Fraud provides another example. Some fraud has always depended on exploiting the cost of verifying information. Counterfeiting, misrepresented quality, and astroturfed reviews all rely on a check being too expensive for the buyer to perform; an agent that can cross-reference certifications, supply-chain records, and review-pattern anomalies at scale erodes the asymmetry that such fraud exploits. But other types of fraud exploit other gaps. Triangulation fraud is the clearest example: an agent poses as a legitimate merchant, delivers a real good to complete the appearance of an ordinary sale, and uses the transaction purely to harvest the paying agent’s credentials for later misuse. The fraud does not need the buyer’s agent to be wrong about anything verifiable. It only needs the transaction to happen at all. And the asymmetry here is structural, not incidental: merchants have decades of underwriting infrastructure, principally through the card networks, for assessing whether a buyer’s intent is legitimate. Buyers, and the agents now acting for them, have no comparable infrastructure for underwriting a seller. The larger point is that large language models do not impact all transaction costs in the same way — they decrease some while increasing others.

Commerce was built, from the beginning of human existence, around a mind capable of representing another's wants and trusting a counterparty it could not fully verify. Machines do not think the way we do. But they do make decisions. At least since the Medallion Fund opened for business in 1988, the party making the purchasing decision on one or both sides of a transaction may not be the kind of mind from which commerce emerged. That is the actual disruption — not that a robot can click “buy.” And although robots are not yet substituting for humans at the checkout, the search and discovery function is increasingly mechanised. Simply put, our role at the table has changed, and with it, the efficient frontier separating what still needs a human-shaped proxy from what no longer does.

This leaves every business with a test that is not comfortable but is, at bottom, Coase's own question asked one layer down. Ask three things. Did your customer choose you because you served as a proxy for something too costly to verify directly, and can a sufficiently good model now verify it instead, faster and without you? Is what you sell already close enough to a number that someone else's model can find the gap between what you're charging and what the moment is actually worth? And when that gap opens, are you the one closing it with a model of your own, or are you waiting to find out, from someone who got there first, what the price should have been? If the honest answer to any of these questions points away from you, you aren't the guest at the dinner table. You are on the menu.

The opinions shared in this article are the author’s own and do not reflect the views of any organization they are affiliated with.

^[1]  The ideas, argument, and editorial judgment reflected in this piece are mine and mine alone. I used Claude (Anthropic) and several other AI-enabled tools (e.g., Google Search) iteratively — for drafting under my direction, restructuring based on my feedback, and fact-checking historical and financial claims. I reviewed, verified, and edited every part of the final text. Many thanks to the Open Banker editorial team for improving the draft. Errors and omissions are mine alone.

Open Banker curates and shares policy perspectives in the evolving landscape of financial services for free.

If an idea matters, you’ll find it here. If you find an idea here, it matters. 

Interested in contributing to Open Banker? Send us an email at [email protected].