Ronald Coase

Foreword

By Edo Segal

The number that rewired my thinking was not twenty. Not the twenty-fold productivity multiplier from Trivandrum, though that number changed how I build. The number was forty-five.

Forty-five percent. That is the share of the American economy that two economists estimated, back in 1986, was devoted not to making things but to the process of making deals about making things. Finding the right person. Writing the specification. Negotiating the terms. Reviewing the output. Managing the handoff. Forty-five percent of everything, consumed by the friction between one mind and another.

I read that number and felt the floor tilt.

Because that friction is exactly what disappeared in Trivandrum. When my backend engineer built a user interface she had never attempted, she did not learn frontend development overnight. She eliminated the transaction that used to sit between her and a frontend developer — the spec, the meeting, the review cycle, the inevitable gap between what she meant and what someone else understood. The machine spoke her language. The middleman vanished. Not the person. The cost.

Ronald Coase asked one question in 1937 that built an entire field of economics: Why do firms exist? His answer was so simple it took the profession fifty years to absorb it. Firms exist because using the market is expensive. Every time you need to find someone, negotiate with them, verify their work, you pay a cost that has nothing to do with the work itself. When those costs are high enough, it becomes cheaper to hire people and coordinate internally than to contract for everything on the open market.

That answer explains my entire career. It explains the teams I built, the managers I hired, the specification documents that consumed weeks, the sprint planning sessions that consumed afternoons. Every one of those structures existed to manage a transaction cost.

Now ask what happens when AI collapses those costs overnight.

The answer is not that firms disappear. It is that the boundary between what belongs inside an organization and what belongs outside it moves — fast, far, and in directions that most leaders have not yet mapped. Coase gives you the map. Not a prediction of what will happen, but a framework for seeing why it is happening, and where the ground will settle after the shaking stops.

This book takes Coase's lens and points it at the world described in *The Orange Pill*. The results are uncomfortable and clarifying in equal measure. They changed how I think about every organizational decision I make.

They might change yours.

-- Edo Segal ^ Opus 4.6

About Ronald Coase

1910–2013

Ronald Coase (1910–2013) was a British-American economist whose work transformed the understanding of why institutions exist and how legal rules shape economic outcomes. Born in Willesden, England, he studied at the London School of Economics before spending the majority of his career at the University of Chicago Law School. His 1937 paper "The Nature of the Firm" introduced the concept of transaction costs — the real-world expenses of searching, bargaining, and enforcing agreements — as the fundamental explanation for why firms exist rather than all economic activity occurring through market exchange. His 1960 paper "The Problem of Social Cost" demonstrated that, in the absence of transaction costs, the initial assignment of property rights would not affect economic efficiency, a proposition that became known as the Coase Theorem and reshaped the fields of law and economics. Coase was awarded the Nobel Memorial Prize in Economic Sciences in 1991. He insisted throughout his career that economics must study real institutions as they actually operate rather than idealized models of how they should, a methodological commitment that gives his framework enduring relevance in periods of rapid institutional change.

Chapter 1: The Question That Built a Field

In the autumn of 1931, a twenty-year-old economics student from the London School of Economics crossed the Atlantic on a traveling scholarship to study the structure of American industry. Ronald Coase did not go to America to test a hypothesis. He went because he had a question that his professors could not answer, and that the entire discipline of economics had, remarkably, failed to ask.

The question was this: Why do firms exist?

It was not an idle curiosity. Economics in 1931 was built on the premise that the price mechanism — the system of markets, prices, and voluntary exchange — coordinated economic activity with an efficiency that no central planner could match. Adam Smith's invisible hand, refined through a century and a half of theoretical development, described a world in which independent actors, each pursuing their own interest, produced outcomes that were collectively optimal. The price system allocated resources. It directed labor. It determined what was produced and in what quantities.

If this was true, and the entire profession believed it was, then the existence of firms was a puzzle. Inside a firm, the price mechanism does not operate. Workers do not negotiate the price of each task with their employer moment by moment. Departments do not bid against each other for resources in an internal market. The manager directs. The employee obeys. The coordination that occurs within a firm is coordination by authority, not by price. It is, in miniature, precisely the kind of central planning that economics argued was inferior to the market.

Why, then, would rational actors voluntarily abandon the market mechanism and submit to hierarchical direction? Why would a skilled carpenter join a construction company rather than contracting independently for each job? Why would an engineer accept a salary rather than selling each hour of work to the highest bidder?

Coase spent months visiting American factories, talking to businessmen, studying how actual enterprises organized themselves. The businessmen, he later recalled, found the question peculiar. Of course firms existed. What kind of question was that? The answer seemed obvious to anyone who had ever tried to run a business. The answer was not obvious to economics, which had somehow built an elaborate theoretical structure on the assumption that all economic coordination happened through markets, without noticing that a very large share of it happened inside organizations that operated by entirely different principles.

The paper Coase wrote in response, published in 1937 as "The Nature of the Firm," provided an answer so simple that it took the profession decades to appreciate its depth. Firms exist because using the price mechanism is costly. There are costs to discovering what the relevant prices are. There are costs to negotiating and concluding a separate contract for each exchange transaction. There are costs to specifying the terms of the agreement, monitoring compliance, and resolving disputes when the terms are violated. These costs — which Coase called "the costs of using the price mechanism" and which later economists would label transaction costs — are real. They consume time, attention, and resources. And when they are sufficiently high, it becomes cheaper to organize activities within a firm, under managerial direction, than to transact for them on the open market.

The boundary of the firm, in this framework, is not arbitrary. It is determined by a comparison of costs. The firm expands when the cost of organizing an additional transaction within the firm is less than the cost of carrying out that transaction on the market. The firm contracts when the reverse is true. The boundary sits at the point of equilibrium, where the marginal cost of internal organization equals the marginal cost of market transaction.

This insight, arrived at by a young man who had simply taken the trouble to look at how actual businesses operated, created an entire field of institutional economics. It earned Coase the Nobel Prize in 1991, fifty-four years after the original paper, by which time the profession had finally absorbed the implications of what he had been saying. In his Nobel lecture, Coase observed with characteristic dryness that the reason his work had taken so long to be recognized was that it dealt with the real world, and economists were, on the whole, not very interested in the real world.

The framework Coase built is not a theory of technology. It does not depend on any particular set of tools, any particular mode of production, any particular historical period. It is a theory of boundaries — a general explanation of why some activities are organized within hierarchies and others are left to markets. The boundary moves whenever the underlying cost structure changes. And this is why the framework, built to explain the industrial firms of the 1930s, turns out to be the most powerful lens available for understanding what artificial intelligence is doing to organizations in 2026.

Consider what The Orange Pill documents. In a room in Trivandrum in February 2026, twenty engineers sat down with an AI coding assistant and discovered that each of them could now perform work that had previously required the combined output of a much larger group. A backend engineer built user interfaces. A designer wrote complete features end to end. The boundaries between specializations, which had seemed as solid as the walls between departments, dissolved in a week.

A Coasian economist observing this scene would not be surprised by the emotional reaction — the mixture of exhilaration and terror that Segal describes. But the economist would focus on something more structural. Each of those engineers had, in the space of a few days, absorbed transactions that previously occurred between specialists. The backend engineer who builds a frontend interface has eliminated the transaction that previously occurred between the backend team and the frontend team: the specification, the handoff, the review, the revision, the negotiation over priorities and timelines. The designer who writes features has eliminated the transaction between design and engineering: the mockup, the specification document, the implementation meeting, the gap between what was designed and what was built.

Each eliminated transaction is a transaction that previously justified organizing those specialists within a firm. The firm existed, in part, to manage the coordination between backend and frontend, between design and engineering, between the person who knew what should be built and the person who knew how to build it. When a single individual can perform both functions, the coordination that justified the organizational structure is no longer necessary.

The Coasian boundary of the firm has shifted inward. Not because the firm has become less efficient at what it does, but because the individual has become vastly more capable. The comparison of costs that determines where the boundary sits — internal coordination versus market transaction — has been altered from the outside. The cost of individual production has fallen so dramatically that the threshold at which internal coordination becomes worthwhile has moved to a much higher level of complexity.

Segal describes a twenty-fold productivity multiplier at one hundred dollars per person per month. Translated into Coasian terms, this means that the cost of production that previously justified a team of twenty, organized within a firm, with managers and meetings and specification documents and all the apparatus of internal coordination, can now be borne by a single individual at a cost that is negligible relative to even a single salary. The entire edifice of organizational coordination that existed to manage the production of that output — the hiring, the onboarding, the performance reviews, the team meetings, the project management tools, the sprint planning sessions — becomes overhead without a corresponding benefit.

This does not mean firms will disappear. Coase himself was careful on this point. The firm does not exist solely because of production costs. It exists because of the full range of transaction costs, and some of those costs — particularly the costs associated with trust, judgment, and the coordination of genuinely complex undertakings — are not reduced by AI in the same way that production costs are.

But the Coasian framework demands that the question be asked freshly. Not "Will firms survive?" — which is a question for pundits — but "Where does the boundary sit now?" Which activities still justify organizational coordination, and which can be performed more efficiently by individuals transacting on the market or producing independently? The answer to this question will determine the shape of the economy for the next generation, and it cannot be derived from first principles. It must be derived from observation of what is actually happening — from studying the real world, as Coase always insisted, rather than the blackboard.

The observation that matters most is the one that Segal returns to throughout The Orange Pill: the collapse of what he calls the imagination-to-artifact ratio. When the cost of converting an idea into a working product approaches zero, the entire structure of economic organization that existed to manage the conversion process is called into question. The project manager, the technical lead, the sprint team, the QA department, the deployment pipeline managed by a dedicated DevOps team — each existed because the conversion was expensive, complex, and required specialized coordination. When the conversion can be accomplished by a single person describing what they want in natural language, the coordination costs that justified each of those roles are no longer incurred.

The firm does not vanish in this scenario. But it shrinks. It retreats to the activities where coordination costs are still high enough to justify organizational structure — activities that, as subsequent chapters will argue, are increasingly concentrated in judgment, direction, and the social functions that markets have never been able to provide.

Coase would not have predicted AI. He was not a futurist. He was an economist who believed that the only way to understand economic institutions was to study them as they actually operated, in specific places, at specific times, solving specific problems. But the framework he built — the simple, elegant, empirically grounded framework that says the boundary of the firm is determined by the comparison of transaction costs — does not need to predict AI. It only needs to observe its effects. And the effects are legible, if one knows where to look: in the engineer who no longer needs a team, in the founder who no longer needs a technical co-founder, in the organization that is discovering, with a mixture of excitement and alarm, that the activities it was built to coordinate can now be performed by its individual members acting alone.

The question Coase asked in 1931 — why do firms exist? — has a new answer, or rather, the same answer applied to new facts. Firms exist because transaction costs make internal coordination cheaper than market transactions. AI has reduced transaction costs in categories that were, for nearly a century, the firm's reason for being. The boundary is moving. The question is where it will settle, and what organizational forms will emerge around it.

That question cannot be answered theoretically. It can only be answered by studying what is actually happening — by visiting the factories, as Coase visited the American factories in 1931, and asking the people inside them why they are organized the way they are. The answers, as Coase discovered ninety-five years ago, will be more interesting than anything the theorists have imagined.

Chapter 2: The Anatomy of a Transaction Cost

Transaction costs, as the concept has developed since Coase's original paper, fall into three canonical categories. Search costs are the costs of finding the right person, resource, or piece of information — the costs of discovering who can do what you need, at what price, with what level of quality. Bargaining costs are the costs of negotiating terms — agreeing on price, schedule, specifications, ownership, and the hundred other details that a contract must settle before work can begin. Enforcement costs are the costs of ensuring compliance — monitoring whether the work meets the agreed standard, resolving disputes when it does not, and pursuing remedies when one party fails to perform.

Carl Dahlman, writing in 1979, consolidated these categories and demonstrated that they could be treated as a unified concept: the resources consumed by the process of exchange itself, as distinct from the resources consumed by the production being exchanged. The distinction matters. When an economist says that markets are efficient, the statement implicitly assumes that exchange is costless — that buyers and sellers find each other instantly, agree on terms without friction, and comply with agreements without monitoring. In the real world, none of this is true. Exchange consumes resources, and those resources are substantial.

How substantial? Estimates vary, but the economic literature has repeatedly found that transaction costs account for a significant share of economic activity in developed economies. John Wallis and Douglass North estimated in 1986 that the transaction sector — the portion of GDP devoted to activities whose primary purpose is facilitating exchange rather than producing goods — accounted for over forty-five percent of U.S. national income. Not production. Transaction. Nearly half the economy dedicated not to making things, but to the process of finding, negotiating, and enforcing agreements about making things.

These estimates are necessarily imprecise. The boundary between transaction and production costs is not always clear, and reasonable economists disagree about where to draw it. But the order of magnitude is not in dispute. A very large share of what knowledge workers do every day — writing specifications, attending meetings, reviewing work, negotiating priorities, resolving disagreements about scope, managing contracts with vendors, conducting performance reviews, onboarding new employees — is transaction activity. It exists not because it produces value directly, but because it facilitates the production of value by others.

AI's effect on these costs is not uniform. It does not reduce all transaction costs equally, and the unevenness of the reduction is what makes the analysis interesting. Some categories collapse almost entirely. Others are reduced but remain significant. Still others are barely touched, or are replaced by new forms of friction that did not exist before.

Search costs, in many domains, approach zero. When an engineer working with an AI coding assistant needs a solution to a technical problem, the search process that previously consumed hours — reading documentation, searching forums, consulting colleagues, evaluating competing approaches — is replaced by a conversation that takes minutes. The AI has, in effect, already conducted the search. It has been trained on the documentation, the forums, the accumulated solutions of millions of practitioners. The cost of discovering the relevant information, which was the first and often the most time-consuming transaction cost in knowledge work, is reduced to the cost of asking a question.

Bargaining costs are reduced through a different mechanism. When the person who conceives a product can also produce it, using AI to handle the implementation, there is no bargain to strike. The specification meeting that previously consumed an afternoon — in which the product manager explained what the product should do, the engineering lead estimated what it would cost, the designer proposed how it should look, and all three negotiated the inevitable compromises between vision and feasibility — is replaced by a single individual describing what they want and receiving it.

This is not a minor efficiency gain. The specification meeting is one of the most expensive recurring transactions in knowledge work, not because of the time it consumes directly, but because of the information lost in translation. The product manager's vision is compressed into a specification document. The specification document is interpreted by an engineer whose understanding of the product manager's intent is necessarily incomplete. The implementation reflects the engineer's interpretation, which diverges from the specification, which itself diverged from the vision. Each translation degrades the signal.

This translation cost, which Coase did not name in his original taxonomy but which The Orange Pill documents extensively, may be the single largest transaction cost in knowledge work. It is the cost of converting intention into a form that someone else can act upon — the cognitive and temporal expense of externalizing a thought precisely enough that a specialist in a different domain can execute it faithfully. Every specification document is a translation. Every requirements meeting is a negotiation conducted in imperfect mutual understanding. Every design review is an attempt to verify that the translation was faithful, and every revision that follows is evidence that it was not.

When Segal describes the natural language interface as "the most time-consuming part of the journey just disappeared," the Coasian translation is direct. The most expensive transaction cost in the production of software — the cost of translating intention from the language of the person who knows what should exist into the language of the person who knows how to build it — has been eliminated. Not reduced. Eliminated. The machine speaks the language of intention. The translation is no longer necessary.

Enforcement costs present a more complex picture. In the traditional firm, enforcement takes the form of quality assurance, code review, testing, and the managerial oversight that ensures the output matches the specification. When an individual produces the output directly with AI assistance, some of these enforcement costs disappear. There is no need to review whether the implementation matches the specification when the person who wrote the specification also produced the implementation. The gap between intent and execution, which quality assurance exists to police, closes to zero.

But new enforcement costs emerge. The AI produces code that works — Segal is clear on this point — but the individual using the AI may not understand the code well enough to verify that it works correctly in all cases, handles edge conditions appropriately, or meets the security and performance standards that production systems require. The enforcement cost shifts from monitoring whether the implementation matches the specification to monitoring whether the AI's output meets standards that the user may not be fully equipped to evaluate.

Dave Friedman, writing in late 2025, argued that the apparent collapse of transaction costs is partly illusory — that costs do not vanish but mutate into new forms. The cost of compute, the cost of accessing capable AI models, the cost of evaluating AI output that is plausible but potentially flawed — these represent what Friedman called "compute friction," a new category of transaction cost that Coase could not have anticipated. The point has merit. The Coasian framework does not assume that transaction costs disappear. It assumes that their magnitude and distribution change, and that organizational forms change in response.

The net effect, however, is clear in direction if not precise in magnitude. The aggregate reduction in transaction costs is enormous. The search costs that consumed hours are reduced to minutes. The bargaining costs that consumed days of meetings and specification-writing are reduced to conversations. The translation costs that degraded signal across every handoff in the production chain are eliminated when the chain is compressed to a single individual. Even accounting for the new costs that emerge — the compute friction, the verification burden, the risk of plausible but incorrect output — the net reduction is large enough to shift the Coasian boundary of the firm significantly inward.

The practical consequence is observable now, in real organizations making real decisions. A technology company that previously employed five frontend engineers, three backend engineers, two designers, a product manager, and a project manager to build a product feature can now accomplish equivalent output with a dramatically smaller team. Not because any of those roles was unnecessary in the old cost structure, but because the transaction costs that justified their existence as separate roles — the handoffs, the specifications, the reviews, the negotiations over scope and priority — have been compressed or eliminated.

The California Management Review published an analysis in April 2025 warning that the uncontrolled adoption of AI within organizations could increase what it called "organizational entropy" — the proliferation of individually optimized but collectively uncoordinated AI-augmented processes. The warning reflects a genuine tension in the Coasian framework: when individuals become capable of independent production, the coordination that previously occurred through organizational hierarchy must be replaced by something, or the organization loses coherence. The firm's response to reduced transaction costs cannot simply be to eliminate coordination. It must be to find new, lower-cost forms of coordination appropriate to the new capability distribution.

This is the institutional design challenge that Coase's framework illuminates with particular clarity. The question is not whether transaction costs have fallen. They have. The question is what organizational architecture matches the new cost structure — what forms of coordination are still necessary, what forms have become unnecessary overhead, and what new forms might emerge to capture the value that the cost reduction makes possible while managing the risks that uncoordinated individual production introduces.

The next several chapters take up this question in detail. But the foundation is the taxonomy presented here: the systematic examination of which transaction costs AI reduces, which it transforms, which it leaves untouched, and which new costs it creates. The analysis proceeds from Coase's insistence that economic reasoning must begin with observable facts, not with theoretical assumptions about how the world ought to work. The facts are visible. The transaction costs are shifting. The boundary of the firm is moving. The only question is where it will settle.

Chapter 3: The Individual as Firm

Alex Finn, a solo builder documented in The Orange Pill, spent the year 2025 constructing revenue-generating software products without writing a line of code by hand. Twenty-six hundred hours of work. Zero days off. No employees, no contractors, no co-founder. One person with an AI assistant and an idea, producing output that, five years earlier, would have required a team of five and a runway of twelve months.

The Coasian question this raises is not whether Finn's achievement is impressive. It is whether Finn, operating alone, constitutes a firm.

In the standard economic taxonomy, Finn is an independent contractor — a market participant who sells output rather than labor, who bears the risk of production rather than receiving a guaranteed wage, and who coordinates his own activities rather than submitting to managerial direction. But the standard taxonomy was built for a world in which the independent contractor's production capacity was limited by the skills of a single individual. A freelance programmer could produce code. A freelance designer could produce designs. Neither could produce a complete product, because a complete product required capabilities that no single individual possessed.

AI changed the production function. The solo builder armed with an AI coding assistant possesses capabilities that previously required a division of labor: architecture, implementation, testing, interface design, deployment. The independent contractor who can do all of these things is not, in any meaningful economic sense, a freelancer. The individual is a production unit whose output matches that of a small firm, achieved without the coordination costs that a small firm incurs.

Coase's framework predicts this outcome with precision. The firm exists because internal coordination is cheaper than market transaction. The individual who needs no coordination — because the individual can perform all necessary functions — represents the limiting case of the Coasian firm: a firm of one, in which coordination costs are zero because there is no one to coordinate with. The cheapest form of coordination is its absence.

But this formulation, while logically correct, misses something important about what is actually happening. The individual is not coordinating with no one. The individual is coordinating with an AI system that functions as something between a tool and a colleague — a system that takes direction, produces output, maintains context across interactions, and responds to feedback with an adaptiveness that earlier tools could not approach. The transaction between the individual and the AI is not a market transaction in the standard sense. There is no negotiation. There is no contract. There is no enforcement problem. The AI does not have interests that conflict with the user's interests. It does not shirk. It does not renegotiate. It does not form a union.

This is a novel economic relationship, and the Coasian framework must stretch to accommodate it. The AI is not an employee — it does not require hiring, training, managing, or compensating in proportion to output. It is not a market counterparty — there is no bargaining, no specification of terms, no risk of non-performance in the contractual sense. It is, in economic terms, a production technology that has absorbed the functions of coordination. The individual directs the AI as a manager directs an employee, but without the costs of management. The individual purchases the AI's services as a buyer purchases from a vendor, but without the costs of contracting.

The result is that the individual-plus-AI unit occupies a position in the Coasian framework that did not previously exist: a production unit with the capabilities of a firm and the coordination costs of a solo actor. The implications of this novel position are significant for the theory of organizational boundaries.

First, it means that the range of activities that can be efficiently performed by individuals on the market, rather than within firms, has expanded enormously. Tasks that previously required organizational coordination — because they required the combined skills of multiple specialists — can now be performed by a single person whose AI assistant provides the missing capabilities. The market absorbs activities that used to belong to the firm.

Second, it means that the type of firm that survives will be different. If individuals can produce, the firm that merely organizes production has lost its economic justification. The surviving firm must provide something that individual production cannot: coordination at a level of complexity that exceeds what a single individual, even an AI-augmented one, can manage; or social functions — trust, mentorship, identity, standards-maintenance — that markets and individuals cannot self-provide.

Third, and most practically, it means that the make-or-buy decision that every organization faces has been restructured from the ground up. This restructuring is the subject of the next chapter, but the foundation is the observation made here: the individual has become a viable alternative to the firm for a class of activities that was, until very recently, firmly within the firm's domain.

Consider the Trivandrum training that Segal describes. Twenty engineers, each discovering that they could perform work across multiple domains. The backend engineer building frontend interfaces. The designer writing complete features. Each individual absorbing transactions that previously occurred between specialists within the firm.

In Coasian terms, each engineer was expanding the boundary of the individual and contracting the boundary of the firm. The transactions that had previously justified organizing backend and frontend into a departmental structure — the handoffs, the specification documents, the design reviews, the sprint planning meetings — were being absorbed into the individual's expanded production function. The firm's internal coordination was being replaced by the individual's AI-augmented self-coordination.

The implications are not symmetric across all levels of capability. The Coasian framework predicts, and the evidence confirms, that the expansion of individual capability is most pronounced for activities that are well-specified and modular — activities where the AI can produce correct output from a clear description. For activities that are ambiguous, context-dependent, or require judgment that draws on tacit knowledge accumulated over years of experience, the individual-plus-AI unit is less capable, and the firm's coordination function retains its value.

Segal makes this observation directly when he notes that the most capable engineers got the most robust output from the AI. The tool amplified existing judgment. It did not replace it. An experienced engineer directing AI-assisted work across multiple domains was producing high-quality output because the engineer's accumulated understanding of systems, quality, and architectural coherence was providing the judgment layer that the AI's production capacity required. A less experienced engineer using the same tool produced more but with less reliability, because the judgment layer was thinner.

This finding is consistent with the Coasian framework in an important way. The firm's coordination function is not merely logistical — it is not merely the scheduling of tasks and the routing of information. At its most valuable, the firm's coordination function is judgmental. The senior engineer who reviews a junior engineer's work, the architect who vetoes a design decision that would create technical debt, the product leader who decides which feature to build next — each is exercising judgment that has been developed through years of organizational experience and that the individual-plus-AI unit cannot yet replicate for itself.

The question, then, is not whether individuals will replace firms. It is at what level of complexity and ambiguity the firm's coordination function remains more efficient than the individual's self-coordination. Below that level, individuals will migrate to the market. Above it, firms will persist, but in a form that reflects their reduced role: smaller, flatter, more focused on the exercise of judgment than on the management of production.

This prediction is already observable. Michael Christen, writing in April 2026, distinguished between two functions that have historically been fused in the managerial role: the routing of information and the exercise of leadership. The first — remembering who is responsible for what, following up on commitments, ensuring that the right information reaches the right person — is precisely the kind of coordination that AI can absorb, and that the individual-plus-AI unit can perform without managerial overhead. The second — setting standards, developing talent, making painful tradeoffs in public view — is the kind of coordination that requires human judgment and human accountability, and that the firm exists to provide.

The emerging organizational forms reflect this unbundling. The "vector pod" described in The Orange Pill — a small group whose function is not to produce but to decide what should be produced — is, in Coasian terms, a firm that has shed its production function and retained only its coordination function. The production is performed by AI-augmented individuals who may or may not be members of the pod. The pod's value lies entirely in its capacity for judgment: its ability to identify which problems are worth solving, to evaluate competing approaches, to maintain quality standards, and to exercise the kind of strategic direction that individual producers, operating alone, may lack the perspective or the incentive to provide.

The individual as firm is real. It is happening now, measurably, in the shift of production from teams to AI-augmented individuals. But the individual as firm is not the end of the story. It is the beginning of a reorganization — a restructuring of the boundary between individual and organization that will produce new forms of coordination to match the new distribution of capabilities.

The Coasian framework does not predict which new forms will emerge. It predicts only that the forms that survive will be those that minimize the sum of transaction and coordination costs under the new conditions. Everything else — the specific structures, the specific distributions of authority, the specific mechanisms of coordination — is an empirical question, to be answered by observing what people actually do, not what theorists think they should do.

The one-person firm is the limiting case. Most economic activity will not migrate there. But the existence of the limiting case changes the calculus for every organizational form between the solo builder and the large corporation, because every firm must now justify its coordination costs against the alternative of individual production. Those that cannot will shrink. Those that can — because they provide coordination, judgment, or social functions that individuals cannot — will persist, but in altered form.

Chapter 4: The Make-or-Buy Decision After Automation

The make-or-buy decision is the operational expression of Coase's theory. Every organization, every day, makes choices about which activities to perform internally and which to purchase on the market. Hire an in-house designer or contract with a design firm. Build a proprietary tool or license an existing one. Train an employee in a new skill or hire someone who already possesses it. Each decision is a comparison of costs: the cost of internal production (including coordination, management, and overhead) against the cost of market procurement (including search, negotiation, quality verification, and the risk of misalignment between what was specified and what is delivered).

For most of the twentieth century, the make-or-buy decision in knowledge work tilted heavily toward make — toward internalizing activities within the firm. The reason was the cost structure of knowledge production. Producing software required programmers who understood the codebase, the architecture, the business logic, and the institutional context. Producing legal analysis required lawyers who understood the client's situation, the regulatory environment, and the precedents that governed the area of practice. Producing marketing required specialists who understood the product, the customer, and the competitive landscape.

In each case, the knowledge required for production was highly specific to the firm. An outside contractor could produce the deliverable, but the contractor lacked the firm-specific context that determined whether the deliverable was actually useful. The cost of transferring that context — of explaining the codebase, the business logic, the client's situation — was a transaction cost that made external procurement expensive relative to internal production, even when the external provider was technically more skilled.

Oliver Williamson, extending Coase's framework in the 1970s and 1980s, formalized this observation as the problem of asset specificity. When the assets required for a transaction — skills, knowledge, relationships, physical equipment — are specific to a particular firm or relationship, market transactions become risky. The party that has invested in firm-specific assets is vulnerable to opportunistic renegotiation by the counterparty. The firm internalizes the transaction to protect against this risk, accepting the costs of internal coordination as the price of avoiding the costs of market opportunism.

Asset specificity explained why firms grew large. The more firm-specific the knowledge required for production, the stronger the case for internal organization. A law firm that understood its client's business was more valuable than any outside counsel, no matter how technically brilliant, because the firm had accumulated the context that made its advice applicable. A software company that had built a proprietary codebase could not easily outsource its maintenance, because the codebase was an asset specific to the company, and any external developer who learned it would acquire bargaining power disproportionate to their market rate.

AI disrupts this logic by reducing the significance of firm-specific knowledge in production. When an AI system can absorb the codebase, the documentation, the architectural patterns, and the business logic of a firm's software in hours rather than months, the firm-specific knowledge that previously justified internal production is no longer exclusively internal. The asset specificity that made outsourcing expensive — the risk that an outside contractor would not understand the firm's particular context — diminishes when the AI can provide that context to anyone who asks.

This is observable now. A contractor equipped with an AI assistant that has been given access to a company's documentation and codebase can produce work that previously required months of onboarding. The knowledge that used to reside in the heads of senior engineers, accumulated over years of working within the firm, is increasingly externalized — captured in documentation, codebases, and data that AI can process and present to anyone. The firm-specific human capital that Williamson identified as the primary driver of vertical integration is being commoditized by the same tool that is reducing production costs.

The practical consequence is a systematic shift in make-or-buy decisions across the economy. Activities that firms internalized because of firm-specific knowledge requirements are migrating to the market — to contractors, freelancers, and AI-augmented individuals who can produce comparable output without the overhead of permanent employment. Activities that firms internalized because of coordination complexity are migrating to smaller, more focused internal teams whose coordination costs are lower because AI has absorbed the mechanical components of coordination.

Consider the sequence described in The Orange Pill when Segal recounts building Napster Station in thirty days. The product required industrial design, optics, audio routing, conversational AI, and software integration. Under the previous cost structure, each of these domains would have required a specialist, and coordinating the specialists would have required project management, regular meetings, specification documents, and the continuous negotiation of tradeoffs between competing design constraints.

The Coasian analysis of this episode is instructive. The thirty-day timeline was possible not because the work was simple, but because the transaction costs between domains had been compressed. The specification-to-implementation transaction that would have consumed weeks of meetings and document review was replaced by conversation with an AI system that could hold the full context of the project. The coordination cost between disciplines — the recurring expense of aligning industrial design decisions with software architecture decisions with audio engineering requirements — was reduced by the ability of a single directing intelligence (Segal's) to maintain coherence across domains with AI assistance.

The episode illustrates a broader pattern: the make-or-buy decision is not simply shifting from make to buy. It is shifting toward a third option that the traditional framework did not contemplate: make yourself, with AI. The architect who previously had to choose between hiring an in-house development team (make) and contracting with a software vendor (buy) can now describe what she wants to an AI system and produce it herself (make yourself). This third option eliminates both the coordination costs of internal production and the transaction costs of external procurement, at the price of accepting whatever limitations the AI introduces in terms of quality, customization, and sophistication.

The limitations are real and should not be minimized. AI-generated output is, as of this writing, excellent for well-specified problems and unreliable for problems that require the kind of tacit, context-dependent judgment that experienced professionals provide. A solo builder can produce a working product, but the product may lack the architectural coherence, the edge-case handling, and the deep domain understanding that a team of experienced specialists would provide. The make-yourself option is most viable for initial versions, for prototypes, for products where speed to market matters more than perfection, and for domains where the AI's training data is extensive and current.

But the boundary of what counts as well-specified is expanding rapidly. Problems that required specialized judgment last year are well-specified problems this year, because the AI's capabilities have improved and because the accumulated experience of millions of users has produced prompting strategies and workflow patterns that extract higher-quality output. The make-yourself option is not static. It is growing, and it is growing into territory that the make-or-buy decision previously allocated firmly to the make-within-a-firm column.

The Software Death Cross documented in The Orange Pill is, in Coasian terms, the market's repricing of firms whose primary value was in the production layer. When code can be produced by individuals at negligible cost, the software company whose value proposition was "we write the code" has lost its transaction-cost justification. The market correctly identifies that the firm's boundary has contracted — that the activities it performs are no longer more efficiently organized internally than on the market — and reprices accordingly.

The firms that retain their value are those whose make-or-buy advantage lies above the production layer: in the data, the integrations, the customer relationships, the regulatory compliance, the institutional trust that cannot be replicated by an individual with a tool, no matter how powerful the tool is. These assets are genuinely firm-specific in Williamson's sense. A customer's twenty years of data in a CRM system is an asset that binds the customer to the firm, not because switching costs are artificially high, but because the data has been accumulated within a specific institutional context and is most valuable when analyzed within that context.

The make-or-buy decision after AI, then, is not a simple story of externalization. It is a restructuring in which the production layer migrates to individuals and markets, the ecosystem layer remains within firms, and the coordination layer — the exercise of judgment about what to produce, for whom, and to what standard — becomes the primary determinant of firm boundaries. Firms that understood their value as residing in the ecosystem and coordination layers will adapt. Firms that understood their value as residing in the production layer will face the arithmetic that every organization now confronts: if an individual can produce what the firm produces, at a fraction of the cost, why does the firm exist?

The answer, for the firms that survive, will be the same answer Coase gave in 1937, applied to different facts: because the costs of coordination, for certain activities, at certain levels of complexity, are still lower within the firm than on the market. The activities that meet this criterion are fewer than they were a year ago. They will be fewer still a year from now. The boundary is moving, and the make-or-buy decision is moving with it, and every organization that fails to reassess its boundary continuously will discover that the market has reassessed it for them — in the form of competitors who produce the same output with fewer people, at lower cost, in less time.

Chapter 5: What Markets Cannot Provide

The previous chapters have traced a trajectory that might appear, at first reading, to point toward the dissolution of the firm. Transaction costs collapse. Individual production capacity expands. The make-or-buy decision shifts toward buy, or toward the novel third option of make-yourself-with-AI. The Coasian boundary contracts inward. The firm shrinks.

If this trajectory were the whole story, the logical endpoint would be an economy of atomized individuals, each a one-person firm, coordinating through markets and platforms without the need for organizational hierarchy. Some commentators have embraced precisely this conclusion. The "Coasean Singularity," as Shahidi, Rusak, Manning, Fradkin, and Horton termed it in their 2025 NBER chapter, describes a condition in which AI agents reduce transaction costs so dramatically that the economic rationale for the firm approaches zero. In such a world, every activity that can be specified can be transacted on the market, and the firm, as a distinct organizational form, becomes unnecessary.

The Coasean Singularity is a useful thought experiment. It is not a description of any world that will actually exist. Coase himself said as much about the frictionless world implied by his own theorem. "I never liked the Coase Theorem," he told Russ Roberts on EconTalk. "It's a proposition about a system in which there were no transaction costs. It's a system which couldn't exist. And therefore it's quite unimaginable." The same skepticism applies with equal force to the Coasean Singularity. Transaction costs will not reach zero. They will fall, dramatically, in the categories documented in Chapter 2 — search, bargaining, translation, enforcement of production quality. But there are functions that markets have never been able to provide efficiently, functions that are not reducible to transaction costs in the standard taxonomy, and these functions will sustain the firm even as its production rationale erodes.

The first such function is trust.

Markets operate on the basis of enforceable contracts. A buyer and a seller agree on terms, and the legal system provides a mechanism for enforcement if either party fails to perform. This mechanism works well for transactions that can be fully specified in advance: a quantity of goods at a stated price, delivered by a stated date, meeting stated quality standards. The enforcement mechanism fails, or becomes prohibitively expensive, when the transaction involves activities that cannot be fully specified.

Most knowledge work cannot be fully specified. A product strategy that will determine the direction of a company for the next three years cannot be reduced to a contract. The judgment calls that a senior engineer makes dozens of times a day — which architectural pattern to use, which edge cases to handle, which technical debt to accept — cannot be enumerated in a specification document. The mentoring that a senior professional provides to a junior colleague, the transmission of tacit knowledge about how the organization works, what matters, what to watch for, cannot be purchased on the market at any price, because the mentor's value lies precisely in the firm-specific understanding that only organizational membership provides.

Trust substitutes for specification. Within a firm, workers trust each other — imperfectly, unevenly, but genuinely — because they share a context, a history, a set of norms, and a mutual investment in the organization's success. This trust enables collaboration on activities that cannot be contracted for, because the parties do not need to specify every detail in advance. They rely on shared understanding, shared commitment, and the expectation that disputes will be resolved through negotiation rather than litigation.

Markets cannot produce this kind of trust. Markets produce the trust of reputation — the expectation that a counterparty will perform because failing to perform will damage their reputation and reduce their future business. Reputational trust is valuable and real, but it is transactional. It is grounded in the calculation of future benefits, not in shared commitment to a common enterprise. It does not extend to the kind of unspecifiable collaboration that the most important organizational work requires.

Segal's account of building Napster Station illustrates the point. The thirty-day sprint was possible not only because AI reduced production costs, but because the team trusted each other — trusted Segal's direction, trusted each other's judgment, trusted that the shared commitment to an audacious goal would be honored even when the work was exhausting and the timeline was impossible. "Human fast trust is not a shortcut," Segal writes. "It is the hardest thing to build and the most valuable thing to have, and it cannot be manufactured or mandated or optimized."

An economy of atomized individuals, each a one-person firm, would lack this trust infrastructure. It would be capable of producing any component that can be specified, but incapable of sustaining the collaborative judgment that complex, ambiguous, high-stakes undertakings require. The firm persists, in part, because it provides a trust environment that the market cannot replicate.

The second function that markets cannot provide efficiently is the transmission of tacit knowledge.

Michael Polanyi, the physical chemist turned philosopher, drew a distinction in 1966 between explicit knowledge — knowledge that can be articulated, documented, and transmitted through language — and tacit knowledge — knowledge that is embedded in practice, acquired through experience, and resistant to articulation. "We can know more than we can tell," Polanyi wrote, and the observation has implications for organizational economics that the Coasian tradition has not fully absorbed.

Tacit knowledge is the knowledge that a senior engineer possesses when she looks at a codebase and feels that something is wrong before she can articulate what. It is the knowledge that a product manager possesses when he evaluates a feature proposal and knows, from years of watching users interact with similar features, that it will not work, even though the data does not clearly support the judgment. It is the knowledge that a master craftsman possesses about the properties of materials — knowledge that was built, as The Orange Pill describes, through thousands of hours of friction, of trial and error, of depositing thin layers of understanding that accumulate into something solid enough to stand on.

Tacit knowledge cannot be purchased on the market because it cannot be specified. A firm cannot write a contract for "the intuition that comes from fifteen years of building production systems." It can hire a person who possesses that intuition, embed that person in an organizational context that allows the intuition to be exercised and transmitted, and create the conditions — mentoring relationships, code reviews, architectural discussions, the daily proximity of junior and senior practitioners — that enable the tacit knowledge to flow from one person to another.

AI can absorb and transmit explicit knowledge with extraordinary efficiency. It can read every document, every codebase, every piece of documentation the organization has ever produced, and make that knowledge available to anyone who asks. But tacit knowledge, by definition, is not in the documents. It is in the heads and hands of the practitioners, and it flows through channels — apprenticeship, collaboration, the shared experience of solving hard problems together — that AI cannot yet replicate.

The firm persists, in part, as a mechanism for the transmission of tacit knowledge. The mentoring relationship, the code review, the architectural discussion — these are not merely social niceties that organizations provide for the comfort of their employees. They are knowledge-transmission mechanisms that the market cannot provide and that AI cannot replace, because the knowledge being transmitted has never been and perhaps cannot be made explicit.

The third function is the maintenance of professional standards and norms.

Markets optimize for price. In a perfectly competitive market, the lowest-cost producer wins. This is efficient when quality is observable — when the buyer can evaluate the product before purchasing and reject substandard work. But quality in knowledge work is often unobservable to the buyer, or observable only long after the purchase. The client who hires a lawyer does not know whether the legal strategy is sound until the case is resolved. The company that deploys AI-generated code does not know whether it handles edge cases correctly until the edge cases occur in production. The student who uses AI to produce an essay does not know whether the essay reflects genuine understanding until the examination reveals that it does not.

In the absence of observable quality, markets tend toward what George Akerlof described as adverse selection — the lemons problem. When buyers cannot distinguish high quality from low quality, they are unwilling to pay the premium that high quality requires, and high-quality producers exit the market or reduce their quality to match the price. The result is a market dominated by low-quality output, not because low quality is preferred, but because the information asymmetry makes high quality unsustainable.

Professional norms — the standards maintained by firms, professional associations, guilds, and communities of practice — are the primary mechanism for preventing this degradation. A law firm maintains standards of practice because the firm's reputation depends on it. An engineering team maintains coding standards because the senior engineers enforce them. The firm provides the institutional context in which norms are established, monitored, and enforced through mechanisms that are less formal and more effective than market contracts: peer pressure, professional pride, the desire to maintain standing within a community of practitioners.

When production migrates to AI-augmented individuals operating on the market, the question of who maintains standards becomes urgent. An individual producing software alone, using AI, has no peer review. No senior engineer looks over the work and says, "This will create problems in six months." No architectural review catches the decision that is locally optimal but globally destructive. The individual may produce output that works today and fails tomorrow, and the market may not detect the failure until the cost has been incurred.

The California Management Review's April 2025 analysis raised precisely this concern. When employees create and deploy their own specialized AI agents, atomization and quality fragmentation follow. Each individual optimizes for their own output, and the collective coherence that organizational standards provided degrades. The firm, in this analysis, is not merely an efficiency mechanism. It is a quality-assurance mechanism, and its dissolution carries risks that the transaction-cost framework, narrowly applied, does not capture.

The fourth function, and perhaps the most resistant to economic formalization, is the provision of professional identity and belonging.

Human beings are not purely economic actors. They are social creatures who derive meaning, purpose, and identity from their membership in communities. For a very large number of people, the firm is the primary community of professional life. It is where colleagues become friends, where mentors shape careers, where shared accomplishments create the bonds that sustain people through difficult periods. The elimination of the firm's production function does not eliminate the human need for professional belonging. It simply leaves that need unmet.

The economic consequences of unmet belonging are real, even if they are difficult to quantify. Workers who lack professional community are less likely to invest in skill development, less likely to maintain professional standards, less likely to engage in the kind of voluntary knowledge-sharing that drives innovation. The social capital that firms generate — the trust, the shared knowledge, the professional norms, the sense of common purpose — is a genuine asset, and its depreciation has genuine costs.

Mark Granovetter, in his influential 1985 paper "Economic Action and Social Structure," argued that economic activity is embedded in social relationships, and that economic theory's tendency to treat actors as isolated individuals pursuing self-interest is a fundamental distortion. The Coasian framework, applied without attention to embeddedness, risks precisely this distortion. It correctly identifies the transaction-cost logic that is shifting the boundary of the firm, but it risks overlooking the social infrastructure that the firm provides and that the market cannot replicate.

The conclusion is not that firms will persist unchanged. The production rationale for the firm is eroding rapidly, and the organizational forms that emerge in response will be different from what exists today. But the conclusion is that firms will persist — transformed, reduced in scope, reorganized around functions that have nothing to do with production efficiency and everything to do with the social conditions that productive work requires.

Trust. Tacit knowledge. Professional standards. Belonging. These are the irreducible social functions of the firm, the functions that remain when AI has absorbed everything else. The firm of the future is, in significant measure, a social institution — a community organized around shared values, shared knowledge, and shared purpose, whose production function has been largely externalized to AI-augmented individuals but whose coordination, quality-assurance, and community functions remain essential.

The Coasian framework, applied honestly, leads to this conclusion: the firm exists because markets are costly, and some of the things that are costly to transact on the market are not goods or services but relationships. The firm is, among other things, a relationship technology. And relationship technologies are not, as yet, subject to the same automation that has transformed production technologies.

Chapter 6: The Coasian Boundary in Motion

The boundary of the firm is not a line drawn once and left in place. It is a frontier, advancing and retreating as the underlying cost structure shifts. Coase understood this. "A firm will tend to expand until the costs of organising an extra transaction within the firm become equal to the costs of carrying out the same transaction by means of an exchange on the open market," he wrote in 1937. The boundary sits at an equilibrium that is continuously recalculated as conditions change.

What has not been sufficiently appreciated, either in the original Coasian literature or in its contemporary extensions, is the speed at which the boundary can move. In previous technological transitions, the shift was gradual enough that organizations could adapt incrementally. The telephone reduced the cost of coordinating across distance, and firms gradually expanded their geographic scope over decades. Email reduced the cost of asynchronous communication, and firms gradually adjusted their meeting structures, their reporting hierarchies, their expectations about response times. Each shift in the boundary was measurable in years, sometimes in decades, allowing institutional structures time to reorganize.

The AI transition is moving the boundary faster than institutions can adapt. The gap between the speed of capability change and the speed of institutional response — what Sebastian Galiani, extending Coasian reasoning to AI governance, called the institutional lag — is the defining structural problem of the current moment. The technology is moving at the pace of software deployment. The institutions are moving at the pace of organizational culture, legal reform, and human psychology. The result is a period of disequilibrium in which the Coasian boundary has already moved, but the organizations on either side of it have not yet reorganized to reflect its new position.

This disequilibrium is visible in the Software Death Cross documented in The Orange Pill. A trillion dollars of market value vanished from software companies in the first eight weeks of 2026. The market was not responding to a decline in the quality of those companies' products. It was responding to a shift in the Coasian boundary — a recognition that the production activities those companies performed could now be performed more cheaply outside the firm, by AI-augmented individuals or by competing firms with dramatically lower cost structures.

The repricing was crude, as market repricing always is. It did not distinguish between firms whose value resided in the production layer — firms that were, in essence, code-writing organizations whose primary asset was the labor of their programmers — and firms whose value resided in the ecosystem layer — the data, the integrations, the customer relationships, the institutional trust that twenty years of enterprise deployment had built. Both categories lost value, because the market had identified the direction of the boundary shift but had not yet calibrated its magnitude for individual firms.

The Coasian analysis of the Death Cross is more precise than the market's reaction. The firms that should lose value are those whose boundary-justification was production: firms that existed because writing software was expensive, and whose competitive advantage was their ability to write it more efficiently than alternatives. These firms are on the wrong side of the boundary. The production activity that justified their organizational structure can now be performed by individuals at negligible cost, and the coordination overhead that the firm incurs — the salaries, the offices, the management layers, the institutional apparatus — is no longer offset by a production advantage.

The firms that should retain value are those whose boundary-justification was never primarily production. Enterprise platforms whose moat is not the code but the data layer that twenty years of customer usage have accumulated. Infrastructure providers whose value is in the reliability, security, and compliance guarantees that individual producers cannot match. Professional services firms whose competitive advantage is the institutional trust — the brand, the track record, the willingness to accept liability — that clients require for high-stakes decisions.

For these firms, the Coasian boundary has not moved beneath them. It has moved around them, contracting the production layer while leaving the ecosystem and trust layers intact. Their organizational form will change — fewer production employees, more coordination and judgment roles, a flatter structure that reflects the reduced need for managerial oversight of production activities — but their economic rationale survives.

The boundary is also moving in directions that create new organizational opportunities. When transaction costs fall, activities that were previously too expensive to coordinate across organizational boundaries become viable as market transactions. A small firm that could not previously afford to hire a full-time data scientist can now access data-science capabilities through an AI-augmented freelancer at a fraction of the cost. A startup that could not previously compete with established players because it lacked the engineering team to build a comparable product can now produce a comparable product with a fraction of the headcount.

The competitive landscape becomes more fluid. The barriers to entry that transaction costs created — the cost of assembling a team, building institutional knowledge, establishing the internal coordination mechanisms that production required — are lower. New entrants can challenge incumbents not by matching their organizational scale but by matching their output with a fraction of the resources. The Coasian boundary has moved inward for the incumbents and outward for the challengers, creating a compression of competitive advantage that the market is still learning to price.

But the fluidity cuts in both directions. Dave Friedman's observation that transaction costs do not vanish but mutate into new forms is empirically grounded and theoretically important. The costs of accessing capable AI models, the costs of evaluating and verifying AI output, the costs of maintaining the data infrastructure that AI requires — these are real costs, and they may favor large organizations that can amortize them across a broader base of activity. The individual who can produce code at negligible marginal cost still faces the fixed costs of the AI subscription, the compute resources for training and fine-tuning, the data pipeline that feeds the AI system, and the expertise to evaluate whether the output is reliable.

Samuel Hammond, in his analysis of AI and power concentration, warned that the Coasian logic could operate in favor of larger, not smaller, organizations. "AI is already driving massive reductions in the cost of monitoring, enforcement, bargaining and so forth," Hammond observed. "But with true AGI, transaction and principal-agent costs have the potential to be driven to near zero; what some economists recently dubbed the Coasean Singularity." The concern is that near-zero monitoring costs favor whoever does the monitoring — and in practice, that tends to be the entity with the most data, the most compute, and the most comprehensive surveillance infrastructure. The Coasian boundary, in this scenario, does not contract the firm. It expands it, because the firm that can monitor and coordinate cheaply can absorb activities that smaller organizations and individuals perform less efficiently.

The historical evidence on this question is ambiguous. Previous transaction-cost reductions have sometimes favored decentralization — the personal computer enabled small businesses to compete with large ones in data processing, for instance — and sometimes favored centralization — the internet enabled platform companies to coordinate millions of transactions at near-zero marginal cost, producing the largest firms in economic history. The direction of the boundary shift depends on which transaction costs fall fastest and which organizational forms are best positioned to exploit the reduction.

The honest Coasian assessment is that both forces are operating simultaneously. The boundary is moving inward for production activities, enabling individuals and small firms to produce what large firms previously monopolized. The boundary is moving outward for coordination and infrastructure activities, enabling large platforms to capture the value of the transactions that individuals and small firms generate. The net effect is not a uniform contraction or expansion of the firm but a restructuring: a hollowing out of the middle, in which mid-sized firms that were too large to be nimble and too small to be platforms find themselves squeezed from both directions.

Galiani's institutional analysis adds a necessary dimension. The boundary of the firm is not determined solely by transaction costs in private markets. It is also shaped by the institutional environment — the property rights, the regulatory framework, the legal infrastructure that determines who bears the costs and captures the benefits of economic activity. "Ronald Coase taught us that markets don't operate in a vacuum — they depend on rules, on clearly defined and enforceable rights," Galiani wrote. When the institutional framework fails to keep pace with the boundary shift, the result is not a smooth reorganization but a period of costly disorder in which property rights are unclear, liabilities are unassigned, and the organizations on both sides of the moving boundary operate in a state of institutional uncertainty.

The AI transition is currently in this state of disorder. Who owns the output of an AI system trained on copyrighted material? Who bears liability when AI-generated code causes a system failure? Who is responsible for verifying the quality of AI-assisted work — the individual who directed the AI, the company that deployed the AI, or the AI provider whose model produced the output? These questions have not been resolved, and the Coasian boundary cannot settle at a stable position until they are.

The boundary is in motion. The direction is legible: inward for production, stable or outward for coordination, trust, and ecosystem functions. The speed is faster than institutions can currently accommodate. The endpoint is not a singularity — not a world without firms — but a restructured landscape in which the firm's production rationale has been largely absorbed by AI-augmented individuals and the firm's remaining rationale is coordination, judgment, and the social functions that the next chapter will examine in detail.

The boundary will continue to move. The organizations that thrive will be those that track its movement continuously, shedding functions that have migrated to the market and concentrating on functions that remain — for now — more efficiently organized within the firm. The qualification is deliberate. What remains within the firm today may migrate tomorrow, and the Coasian discipline requires that the question be asked freshly at every cycle: Does this activity still justify organizational coordination? Or has the boundary moved past it?

Chapter 7: New Organizational Forms for New Capabilities

If the hierarchical firm was the organizational solution to high transaction costs, the question that follows the reduction of those costs is not whether the firm disappears — Chapter 5 established that it does not — but what forms of organization emerge to match the new distribution of capabilities between individuals and institutions.

The question is not theoretical. New organizational forms are already appearing. They are appearing in the way that organizational forms always appear — not through deliberate design by planners but through the trial-and-error experimentation of practitioners who face practical problems and improvise solutions. The Coasian framework does not predict which specific forms will succeed. It predicts the criteria by which they will be evaluated: the forms that survive will be those that minimize the sum of transaction costs and coordination costs under the new conditions.

Four organizational forms are currently visible. Each represents a different balance between individual AI-augmented production and collective coordination. Each has specific strengths and specific failure modes. None is likely to be the final answer. All are experiments, and the experimental period will last years.

The first form is the guild. A guild, in the medieval sense, was a community of practitioners organized around a shared craft. It provided training through apprenticeship, maintained quality standards through peer review, offered social belonging through membership, and regulated entry to the profession through credentialing. The guild did not employ its members. Members were independent producers who operated on the market, selling their services to customers. The guild's function was not production but the maintenance of the conditions — skills, standards, trust, identity — under which production could occur at a reliable level of quality.

The relevance of the guild model to the AI economy is direct. When individual producers can create at the level that previously required firms, the question of quality assurance becomes urgent. Who reviews the AI-augmented individual's output? Who maintains the standards that prevent the market from degrading into Akerlof's lemons problem? Who transmits the tacit knowledge that senior practitioners possess and junior practitioners need?

A guild of AI-augmented developers, organized as a community of practice rather than a firm, could provide these functions without the overhead of hierarchical management. Members would be independent producers, taking clients and building products on their own, but embedded in a community that provided code review, architectural guidance, quality certification, and the mentoring relationships through which tacit knowledge flows. The guild would not direct its members' work. It would certify the conditions under which the work was done, providing clients with a signal of quality that the individual producer, operating alone, could not credibly offer.

This form has historical precedent and contemporary expression. Open-source communities already function as partial guilds, maintaining code standards through peer review and transmitting knowledge through collaborative development. Professional associations in law, medicine, and accounting perform guild functions of credentialing and standard-maintenance. The AI economy may produce a proliferation of such structures, adapted to the specific needs of AI-augmented knowledge work: communities that certify not just the practitioner's skill but the practitioner's judgment in directing AI tools, the ability to evaluate AI output critically, and the discipline to maintain quality standards when the tool makes quantity so easy to produce.

The failure mode of the guild is insularity. Medieval guilds became protectionist, restricting entry to protect incumbents, resisting innovation that threatened established practices, and accumulating regulatory power that served the guild's interests at the expense of the public. Any contemporary guild structure must guard against this tendency, which is to say it must build in mechanisms for admitting new members, adopting new practices, and subjecting its own standards to external scrutiny.

The second form is the studio. A studio is a small group — typically fewer than ten people — organized around a shared creative or productive vision. Studios have been the dominant organizational form in architecture, film, design, and game development for decades. The studio model assumes that the best work is produced by small, tightly coordinated teams whose members know each other well enough to collaborate without the overhead of formal specification and review.

AI amplifies the studio model by expanding the productive capacity of each member. A studio of five AI-augmented individuals can produce output comparable to a conventional team of fifty, while maintaining the tight coordination and shared understanding that are the studio's competitive advantage. The coordination costs are low because the team is small. The production capacity is high because each member is AI-augmented. The result is an organizational form that combines the firm's coordination advantages with the individual's production efficiency.

The "vector pod" described in The Orange Pill is a variant of the studio model. A small group whose function is to decide what should be built, with AI handling the production. The pod exercises collective judgment — evaluating problems, identifying solutions, making tradeoffs — while the production is distributed to individual members or to AI systems operating under the pod's direction. The pod's value lies entirely in its capacity for judgment, not in its capacity for production.

The failure mode of the studio is scalability. Studios work well for projects that a small team can direct. They work poorly for undertakings that require coordination at a scale that exceeds the cognitive capacity of a small group. A studio can build a product. It cannot build an operating system, a global logistics network, or a regulatory compliance infrastructure. For these undertakings, a larger organizational form is needed — one that provides coordination mechanisms that scale beyond the personal relationships that the studio depends on.

The third form is the platform. A platform is an infrastructure that enables independent producers to coordinate at scale without organizational hierarchy. Amazon Marketplace, Uber, Airbnb, GitHub — each is a platform that connects individual producers with customers, provides the transaction infrastructure that enables exchange, and maintains quality through rating systems, review mechanisms, and algorithmic matching.

The platform model is the most natural extension of the Coasian logic in the AI economy. If transaction costs are what justify the firm, and AI reduces transaction costs, then the organizational form that emerges should be one that facilitates market transactions rather than one that replaces them with internal coordination. The platform does exactly this. It reduces the remaining transaction costs — search, matching, payment, reputation — to a level that makes market transactions viable for activities that previously required organizational coordination.

AI-augmented platforms are already emerging. Marketplaces where AI-augmented freelancers offer services that previously required firms — product development, legal analysis, financial modeling, design — with the platform providing quality assurance through rating systems, credential verification, and AI-assisted evaluation of output quality. The platform's competitive advantage is not production but facilitation: reducing the costs that individual producers and their clients would otherwise incur in finding each other, agreeing on terms, and verifying quality.

The failure mode of the platform is extraction. Platform operators occupy a position of structural power — they control the infrastructure through which transactions occur — and the economic incentive to exploit that position is strong. Friedman's observation that transaction costs mutate into "access rents" is particularly applicable to the platform model. When the platform becomes the only viable channel for connecting producers with customers, the platform can extract rents that reproduce, in a different form, the transaction costs that the platform was supposed to eliminate. The Coasian efficiency gain is captured by the platform operator rather than distributed to the participants.

The fourth form is the network — a loosely structured association of independent producers who coordinate through informal relationships, shared norms, and ad hoc arrangements rather than through hierarchical authority or platform infrastructure. Networks have always existed in economic life. The clusters of small firms in Italian industrial districts, the collaborative relationships among Silicon Valley startups, the informal knowledge-sharing among professionals at conferences and in online communities — each is a network that provides coordination benefits without the costs of formal organization.

AI may strengthen the network form by reducing the coordination costs that previously limited its effectiveness. When AI can handle the mechanical components of coordination — scheduling, information routing, progress tracking, document management — the human components of network coordination — trust, shared understanding, mutual commitment — become more salient and more valuable. The network's competitive advantage over the firm is its flexibility; over the platform, its human depth. Members coordinate not because an algorithm matches them but because they know and trust each other, and the coordination is shaped by shared professional judgment rather than by platform incentives.

The failure mode of the network is fragility. Networks depend on voluntary participation and informal norms, and they dissolve when participants defect — when the incentive to free-ride on the network's knowledge-sharing exceeds the incentive to contribute, or when the informal norms that maintain quality are not enforced because there is no mechanism of enforcement. Networks that persist tend to develop more formal structures over time — credentialing, governance, explicit rules — which brings them closer to the guild or the firm.

No single form will dominate. The Coasian framework predicts pluralism: different organizational forms will be efficient for different activities, at different scales, in different institutional contexts. The guild will be most effective where quality assurance and tacit knowledge transmission are paramount. The studio will be most effective where tight coordination and shared creative vision matter most. The platform will be most effective where scale and matching efficiency are the primary requirements. The network will be most effective where flexibility and trust among known partners are the critical resources.

The large hierarchical firm will persist as well, for activities where the coordination requirements exceed what guilds, studios, platforms, and networks can provide — activities like maintaining global supply chains, operating regulated infrastructure, and managing the institutional relationships that governments and large customers require. But the hierarchical firm will be a smaller share of the organizational landscape, and the hierarchical firm that does persist will be organized differently: flatter, leaner, more focused on coordination and judgment, with its production function largely externalized to AI-augmented individuals operating in guilds, studios, platforms, and networks.

The transition between organizational forms is not costless, and the Coasian framework insists that these transition costs be accounted for. Workers who have built careers within hierarchical firms must find their place in the new organizational landscape. Skills that were valuable within the firm — navigating bureaucracy, managing up, playing organizational politics — may be less valuable in the guild or the studio. Skills that were undervalued within the firm — independent judgment, self-direction, the ability to maintain quality without external oversight — may become essential. The transition will produce winners and losers, and the institutional challenge is to build structures that cushion the transition for the losers while enabling the winners to realize the gains.

The organizational experiments are underway. The verdict will be empirical, as Coase would have demanded. Not which form theorists prefer, but which forms practitioners adopt, sustain, and find workable in the actual conditions of the AI economy.

Chapter 8: The Problem of Social Cost in the AI Economy

In 1960, twenty-three years after asking why firms exist, Coase asked a second question that proved equally consequential: What should be done when the activities of one party impose costs on others?

The question concerned what economists call externalities — costs or benefits that fall on parties who did not choose to incur them. A factory that pollutes a river imposes costs on the downstream fishermen. A beekeeper whose bees pollinate a neighbor's orchard provides benefits the neighbor did not pay for. The standard economic response, developed by Arthur Pigou and dominant for decades before Coase's paper, was that the government should intervene: tax the polluter, subsidize the beekeeper, correct the market failure through regulatory action.

Coase's insight was that the problem was not market failure but institutional design. If property rights were clearly defined and transaction costs were low, the parties could negotiate a solution that was efficient regardless of how the rights were initially assigned. The downstream fishermen could pay the factory to reduce pollution, or the factory could pay the fishermen for the right to pollute, and either arrangement would reach the same efficient outcome. The government intervention that Pigou recommended was unnecessary — provided that transaction costs were low enough to permit negotiation.

The qualification was crucial, and Coase insisted on it more than his interpreters typically acknowledged. In the real world, transaction costs are not low. Negotiation is expensive. Information is asymmetric. Property rights are ambiguous. And when transaction costs are high, the initial assignment of rights matters enormously, because the parties cannot negotiate their way to an efficient outcome. The practical implication of the Coase theorem is not that government intervention is unnecessary. It is that institutional design — the assignment of property rights, the structure of liability rules, the mechanisms for dispute resolution — determines outcomes in ways that economic theory had not adequately appreciated.

The AI economy generates externalities at a scale and speed that existing institutional frameworks are not equipped to handle. The costs are real, they fall on identifiable parties, and the property rights governing them are ambiguous or undefined. A Coasian analysis of these externalities does not begin by asking whether government should intervene. It begins by asking how property rights are currently assigned, who bears the costs under the current assignment, and whether a different assignment would produce better outcomes at lower transaction costs.

The most prominent externality is the training-data problem. Large language models are trained on vast corpora of text, code, images, and other creative works, much of it produced by individuals and organizations who were not compensated for the use and did not consent to it. The AI companies benefit from the training. The consumers benefit from the capabilities the training produces. The original creators bear a cost: their work has been used to build a competing capability that may reduce the market value of their future output.

Under the current assignment of property rights, this cost falls entirely on the creators. The legal framework governing the use of publicly available data for AI training is unsettled, and the practical reality is that creators have no effective mechanism for preventing the use of their work or for extracting compensation after the fact. The transaction costs of enforcing copyright against AI training are prohibitive: the scale of the training data is measured in billions of documents, the causal connection between any individual document and the model's output is diffuse and difficult to establish, and the legal frameworks that might govern the relationship were designed for a world in which copying meant reproduction, not statistical learning.

The Coasian question is not whether creators deserve compensation — that is a normative question that economics cannot answer — but whether a different assignment of property rights would produce a more efficient outcome. If creators held a clearly defined right to compensation for AI training use, the AI companies would face a cost they currently externalize. This cost would be reflected in the price of AI services, which would be higher. The higher price would reduce demand, which would reduce the scale of training, which would reduce the capability of the models. The question is whether the sum of these effects — higher prices, reduced capability, compensated creators — produces a better or worse outcome than the current arrangement — lower prices, higher capability, uncompensated creators.

The Coasian framework does not provide a clean answer. It provides a method: compare the costs of different institutional arrangements, including the transaction costs of implementing and enforcing each arrangement, and choose the one that produces the greatest net benefit. The comparison is empirical, not ideological, and it requires detailed knowledge of the actual costs involved — knowledge that, in the current period of institutional uncertainty, nobody fully possesses.

A second externality concerns the labor-market effects of AI adoption. When AI enables one person to produce the output of a team, the remaining team members bear a cost that they did not choose to incur. The cost is not imposed by the individual who adopted the tool. It is imposed by the technological capability itself, and it falls on workers whose skills have been devalued not through any failure of their own but through a shift in the production technology that their skills were calibrated to serve.

The historical pattern documented in The Orange Pill — the Luddites, the framework knitters, the displacement that accompanies every major technological transition — is, in Coasian terms, a pattern of externalities. The gains of the transition accrue to the adopters and the consumers. The costs accrue to the displaced workers. The gains are concentrated and immediate. The costs are dispersed and extended over time — over careers, over communities, over generations.

The Luddite episode is instructive because it illustrates what happens when the institutional framework fails to assign the costs of transition in a way that is both efficient and equitable. The Luddites bore the full cost of the transition to mechanized production. The factory owners captured the full benefit. The institutional framework — the legal system, the political system, the social safety net — did not provide mechanisms for redistributing the costs, and the result was decades of immiseration that was inefficient as well as unjust, because the human capital of the displaced workers was wasted rather than redirected to productive use.

Coase would not have endorsed the Luddites' response — the destruction of machines is the elimination of a productive asset, which is inefficient by any standard. But Coase would have recognized that the problem was institutional, not technological. The machines were not the cause of the suffering. The absence of institutional mechanisms for managing the transition was the cause. Property rights, liability rules, social insurance, retraining infrastructure — these are the instruments of institutional design that determine whether a technological transition produces broad benefit or concentrated harm.

Acemoglu and Johnson, in Power and Progress, documented this pattern across a thousand years of technological transitions. The technology does not determine the distributional outcome. The institutions do. Transitions that were accompanied by institutional innovation — labor laws, educational systems, social insurance, redistributive mechanisms — produced broadly shared gains. Transitions that were not accompanied by institutional innovation produced concentrated gains and dispersed losses, often for decades.

The AI transition is currently proceeding without adequate institutional innovation. The technology is advancing at the speed of software deployment. The institutional framework is adapting at the speed of legislative process, regulatory procedure, and judicial interpretation. The gap between these speeds is the practical expression of Galiani's institutional lag, and it is the gap in which the externalities of the AI transition accumulate.

A third externality, less discussed but potentially more consequential, concerns the effect of AI on the institutional infrastructure of knowledge production. When AI can produce competent analysis in any domain, the market for human analysis in that domain contracts. The contraction reduces the economic incentive for individuals to invest in developing deep expertise. The reduction in investment reduces the supply of deep expertise. The reduction in supply reduces the quality of the training data available for the next generation of AI models, because the models are trained on human-produced work, and the quality of the work available for training depends on the depth of the expertise that produced it.

This is a negative feedback loop — an externality that operates over time rather than across parties. The current generation of AI models was trained on work produced by experts who developed their expertise through years of investment that the pre-AI economy rewarded. If the AI economy ceases to reward that investment, the next generation of experts will be smaller and less skilled, and the training data available for future AI models will be correspondingly thinner.

The Coasian analysis of this problem focuses on the assignment of rights and the design of incentives. The current institutional framework does not assign value to the positive externality that deep human expertise provides as training data for AI systems. A creator who produces high-quality work that is used to train an AI model receives no compensation for the training value and faces reduced market compensation for the competitive output the model produces. The incentive to produce high-quality work is doubly undermined: the reward is reduced and the cost is borne without compensation.

A different institutional design might assign a property right to the training value of creative and expert work, creating a mechanism for compensation that would maintain the incentive to invest in deep expertise. The transaction costs of implementing such a mechanism are substantial — identifying which works contributed to which model outputs, quantifying the contribution, administering the compensation — but the costs of not implementing it may be larger, measured in the long-term degradation of the human expertise that the AI economy depends upon.

The Coasian conclusion is institutional, not technological. The externalities of the AI economy are not inherent in the technology. They are consequences of the current assignment of property rights and the current design of institutional mechanisms for managing the costs of technological transition. A different assignment, a different design, would produce different outcomes. The question is not whether to intervene but how — which rights to assign, which mechanisms to build, which institutional innovations to pursue — taking into account the transaction costs of each alternative and the practical constraints of implementation in a period of rapid technological change.

Coase, in his 1992 Nobel lecture, observed that economists had spent too much time studying the allocation of resources in abstract markets and too little time studying the institutional structure within which allocation actually occurs. "It is the institutional structure of production that I am concerned with in this lecture," he said. The observation applies with particular force to the AI economy, where the most consequential decisions are not about the allocation of resources but about the design of the institutions — the property rights, the liability rules, the regulatory frameworks, the social safety nets — within which the allocation occurs.

These decisions are being made now. They are being made in legislatures and courtrooms and regulatory agencies, but also in corporate boardrooms and open-source communities and the daily practices of millions of individuals who are adopting AI tools without a clear institutional framework to govern their use. The decisions will determine whether the AI transition follows the pattern of transitions that produced broadly shared gains or the pattern of transitions that produced concentrated gains and dispersed losses. The technology does not determine the outcome. The institutions do. And the institutions, at this moment, are not keeping pace.

Chapter 9: The Future of Coordination

The argument of the preceding eight chapters can be compressed into a single proposition: AI has not eliminated the need for coordination. It has eliminated the need for a particular kind of coordination — the coordination of production — while intensifying the need for a different kind — the coordination of direction.

The distinction is not semantic. It corresponds to two genuinely different activities that have been bundled together within the firm for so long that most practitioners cannot tell them apart. Production coordination is the management of who does what, when, and how: scheduling tasks, assigning roles, routing information between specialists, monitoring progress, resolving bottlenecks. Direction coordination is the exercise of judgment about what should be done at all: identifying which problems are worth solving, evaluating which approaches are worth pursuing, deciding when to persist and when to abandon, maintaining coherence between the short-term actions of individual contributors and the long-term trajectory of the enterprise.

In the traditional firm, both types of coordination were performed by the same organizational apparatus — the management hierarchy. The project manager who scheduled sprints also made judgment calls about scope. The engineering director who assigned tasks also decided which technical direction to pursue. The CEO who organized the company also determined its strategy. The bundling of production coordination and direction coordination was efficient when both were expensive, because the same organizational structure could economize on both sets of costs simultaneously.

AI unbundles them. Production coordination — the scheduling, the routing, the monitoring, the mechanical management of who is working on what — is precisely the kind of repetitive, information-intensive, rule-governed activity that AI performs well. Jacques Bughin, analyzing AI's effect on organizational structure, identified this unbundling with precision: "Authority can be embedded in software. Memory can be persistent. Follow-up can be automatic. Coordination no longer vanishes when a conversation ends." The production-coordination function of the manager is absorbable by AI systems that track progress, route information, maintain context, and flag deviations from plan — all without the cognitive limitations, the political biases, and the variable attention that human managers bring to these tasks.

Direction coordination — the exercise of judgment about what to build, for whom, and why — is not absorbable. Not because AI cannot process the relevant information. It can process more of it, faster, than any human. But because direction requires something that information processing does not provide: a commitment to an outcome. A willingness to be wrong. An accountability that is personal rather than computational.

When a product leader decides to pursue a particular market, the decision is not a calculation. It is a bet, informed by data but determined by judgment — by the leader's understanding of the customer, the competitive landscape, the team's capabilities, and the dozens of intangible factors that data captures imperfectly and that the leader synthesizes through experience, intuition, and the willingness to accept responsibility for being wrong. The decision can be aided by AI, which can provide analysis, surface patterns, and test assumptions. But the decision itself — the commitment to one path rather than another, with all its consequences — is an act of human judgment that cannot be delegated to a system that bears no consequences for error.

Michael Christen drew this distinction sharply: "The point is not to replace managers with machines, but to unbundle two jobs that have been fused. One is the routing of information — a task ripe for automation. The other is leadership: setting standards, developing talent and owning painful trade-offs in public. The first can be coded. The second cannot." The Coasian translation is direct. The firm of the future exists not to coordinate production — that function has been absorbed by AI — but to coordinate direction: to house the judgment function, to create the conditions under which judgment can be exercised well, and to provide the accountability structures that make judgment consequential.

The practical implications are already visible. The vector pod described in The Orange Pill is a direction-coordination unit. It does not produce. It decides. Its members spend their time talking to users, analyzing markets, debating strategy, and producing specifications that AI-augmented individuals execute. The pod's value is measured not by its output — it produces no code, no designs, no deliverables in the traditional sense — but by the quality of its decisions: whether the things it chose to build were the right things, whether the specifications it produced were coherent, whether the direction it set was followed through to outcomes that served users well.

This organizational structure inverts the traditional ratio of production to direction. In a conventional technology company, the majority of employees are production workers — engineers, designers, testers — and a small minority are direction workers — product managers, executives, strategists. In the AI-augmented organization, the ratio reverses. A small number of AI-augmented individuals, or AI systems operating with minimal human oversight, handle production. The majority of the organizational effort is devoted to direction: deciding what to build, evaluating whether it was built well, maintaining the standards that distinguish a good product from a functional one, and sustaining the relationships — with customers, with partners, with the broader community — that production alone cannot maintain.

This inversion has consequences for hiring, for compensation, for career development, and for organizational culture. The skills that the direction-focused firm values are not the skills that the production-focused firm valued. Technical proficiency — the ability to write code, to design interfaces, to build systems — was the primary hiring criterion in the production-focused firm. In the direction-focused firm, technical proficiency is a necessary but insufficient condition. The primary hiring criteria are judgment, taste, the ability to synthesize information across domains, the capacity to make decisions under uncertainty, and the social skills — communication, persuasion, empathy, the ability to understand what users need rather than what they say they want — that direction requires.

The compensation structure shifts accordingly. In the production-focused firm, the highest-paid individual contributors were the most technically skilled — the senior engineer, the principal architect, the staff designer. In the direction-focused firm, the highest-paid contributors are those with the strongest judgment — the product leaders who consistently choose the right problems, the strategists who anticipate market shifts, the creative directors whose taste determines whether the product delights or merely functions.

Career development shifts as well. The traditional career ladder in technology ran from junior engineer to senior engineer to staff engineer to principal engineer, with each rung representing deeper technical mastery. The direction-focused firm requires a different ladder: from individual contributor (AI-augmented production) to judgment contributor (evaluation and quality assurance) to direction contributor (deciding what to build) to coordination contributor (maintaining the organizational and social infrastructure that makes good direction possible). The transition between rungs requires not deeper technical skill but broader perspective, stronger judgment, and greater facility with the social and institutional dimensions of productive work.

The cultural shift may be the most significant and the most difficult. Production-focused organizations developed cultures organized around output: lines of code written, features shipped, deadlines met. These metrics were imperfect but measurable, and the measurability provided a shared framework for evaluating contribution. Direction-focused organizations must develop cultures organized around judgment quality: decisions made, outcomes achieved, problems identified before they became crises, opportunities seized that others missed. These metrics are harder to measure, slower to materialize, and more subjective in their evaluation.

The difficulty of measuring judgment quality creates a genuine organizational risk. When output is the metric, contribution is visible. When judgment is the metric, contribution is ambiguous. The product leader whose decision to pursue a particular market results in success eighteen months later may be rewarded — or the organization may have forgotten who made the decision, or may attribute the success to the execution team rather than the direction team. The incentive structures that sustained the production-focused firm do not transfer cleanly to the direction-focused firm, and the organizational design challenge of creating incentive structures appropriate to judgment work is substantial and unsolved.

The Coasian framework clarifies what is at stake. The firm exists because coordination is costly, and the firm persists because the coordination it provides is more efficient than the alternatives. If the direction-focused firm cannot develop metrics, incentive structures, and cultural norms that effectively reward good judgment, it will fail — not because the direction function is unnecessary but because the organizational form that houses it is inefficient. Direction will migrate to other forms: to studios, to guilds, to the informal networks of trusted advisors that founders have always relied upon when formal organizations failed them.

The future of coordination, then, is not guaranteed to be organizational in the traditional sense. It may be that the direction function is best housed in small, informal, high-trust groups rather than in formal firms. It may be that the guild model — communities of practice that provide standards, mentorship, and peer evaluation without hierarchical management — proves more efficient for direction coordination than the corporate model. It may be that the platform model evolves to provide direction services: marketplaces where organizations can access strategic judgment the way they currently access freelance production.

The Coasian prediction is that the forms that survive will be those that minimize the total cost of direction coordination, including the costs of finding good judgment (search), agreeing on terms (bargaining), and ensuring that the judgment is actually followed through (enforcement). These costs are not trivial. Good judgment is rare, difficult to identify in advance, and expensive to verify after the fact. The organizational forms that solve these problems most efficiently will define the economic landscape of the coming decades.

What can be said with confidence is that the irreducible core of the firm, stripped of its production function, is the coordination of direction — the institutional capacity to decide what should exist in the world and to hold people accountable for the quality of those decisions. This capacity requires trust, shared context, institutional memory, and the social infrastructure that Chapters 5 and 7 described. It does not require the hierarchical management of production. It does not require the departmental silos that the production-focused firm created to manage the division of labor. It does not require the project management apparatus that existed to coordinate handoffs between specialists.

It requires people who can think clearly about what matters, who can evaluate options with rigor and imagination, who can make decisions under uncertainty and accept responsibility for the outcomes, and who can sustain the relationships — with colleagues, with customers, with the broader community — that give direction its grounding in reality rather than in abstraction.

The firm of the future is a judgment institution. Its product is decisions. Its competitive advantage is the quality of those decisions. And the Coasian boundary that determines its scope is the point at which the cost of making decisions within the institution exceeds the cost of making them on the market — through freelance advisors, through consulting engagements, through the informal networks that have always supplemented and sometimes replaced the formal organization.

That boundary will be determined, as Coase always insisted, not by theory but by observation — by studying what people actually do, how they actually organize, and which arrangements actually produce the outcomes that the participants are seeking. The organizational experiments are underway. The results will take years to materialize. The Coasian framework does not predict the winner. It only insists that the winner will be the form that coordinates direction most efficiently, taking into account all the costs — economic, social, institutional — that coordination entails.

Epilogue

The invoice was four hundred and twelve dollars.

That was what Anthropic charged for the Claude usage that produced the core architecture of a product feature at Napster — a feature whose development, under the old cost structure, I would have estimated at somewhere between forty and sixty thousand dollars when you accounted for engineering time, project management overhead, specification-writing, review cycles, and the inevitable revisions that follow every translation from what you meant to what someone else understood you to mean.

Four hundred and twelve dollars versus sixty thousand. I stared at the number and understood, with the clarity that comes from seeing your own assumptions exposed as arithmetic, that I was looking at a transaction cost collapsing in real time.

That is what Coase's framework gave me. Not a theory. A lens. A way to see the structure beneath the surface of what was happening in every room I walked into, every team meeting I attended, every quarterly planning session where the old assumptions about headcount and timeline and budget sat on the table like artifacts from a civilization that had already fallen.

The firm exists because transactions are costly. That sentence, written by a twenty-seven-year-old in 1937, explained my entire career. It explained why I had spent decades assembling teams — not because teams were inherently better at producing things, but because the cost of coordinating production across independent actors was higher than the cost of coordinating it within an organization. It explained the project managers, the sprint planners, the specification documents, the design reviews. Each was a mechanism for managing a transaction cost. Each was overhead that the firm accepted because the alternative — market transactions between independent specialists — was more expensive still.

And it explained why the ground was moving. When I watched my engineers in Trivandrum reach across domain boundaries in a week, when I watched a backend developer build user interfaces she had never attempted because the translation cost between her expertise and the frontend domain had dropped to the cost of a conversation — I was watching the Coasian boundary move. Not metaphorically. Structurally. The transactions that had justified organizing those specialists into departments, that had justified the meetings and the handoffs and the management layers, were being absorbed into individuals whose expanded production capacity made the organizational coordination unnecessary.

What stayed with me longest from this intellectual journey was not the prediction that firms will shrink, though they will. It was the identification of what remains after the shrinking. Trust. Tacit knowledge. Professional standards. The capacity for judgment about what deserves to exist. These are the functions that markets cannot provide and that AI cannot perform, and they are, it turns out, the functions that matter most.

I think about the vector pods we have been building at Napster — small groups whose purpose is not to produce but to decide. Their value lies entirely in the quality of their judgment: whether the things they choose to build are the right things, whether the direction they set serves the people we are trying to serve. That judgment cannot be automated. It can be informed by AI, augmented by AI, tested against AI-generated alternatives. But the commitment to a direction — the willingness to say "this is what we should build, and I accept responsibility for the consequences" — is a human act. It requires stakes. It requires caring about the outcome in a way that a system optimizing a loss function does not and cannot.

The Coasian framework is cold. It does not care about human flourishing. It cares about costs. But the coldness is clarifying. It strips away the sentimentality that clouds organizational thinking and forces you to ask the hard question: Does this structure earn its existence? Does the coordination it provides justify the costs it incurs? Or has the boundary moved past it, leaving it stranded on the wrong side of a calculus that has already been recalculated by every competitor, every customer, every individual who has discovered that they can do alone what your organization required twenty people to do together?

The answer, for many organizational structures that exist today, is that the boundary has already moved past them. They persist through institutional inertia, through contracts that have not yet expired, through the human reluctance to acknowledge that the ground has shifted. But the arithmetic is patient. It will catch up.

For the structures that earn their existence — the ones that provide trust, judgment, mentorship, the social infrastructure that productive work requires — the Coasian message is different. These structures are more valuable now, not less, precisely because everything around them has changed. When production is cheap, direction is precious. When anyone can build, the question of what to build becomes the only question that matters. And the organizational forms that house that question well — that create the conditions for good judgment, that transmit the tacit knowledge that judgment requires, that maintain the standards that separate good work from merely functional work — these forms are the irreducible core of economic organization.

A twenty-year-old crossing the Atlantic in 1931 to ask a question his professors could not answer. The question turned out to be the most important question in organizational economics, and its answer — transaction costs — turned out to be the key to understanding not just why firms exist but what they become when the costs that created them dissolve.

The costs are dissolving. The firms are transforming. The question Coase asked is alive again, and it will not be answered by theory. It will be answered by people building things — building organizations, building products, building the institutional structures that determine whether this transition produces broad benefit or concentrated harm.

The young economist who visited the factories would have insisted on nothing less.

-- Edo Segal