AI fits into this: computer intelligence calculate, human entrepreneurs judge. Machines operate on information that can be represented formally; entrepreneurs rely on experience, intuition, and tacit knowledge. Algorithms may improve prediction and therefore push back the boundary of uncertainty, but genuine uncertainty remains the province of human judgment. This is an attractive division of labor, and variants of it can be found in contemporary discussions of AI and entrepreneurship. But, perhaps the premise is wrong and there is a sense in which machine intelligence can possess tacit knowledge.

A fascinating recent paper by the philosopher Céline Budding, published in Philosophy of Science, explains why we should take that possibility seriously. Budding does not claim that LLMs literally think like human beings, possess consciousness, or hold beliefs in anything like the ordinary sense. Her argument is narrower and, for precisely that reason, more interesting. Drawing on Martin Davies’s philosophical account of tacit knowledge, she argues that at least some large language models may contain internal causal structures that warrant attributing to them a form of tacit knowledge.

The argument matters for entrepreneurship because tacitness is often highlighted as a key reason why judgment cannot be reduced to an algorithm (a position I have not defended myself, however; more below …). If tacit knowledge is no longer exclusively human, tacitness by itself cannot establish the distinctiveness of entrepreneurial judgment. The arrival of generative AI therefore gives entrepreneurship scholars an opportunity to ask a question that is more fundamental than whether ChatGPT can write a business plan. What exactly is judgment, and what distinguishes it from other forms of intelligent behavior?

Between rules and behavior

Budding begins with a familiar problem. What could it mean to say that a language model ”knows” that Paris is the capital of France? One possibility is explicit representation. Traditional symbolic AI systems can contain propositions or rules that are stored in something resembling a database and retrieved when needed. In such a system, knowledge is explicit in a fairly strong sense. A rule exists somewhere in the system and plays an identifiable causal role in producing the answer.

Neural networks are not like this. Their representations are distributed across large numbers of parameters. We cannot normally inspect a particular location in the model and find the proposition ”Paris is the capital of France” sitting there waiting to be retrieved. This fact has encouraged a skeptical interpretation according to which an LLM does not really know anything. It merely behaves as if it knows. If prompted appropriately, it produces ”Paris,” but successful input-output behavior does not establish that the system possesses knowledge.

Following Davies Budding rejects the forced choice between these two alternatives. There is an intermediate possibility. A system may lack explicit representations of rules while nevertheless containing internal structures that systematically and causally generate its behavior. Davies calls this tacit knowledge. It is more than conformity to a rule because the attribution is grounded in the internal causal organization of the system. But, it is less than explicit knowledge because the relevant rule cannot simply be read off from a symbolic database or program.

This is where things become difficult for the too easy human-machine distinction. Tacit knowledge has long been associated with precisely those things humans can do without being able to state how they do them. Michael Polanyi’s famous dictum that ”we can know more than we can tell” captures the basic intuition. An experienced entrepreneur may recognize that a prospective partner cannot be trusted, that a technology is finally commercially viable, or that a market is about to change without being able to specify an algorithm that produces the judgment. Tacitness seems to place such knowledge beyond the reach of machines because computers supposedly require explicit instructions. Budding’s argument suggests that this inference is too quick.

Machines that know more than they can tell?

Davies originally doubted that connectionist neural networks could satisfy his conditions for tacit knowledge. Budding argues that the architecture of modern transformer models changes the picture. In particular, embeddings allow semantically related inputs to be represented close to one another in a high-dimensional space. The representations need not be identical. It may be enough that the model recognizes semantic similarity in a way that channels related inputs through similar causal processes.

The empirical evidence Budding considers comes largely from work by Meng et al. on locating and editing factual associations in language models Their ROME method identifies model components associated with particular factual relationships and intervenes on them. A model that associates the Eiffel Tower with Paris can, in the stylized example Budding uses, be edited so that it associates the Eiffel Tower with Rome. The interesting result is not simply that the model subsequently produces ”Rome.” The intervention can generalize across different formulations of the same underlying relationship.

This generalization is important. If the system had merely memorized every sentence separately, changing one input-output association should not systematically change semantically related ones. But, Meng et al. find substantial generalization alongside relatively high specificity: editing one factual association affects paraphrases and related prompts while largely leaving unrelated associations intact. Budding interprets this as preliminary evidence for what Davies calls a ”causal common factor”: some internal structure appears to participate systematically in producing a family of semantically related outputs.

The evidence is far from decisive. Budding discusses problems of polysemanticity, distributed representation, imperfect specificity, and recent challenges to causal-tracing methods. She is careful to describe the conclusion as preliminary. But, the conceptual point survives even if particular mechanistic interpretations of today’s LLMs turn out to be wrong. There is no obvious reason to define tacit knowledge in a way that makes biological implementation a necessary condition. If tacit knowledge consists in causally effective internal structures that systematically guide behavior without taking the form of explicitly represented rules, artificial systems may qualify.

Tacit knowledge isn’t judgment

In the judgment-based approach to entrepreneurship, judgment refers to decision-making under uncertainty; see my earlier post. Its intellectual ancestry lies in Knight’s distinction between risk and uncertainty. Under risk, the possible outcomes and their probabilities are sufficiently well specified to permit calculation. Under Knightian uncertainty, they are not (or, alternatively, they are only well-specified to the entrepreneur). The entrepreneur must act despite the absence of an objectively correct procedure for choosing among alternatives.

Judgment therefore cannot simply be equated with possessing information. Nor can it be reduced to sophisticated prediction. Entrepreneurs form beliefs about an uncertain future and commit resources on the basis of those beliefs. Different entrepreneurs can confront the same observable circumstances and exercise different judgments because the future value of heterogeneous assets is not given independently of entrepreneurial conjecture. Markets then provide an ex post test of decisions that could not be settled ex ante by calculation.

This is where Budding helps clarify the argument. An LLM could possess tacit knowledge in her sense without possessing judgment in the Knightian sense. Knowing tacitly that the Eiffel Tower is in Paris, even if that knowledge is genuinely represented by an internal causal structure, says nothing about whether the system can decide whether to acquire a factory, launch an unfamiliar product, reorganize a company, or enter a market whose future structure is unknown. The distinction is not between explicit and tacit cognition. It is between knowledge and judgment.

That matters because entrepreneurship scholars sometimes slide between these concepts. Hayek’s dispersed knowledge problem, Polanyi’s tacit knowledge, and Knight’s uncertainty are related, but they are certainly not interchangeable. The entrepreneur mayindeed rely on tacit knowledge when exercising judgment, but the need for judgment does not arise merely because some relevant knowledge is tacit. It arises because the economic problem is open-ended and its resolution requires committing resources before the relevant future is known. In other words, while machines may increasingly cross the tacitness boundary (supposing Budding is right), that doesn’t mean they have crossed the judgment boundary.

What, then, is left for the entrepreneur?

This reframing also suggests that the interesting research program is not to identify some cognitive ability that machines can never possess. I am not sure such arguments have had a good historical track record. The safer and theoretically more productive question is to identify the properties of entrepreneurial judgment without building human exclusivity into the definition.

One candidate is conjecture (based on some abductive operation). Entrepreneurial action does not merely predict which known state will occur. Entrepreneurs imagine possible uses of resources, possible products, possible organizational forms, and possible future states that are not simply presented as alternatives in a predefined choice set. In the language of entrepreneurship theory, opportunities need not exist as objective objects waiting to be detected. They may emerge through entrepreneurial imagination, experimentation, and action.

A second candidate is ownership and resource commitment. Judgment as Peter Klein and I have discussed it is closely connected to ownership because markets for judgment are incomplete. If I could simply sell you my judgment that a new technology will succeed, without either of us committing resources, the economic problem would remain unresolved. Exercising entrepreneurial judgment means arranging and committing heterogeneous assets under uncertainty. The entrepreneur is not merely producing an answer to the question ”What will happen?” Rather, she is deciding what resources should be assembled and controlled given a conjecture about what might happen.

This makes the comparison with LLMs sharper. Budding’s model has internal causal structures that may warrant calling some of its representations knowledge. Still, the model she analyzes is still being prompted, evaluated, and intervened upon by external agents. Its factual representations can affect its outputs, but the paper provides no basis for saying that the model originates entrepreneurial projects, acquires property rights, decides which resources to expose to uncertainty, or bears the economic consequences of those commitments. Budding herself deliberately brackets the stronger question of whether tacit knowledge warrants attributing beliefs, concepts, or other richer cognitive states to the model. This doesn’t settle whether future AI systems could exercise judgment; rather, it tells us why Budding’s argument alone does not establish that they can.

AI as a test for entrepreneurship theory

There is a broader lesson here. AI is interesting for entrepreneurship research because it provides a kind of stress test for our theoretical concepts.

Many associate explicit knowledge with machines and tacit knowledge with humans. Deep learning makes that dichotomy increasingly difficult to maintain. If Budding is right, artificial systems can acquire causally effective knowledge that was never explicitly programmed into them and that cannot simply be read off from their architecture. The traditional argument that entrepreneurial judgment must remain human because it relies on tacit knowledge therefore becomes much weaker.

Still, the judgment-based approach (or other entrepreneurship theory) does not ultimately need the claim that machines cannot possess tacit knowledge. Its deeper claim is that economic action takes place in a world in which the future cannot be reduced to a set of known states with known probabilities, and in which somebody must nevertheless decide how heterogeneous resources should be used. Prediction can improve dramatically without eliminating that problem. Tacit artificial knowledge can become extraordinarily sophisticated without making the future determinate.

The usual postmortem calls this a branding mistake. In a new paper forthcoming in Strategic Management Review, Eric Mota and I describe the problem as something more specific and more general: a sequencing failure. The visual identity choice was made before the choice it was supposed to express. No amount of design talent fixes that, because the later choice had nothing to reinforce.

Subscribe now

The problem with “alignment”

Strategy research has explored this issue for fifty years under various labels: strategic fit, alignment, consistency, congruence, integration, coherence. These constructs share a common feature that limits their usefulness: they describe a state that organizations exhibit or don’t. We can recognize coherence when we see it, usually in a firm that is already performing well. It’s less clear how managers can know in advance if a given set of strategic choices work well together.

Moreover, most of these constructs are non-hierarchical. They tell you that choices shouldn’t contradict each other, which is surely true, but not very helpful. Managers need to know which choice comes first, which choice should take precedence when two of them conflict, or what to do when a currently attractive option might counteract something the firm committed to a few years ago.

We did an interpretive synthesis across strategic management, marketing, product development, and resource allocation—literatures that have been working on pieces of this problem in parallel without much contact—and asked a narrower question: which choices actually establish coherence, and in what order?

Seven choices, one anchor

Our answer is a set of seven elements: positioning, customer identity, corporate visual identity, stakeholder messaging, salient features, feature exclusion, and strategic expenditure. These elements operate on different levels. Positioning is the anchor—it defines the territory the firm intends to occupy and, crucially, what it will not do. Customer identity follows, because identity claims have to be credible given the firm’s position (e.g., you can’t characterize customers into connoisseurs while positioning as a mass-market, accessible provider). Communication follows both, because a firm can’t consistently express what it hasn’t defined. Product development follows communication, because features and exclusions supply the evidence for the claim. Resource allocation comes last and enforces the whole package.

The last one deserves emphasis, and connects the argument to my previous work on capital and judgment. Bower showed decades ago that funding patterns reveal a firm’s real strategy regardless of what the strategy document says. The reason, in the language Nicolai Foss and I have used, is that strategy isn’t “real” until someone with authority over assets makes an irreversible commitment of those assets to a particular combination. Ghemawat’s emphasis on strategic commitment and Lachmann’s point about capital complementarity are getting at the same point: heterogeneous assets have value only in specific combinations, and choosing a combination forecloses others. Southwest’s single aircraft type isn’t a cost decision that happens to fit the strategy, it is the strategy.

Reinforcement and its discontents

Coherence, on our account, isn’t the absence of contradiction, but the presence of recursive reinforcement—later choices actively strengthening the anchor rather than simply tolerating it. We identify four pathways for coherence: proof (features demonstrate the claim), amplification (repetition raises salience), constraint escalation (commitments raise the cost of reversal), and credibility locking (accumulated choices make opportunistic repositioning implausible). Nike’s athlete endorsements, product-technology spending, and messaging all circle back to make the same “winner’s attitude” position more credible (and, importantly, harder to imitate—an isolating mechanism, in the language of Lippman and Rumelt).

A point we don’t address in the paper, but future research may consider, is whether recursive reinforcement can create core rigidities that make it difficult to innovate and adapt. Actions that lock in a valuable position can make it harder to see when that position has stopped being valuable (Christensen’s “resource allocation trap”). We say in the paper that discipline is not rigidity, but it may be difficult to tell these apart ex ante. (Put differently, the ability to tell them apart can be understood as a type of superior entrepreneurial judgment.)

What our framework doesn’t claim

A couple of caveats to our argument. First, we’re trying to explain coherence (how a firm’s internal choices reinforce each other), not fit (whether the resulting strategy matches its environment). A beautifully coherent buggy-whip manufacturer is still a buggy-whip manufacturer. These are different problems which the literature sometimes confuses. Second, our framework is conceptual and we don’t suggest specific techniques for identifying and measuring the seven attributes (e.g., how to code separately for positioning-identity fit, positioning-feature fit, and positioning-resource fit and then asking whether particular profiles predict performance). Post-merger integration may offer opportunities for natural experiments in which positioning gets reset and we can see how long it takes for recursive reinforcement to reappear.

In short, firms don’t usually fail because one decision was wrong, but because a bunch of individually defensible decisions were made in parallel, by different people, under different pressures, before anyone settled what the firm was committed to being. The fix is not coordination per se (though that is clearly important!), but being willing to make the anchoring commitment first, and making other choices to match. We hope our framework helps advance thinking about this kind of coordination.

Pedro Bento’s 2014 American Economic Journal: Macroeconomics paper, “Competition as a Discovery Procedure: Schumpeter Meets Hayek in a Model of Innovation,” approaches this old problem from a angle that is novel from the point of view of standard economics (it may be more familiar to those who know the work of Nelson and Winter … ). Bento argues that innovation is not simply an investment whose payoff is uncertain, but a search process in which the quality of the technological opportunities themselves is unknown before firms experiment with them. Once that distinction is introduced, the number of independent innovators becomes economically significant in its own right. Competition matters not merely because it changes each firm’s incentive to innovate, but because it changes how many technological experiments the economy conducts.

That sounds like a modest modification of the standard innovation problem. It is not. It introduces into Schumpeterian growth theory a Hayekian function for competition and, in the process, separates three relationships that are too easily conflated: competition and innovation at the firm level, competition and technological progress at the industry level, and the effect of competition policy on long-run growth. Bento’s most interesting result is that these relationships need not even have the same sign.

The usual Schumpeterian question

The modern literature on competition and innovation has inherited a recognizably Schumpeterian problem. Innovation requires costly effort today in return for an uncertain stream of rents tomorrow. Market structure matters because it changes both the rewards from successful innovation and the consequences of failing to innovate. Competition can therefore sharpen incentives, but it can also erode the rents that make innovative investment worthwhile.

Aghion, Bloom, Blundell, Griffith, and Howitt’s (2005) paper in the QJE showed how these forces can generate an inverted-U relationship between product-market competition and innovation. At low levels of competition, an increase in competition can induce firms to innovate in order to escape their rivals. As competition becomes sufficiently intense, however, the incremental rents from innovation shrink and innovative effort eventually falls. The attraction of the argument is precisely that it escapes the sterile choice between “competition is good for innovation” and “market power is good for innovation.” Both can be true over different parts of the relevant range.

Bento retains this incentive problem but adds another margin. In the usual formulation, competition affects how intensively firms pursue innovation; in Bento’s formulation, competition also affects the number of independent technological draws available to the economy. That matters because innovations differ in their realized productivity, and firms do not know ex ante which experiment will generate the superior outcome. The market is therefore not merely eliciting effort toward a known technological objective; it is sampling from a set of uncertain technological possibilities. Enter Friedrich von Hayek.

Intensive and extensive margins of experimentation

Economists often invoke Hayek’s description of competition as a “discovery procedure,” but usually (in my experience) rhetorically rather than substantively, i.e., as a general, loose claim that markets are dynamic, entrepreneurs are clever, or decentralized economies somehow generate information. Bento adds real content to the rhetorics.

Suppose innovative outcomes are uncertain and heterogeneous. A firm undertakes research, but neither it nor anyone else knows beforehand exactly how productive the resulting innovation will be. If several firms undertake independent innovative activities, society obtains several draws from the relevant distribution of technological possibilities. The technology that ultimately determines productivity is selected from this collection of realized innovations. Increasing the number of firms therefore changes the distribution of the best available technology even if the innovative effort of each individual firm does not increase.

The distinction is between the intensive and extensive margins of experimentation. Schumpeterian reasoning directs attention primarily toward the intensive margin: how much innovative effort does each firm undertake given the rents available to a successful innovator? The Hayekian mechanism adds an extensive margin: how many independent experiments are undertaken, and hence how many opportunities does the economy have to discover an unusually productive innovation?

This is not merely “more firms means more innovation.” Rather, the point is that, under technological uncertainty, the social value of an additional experiment depends partly on the order statistic it generates. Society cares about the best technology discovered, not simply the average quality of all research projects undertaken. Increasing the number of independent draws can therefore improve expected technological performance even while the amount of innovative activity undertaken by the representative firm declines. That is the key idea in the paper.

The average innovator and the technological frontier

Consider what happens as competition increases. Greater competitive pressure changes the rents associated with innovation and therefore changes the innovative effort chosen by each firm. Industry productivity depends on more than the innovative effort of the representative firm; it also depends on the number of firms generating independent innovations and on the quality of the best innovation among them. More competition can mean a larger pool of realized technologies from which the frontier technology emerges.

Suppose an industry moves from a relatively concentrated structure to one containing substantially more independent innovators. Competitive pressure may reduce the resources devoted to innovation by each firm. The expected innovative output of the representative firm may fall. But, with sufficiently many additional experiments, the expected quality of the best realized innovation can rise. Industry productivity growth may therefore increase precisely over a range in which firm-level innovative effort is declining.

This allows Bento to neatly reconcile seemingly conflicting empirical possibilities. Competition may display an inverted-U relationship with firm-level innovation while exhibiting a positive relationship with industry-level productivity growth. There is no paradox once one recognizes that firm-level innovative effort and the economy’s rate of technological discovery are different objects.

This matters empirically. R&D expenditure per firm, patenting per firm, or similar measures of innovative intensity do not necessarily provide sufficient statistics for the technological consequences of market structure. A concentrated industry may contain firms undertaking large research programs, while a less concentrated industry may contain a wider set of independent technological experiments. If realized innovation quality is heterogeneous, the latter structure may generate faster movement of the technological frontier even when its representative firm appears less innovative.

Competition as an epistemic institution

As already suggested, a key aspect of Bento’s paper is that competition does more than alter incentives or resource allocation. It also performs an epistemic function: it helps determine what the economy comes to know about its technological possibilities. If nobody knows which technology will prove best, allowing several independent experiments can reveal something that no participant knew beforehand. The relevant knowledge is generated through the competitive process rather than merely transmitted through it.

This is close to Hayek’s conception of competition as a discovery procedure. Hayek’s point was not simply that markets aggregate information already possessed by dispersed individuals (the focus of his 1945 paper). Competition also produces knowledge through action: prices are tried, products introduced, technologies tested, and conjectures about consumer demand or productive possibilities exposed to market selection. (the focus of his 1946 paper and o/c his 1968 paper ). Some economically relevant knowledge becomes available only because competing actors attempt different things.

Bento captures a tractable version of this idea within his model. Firms undertake uncertain technological experiments, and competition affects how many independent attempts occur. The best realized technology can then determine industry productivity. Competition matters not merely because it changes the innovative effort of existing firms, but because it changes the experimental process through which superior technologies are discovered.

There is, however, a limit to how Hayekian this mechanism can be. Bento represents discovery as stochastic technological search: firms draw innovations whose quality is uncertain beforehand. Hayek’s stronger argument concerns circumstances in which economic actors may not even know the relevant possibilities in advance. Discovery then involves more than learning which outcome is realized from a given distribution; it involves forming and testing conjectures about opportunities whose value, and sometimes whose very existence, cannot yet be specified.

This doesn’t weaken Bento’s result. The simplification is what makes the discovery mechanism analytically tractable. But, it raises a further question that the model itself need not answer: where do genuinely different experiments come from? One way of approaching that question is through the judgment-based account of entrepreneurship, which shifts attention from uncertain technological draws to the heterogeneous judgments that lead economic actors to commit resources to different visions of the future.

From technological draws to entrepreneurial judgment

The JBA offers something that is largely missing from Bento’s otherwise elegant Hayekian mechanism: an account of where heterogeneous experiments come from. In Bento’s model, competing firms generate independent technological draws which suffices to establish the formal result. If innovation quality is uncertain, increasing the number of independent experiments can increase the expected quality of the best realized technology even when innovative effort per firm declines. Competition therefore has a discovery value independent of its effect on the incentives of the representative firm.

However, from a judgment-based perspective, entrepreneurs do not simply draw technologies from an urn. They form judgments about which technologies to pursue, which consumer wants will become important, which resources are undervalued in their present uses, and which combinations of heterogeneous assets might generate future value. They then acquire or control resources and expose those judgments to the market process.

This distinction matters because the judgment-based approach, following Knight, reserves a central role for uncertainty that cannot simply be represented as risk with known probabilities. If the entrepreneur knew the complete set of technological alternatives and the probability distribution governing their returns, the decision problem would largely be one of optimization. Judgment becomes important when neither probabilities nor, in the stronger case, the relevant future possibilities themselves can be specified adequately in advance.

Seen this way, Bento’s competing innovators can be interpreted as entrepreneurs exercising rival judgments about the technological future. Their investments are not merely multiple attempts at solving the same known problem. They are conjectures about what the problem is, which resources will become valuable, and which combinations of those resources will produce value in circumstances that do not yet exist.

This gives competition an even stronger discovery function than Bento’s formal model requires. Competition does not merely increase the number of draws from a distribution of technological outcomes. It permits different theories of the future to be enacted simultaneously.

That is precisely where ownership becomes important in the JBA. Judgment is economically consequential because entrepreneurs can acquire, combine, and deploy heterogeneous assets according to their own beliefs. Ownership provides the entrepreneur with the authority to rearrange resources without having to establish beforehand that the proposed use is objectively superior. Indeed, if superiority could already be demonstrated, much of the entrepreneurial problem would have disappeared.

A decentralized ownership structure therefore permits disagreement to become productive experimentation. Entrepreneurs who hold different judgments can organize resources differently and allow subsequent market outcomes to discriminate among their conjectures. Profit and loss do not eliminate uncertainty ex ante; they provide an ex post selection mechanism among resource combinations generated under uncertainty.

This puts additional flesh on Hayek’s idea of competition as a discovery procedure. What competition discovers is not simply which of several known firms has lower costs. It discovers which entrepreneurial judgments about unknown resource uses survive contact with consumers, technologies, rivals, and complementary investments.

Bento and the JBA therefore illuminate different parts of the same process. Bento provides the formal logic for why a plurality of independent experiments can improve technological outcomes. The JBA explains the microfoundations of that plurality: heterogeneous capital, heterogeneous beliefs, entrepreneurial judgment, ownership, and the authority to commit resources under uncertainty.

The synthesis also sharpens the policy problem. Concentration may matter not simply because it changes conventional measures of market power or R&D incentives, but because it can alter the number of genuinely independent judgments being exercised over productive resources. Ten nominally separate research projects inside one corporation are not necessarily equivalent to ten entrepreneurs risking resources on competing visions of the future. Internal projects ultimately confront a common hierarchy, a common capital-allocation process, and potentially a common strategic conception of what the organization should become.

Conversely, merely increasing the number of firms does not guarantee meaningful experimental diversity. If firms imitate one another, rely on the same information, respond to identical subsidies, or pursue the same technological trajectory, an industry may contain many legal entities but relatively few independent entrepreneurial conjectures. The economically relevant variable is therefore not firm count as such, but something closer to the number and heterogeneity of judgments that can actually command resources.

This suggests a more explicitly entrepreneurial interpretation of Bento’s discovery mechanism. The social value of competition lies partly in preserving the possibility that somebody is wrong in a different direction. Under genuine uncertainty, we do not know which direction will eventually prove wrong, and we therefore cannot rationally select the winning experiment in advance. What decentralized entrepreneurship provides is a mechanism for maintaining rival conjectures long enough for economic experience to select among them.

The implication is subtle but important. Entrepreneurship is valuable not because entrepreneurs know the future better than everyone else. It is valuable because an entrepreneurial economy permits people with different judgments about an unknowable future to act on those judgments. Competition then becomes the institutional process through which those judgments are tested.

What should we measure?

Bento’s argument becomes harder to operationalize once we take entrepreneurial judgment seriously. If competition promotes discovery by sustaining multiple experiments under uncertainty, what exactly should we measure? Conventional indicators—R&D expenditure, patents, firm counts, concentration ratios—capture innovative inputs or market structure reasonably well. They tell us much less about the diversity and independence of the judgments underlying innovative activity.

Bento naturally directs attention to the number of independent technological experiments. The JBA suggests a further question: what makes those experiments genuinely independent? From a judgment-based perspective, what matters for discovery may therefore be not simply how many firms or research projects exist, but whether decision-makers are able to act on genuinely different judgments about an uncertain future.

Firm counts are an imperfect proxy for this. Ten nominally independent firms may pursue essentially the same technological trajectory, imitate one another, or respond similarly to common incentives. Conversely, a large firm may sustain competing research programs based on quite different technological conjectures. Organizational boundaries nevertheless matter because projects within a firm ultimately confront a common hierarchy and internal capital-allocation process. Common ownership can facilitate experimentation by allowing resources to be redeployed among projects, but it can also narrow experimentation if rival conjectures are screened out before markets have had much opportunity to test them.

Ideally, then, we would like to know how much genuinely independent judgment exists within an industry: how many decision-makers can commit resources to distinct technological conjectures, and how different those conjectures actually are. Patent similarity, technological distance among research programs, inventor mobility, venture-capital portfolios, and the organization of internal capital markets might provide partial evidence, but I don’t think there is any obvious sufficient measure.

The case for pluralism

The deepest implication of Bento’s paper is that the welfare consequences of competition cannot be inferred solely from what competition does to the innovative effort of the representative firm.

A concentrated industry can contain formidable research organizations, exploit economies of scale in R&D, and offer large prizes for successful innovation. Those are genuine advantages. However, concentration can simultaneously reduce the number of independent technological bets being placed. Whether the resulting system generates faster technological progress depends on the balance between the scale and incentive advantages emphasized by Schumpeterian reasoning and the experimental diversity emphasized by the Hayekian mechanism.

That is also why the paper cannot really be easily claimed by any either side of contemporary competition-policy debates. It provides no theoretical warrant for the proposition that breaking firms up necessarily accelerates innovation. Equally, impressive R&D spending by dominant firms does not settle the case for concentration. Both arguments focus on only one part of the innovation process.

Rather, the paper suggests that the real question is institutional. What arrangements preserve sufficient rents to induce costly experimentation while maintaining enough independent centres of decision-making to generate technological variety? How do entry barriers, intellectual-property rules, merger policy, research subsidies, capital markets, and organizational boundaries alter not simply total innovative expenditure but the distribution of experiments? And, under what technological conditions does the value of an additional independent draw outweigh the loss of innovative intensity induced by stronger competition? Those questions are perhaps tougher than asking whether competition increases R&D. They are also closer to the economics of innovation under genuine uncertainty and entrepreneurial judgment.

This distinction matters because Demsetz is frequently misunderstood as offering merely another variation on the Coasian question: why do firms exist instead of markets? However, his argument is subtler than this. He does not view firms and markets as opposed mechanisms of coordination, nor does he think firms emerge because markets somehow fail. Instead, he argues that firms are themselves products of the division of labor.

Markets coordinate exchange, but specialization must be embodied in actual productive units. Knowledge does not exist abstractly inside “the market.” It exists only in individuals and organized groups. Firms arise because specialization requires institutions capable of acquiring, maintaining, and deploying costly knowledge. Arguments about specialized assets, moral hazard and team production, etc. are not part of the reasoning. In management research, this perspective is known as the “knowledge-based view of the firm,” after Robert Grant’s highly cited (30k+ GS cites) paper. But, its arguably first systematic (and indeed most systematic) proponent is Demsetz’s 1988 paper. Demsetz pursued this theme in a number of subsequent papers throughout the 1990s that essentially restate and meditate on the argument in the original paper.

Perfect decentralization and the absence of management

The paper begins with a reinterpretation of standard neoclassical economics. What economists call “perfect competition,” Demsetz argues, is more accurately understood as a model of “perfect decentralization.” The classical model abstracts almost entirely from authority, management, uncertainty, and informational asymmetry. Individuals maximize utility or profits under conditions where production techniques and prices are freely known. Under such assumptions, there is very little meaningful role for management. What one firm can do at the level of management, another firm can do equally well. The real tasks of management — that is, exercising judgment under uncertainty, coordinating activities in response to unanticipated chane, directing labor— disappear because the model assumes away the informational conditions that make these tasks necessary in the first place.

Demsetz’s criticism is therefore not merely that markets involve transaction costs. His deeper complaint is that the standard theory suppresses the problem of costly knowledge altogether. If knowledge were genuinely free, universally available, and costless to apply, then there would indeed be little need for firms as organizational structures. But actual economic life is characterized by highly specialized and unevenly distributed knowledge. Economic organization must therefore reflect the costs of acquiring, maintaining, and using this knowledge.

This is the foundation of Demsetz’s specialization theory of the firm. The central economic problem is not simply exchange. It is the organization of specialized knowledge.

Firms within the market order

Importantly, Demsetz’s argument remains deeply compatible with the Smithian and Hayekian defense of decentralized markets. He does not reject the price system. On the contrary, he repeatedly emphasizes that specialization and exchange are inseparable. The firm in the theory of price, he notes, “does not consume what it produces, it sells to others.” Firms therefore do not replace markets. They participate within them. The division of labor generates both specialized production and exchange between specialists.

This point is essential because Demsetz is often read too quickly through a Coasian framework. Coase’s great insight was that markets are not costless to use, and that firms emerge when internal coordination becomes cheaper than market exchange. Demsetz does not deny this, but he believes the transaction cost framework has become too dominant and too narrow. In particular, he argues that economists have overstated the distinction between “market coordination” and “managed coordination.”

His critique of transaction cost economics is devastatingly simple. Firms themselves rely extensively on market exchange. They purchase inputs from outside suppliers, hire labor through contracts, and coordinate activities through negotiated relationships. Meanwhile, market exchange itself often involves continuity, direction, and long-term association. The clean conceptual distinction between “firm” and “market” therefore begins to dissolve.

Indeed, Demsetz argues that even in a hypothetical world of zero transaction costs, firms would still exist. This is one of the paper’s most radical claims. Transaction cost theory often implies that firms are responses to costly market exchange. But Demsetz insists that productive cooperation among specialists would persist even if negotiation itself were costless. The reason is that specialization requires organized deployment of knowledge, not merely cheap contracting.

Specialization and costly knowledge

This insight allows Demsetz to shift the theory of the firm onto a different foundation entirely. The crucial fact about knowledge is that it is both costly and specialized. Individuals can learn deeply only within narrow domains. Yet the productive use of specialized knowledge requires integration with the complementary expertise of others. Modern economic organization therefore emerges from the asymmetry between the economics of learning and the economics of use.

Demsetz formulates this point with remarkable clarity: “Firms and industries must form a pattern of economic organization that takes account of the need for acquiring knowledge in a more specialized fashion than the manner in which it will be used.” This sentence contains the core of the theory.

Knowledge is acquired most efficiently through specialization. But the use of knowledge cannot require everyone to master every relevant body of expertise, because that would destroy the gains from specialization itself. Society therefore requires mechanisms that allow specialists to deploy their knowledge without transferring the underlying expertise to everyone else.

“Direction substitutes for education”

This is where Demsetz introduces one of his most original ideas: “direction substitutes for education.” Rather than educating every participant in the entire body of knowledge relevant to production, organizations economize by transmitting instructions. Workers, suppliers, and consumers can use products and services without understanding in detail how they are produced. The user of industrial chemicals need not become a chemical engineer. Construction firms need not master steel production. Knowledge remains specialized, while instructions permit coordinated use by non-specialists.

This idea transforms the role of management. In Demsetz’s framework, management is not fundamentally about legal authority or command in the political sense. It is an informational function. Managers direct activities because they possess knowledge that others do not possess. Organizational direction economizes on the costs of transmitting specialized expertise through formal education.

Vertical integration and the boundaries of the firm

The implications for vertical integration are especially important. Demsetz argues that the boundaries of the firm are determined not simply by transaction costs or opportunism, but by the economics of conserving expenditures on knowledge. Production remains integrated within a single firm only so long as it is efficient for one organization to maintain the relevant stock of expertise. Once downstream uses become too diverse, however, it becomes uneconomical for a single firm to possess all the required knowledge. At that point, products are sold to other firms possessing specialized downstream capabilities.

This is an extraordinarily sophisticated account because it explains both integration and fragmentation through the same logic. Specialization creates firms, but specialization also creates markets between firms. The modern economy becomes a network of specialized repositories of knowledge linked together through exchange.

Demsetz and Conner & Prahalad

The influence of Demsetz’s argument can be seen clearly in later knowledge-based theories of the firm, especially the work of Conner and Prahalad. Their 1996 paper (which was written as a response to a paper by me), “A Resource-Based Theory of the Firm: Knowledge versus Opportunism,” can be read as an extension of Demsetz’s core insight. Like Demsetz, they argue that firms exist not primarily because of opportunism, but because knowledge is costly, specialized, and unevenly distributed across individuals. They also adopt Demsetz’s central epistemic claim that “direction substitutes for education.”

Their distinctive contribution is to sharpen this idea into an explicit Simonian theory of authority. Under market contracting, individuals retain autonomous judgment and must internalize knowledge before acting upon it. Under firm organization, by contrast, managers can direct employees before employees fully understand or agree. Conner and Prahalad call this “knowledge substitution”: the manager’s judgment substitutes for the employee’s own.

Demsetz and Cheung

Another important intellectual parallel is Steven Cheung’s “The Contractual Nature of the Firm” (1983), which, like Demsetz, dissolves the sharp distinction between firms and markets (see my earlier post on Cheung’s paper). Cheung argues that what economists call “the firm” is not a discrete institutional category but a particular pattern of contractual substitution under positive transaction costs. Where Demsetz emphasizes the economics of specialized knowledge, Cheung emphasizes the economics of measurement and pricing. In Cheung’s account, firms emerge because directly measuring and pricing every productive contribution is prohibitively costly; proxy measurements such as wages therefore replace direct pricing, and managerial direction expands accordingly. Yet despite these differences, the two theories converge in striking ways. Both reject rigid firm-market dichotomies, both treat organization as a continuum of contractual arrangements, and both argue that “direction” arises not because hierarchy is inherently superior, but because informational constraints make decentralized pricing incomplete. Cheung’s claim that delegation of use rights increases when direct pricing becomes difficult closely parallels Demsetz’s later insight that “direction substitutes for education.” Taken together, the two papers suggest that the modern firm is best understood not as a categorical alternative to markets, but as an evolving organizational response to the combined problems of costly knowledge, costly measurement, and incomplete pricing.

The problem of defining the firm

However, the very elegance of this framework creates several conceptual problems. The first is that Demsetz gradually broadens the concept of the firm until it risks losing analytical distinctiveness. He explicitly treats one-person enterprises as firms and repeatedly emphasizes that “firm-like” organization exists in degrees rather than as a sharp institutional category. Independent contractors, employees, and long-term collaborators may differ less in economic substance than legal terminology suggests.

This flexibility gives the theory considerable descriptive realism. Actual economic organization does blur the boundaries between markets and firms. But the move also creates a serious analytical problem. If firms and markets differ only continuously, then what exactly is the explanandum? Traditional theories of the firm attempt to explain why some activities occur within firms rather than through market contracting. Demsetz instead partially dissolves the distinction itself.

At times the theory comes close to tautology. Specialists produce for others; firms are specialized productive units; therefore firms exist because specialization exists. But if “firm” simply means “specialized producer,” then the theory risks explaining firms by presupposing them conceptually. Demsetz himself seems aware of this difficulty when he admits that “we have no clear notion of firm-like contractual arrangements.”

The limits of direction

A second problem concerns the role of direction. Demsetz’s claim that “direction substitutes for education” is both powerful and incomplete. Direction can economize on knowledge transfer only up to a point. Effective direction itself often requires tacit understanding, local judgment, and contextual interpretation. Particularly in innovative or uncertain environments, compressed instructions cannot fully substitute for distributed expertise.

Here the theory encounters a tension with Hayekian ideas about dispersed knowledge. Demsetz recognizes that knowledge is decentralized and costly. But, Hayek’s critique of central planning rested precisely on the inability of centralized direction to mobilize local tacit knowledge effectively. Demsetz avoids advocating comprehensive planning, but his emphasis on direction risks understating the extent to which decentralized judgment remains necessary even within firms.

Predictive weaknesses

A third problem concerns prediction. Because specialization simultaneously generates integration and fragmentation, the theory may struggle to produce determinate organizational predictions. The same logic that explains the existence of firms also explains outsourcing, modular production, and inter-firm markets. Firms internalize activities when knowledge complementarities favor integration; they disintegrate when downstream uses become too diverse.

This flexibility may accurately describe modern capitalism, but it weakens predictive precision. The theory can sometimes explain almost any organizational arrangement retrospectively.

There is also a deeper sociological limitation. Demsetz consistently treats organizational forms as responses to informational and productive efficiency. Long-term employment relationships, for example, survive because continuing association allows firm-specific knowledge to accumulate cheaply. But, real firms persist for many reasons that cannot be reduced to knowledge conservation: political bargaining, financing structures, regulation, organizational inertia, and historical accidents all matter substantially.

The theory therefore risks excessive functionalism. It often moves too quickly from “this arrangement economizes on knowledge costs” to “therefore this arrangement exists.” However, institutions are not always selected efficiently, nor do they always survive because they minimize informational costs.

Authority and organization

Finally, Demsetz underdevelops the problem of authority itself. By emphasizing continuity, specialization, and informational direction, he tends to minimize the distinctive legal and organizational powers associated with firms. Employers can reassign tasks, discipline workers, restructure production, and impose routines in ways ordinary market relationships frequently cannot. Direction inside firms is not merely informational; it is also institutional and juridical.

None of these criticisms diminish the importance of the paper. If anything, they reveal how ambitious it is. Demsetz is attempting to relocate the foundations of the theory of the firm away from narrow concerns with opportunism and toward a broader account of economic organization under conditions of costly knowledge.

Conclusion

The enduring significance of “The Theory of the Firm Revisited” lies precisely here. The firm is not fundamentally a substitute for markets. It is a specialized repository of knowledge embedded within a decentralized system of exchange. Markets coordinate transactions, but firms embody expertise.

That insight remains remarkably relevant today. Modern economies are increasingly organized around specialized human capital, modular supply chains, and knowledge-intensive production. The central economic problem is often not the allocation of homogeneous inputs, but the integration of dispersed expertise. In such an environment, Demsetz’s emphasis on specialization and costly knowledge appears extraordinarily prescient.

At the same time, the ambiguities of the theory remain unresolved. The more convincingly Demsetz dissolves the boundary between market and firm, the more difficult it becomes to explain what makes firms institutionally distinctive. His theory brilliantly links specialization to organized production, but less clearly why that organization takes the particular forms we call firms.

That tension may ultimately reflect a deeper truth. Perhaps firms are not sharply bounded institutions standing opposite markets, but evolving patterns of specialized cooperation embedded within the market order itself. If so, Demsetz’s theory succeeds not despite its blurred boundaries, but because modern capitalism increasingly blurs them as well.

Running through all these debates is a broader issue about what higher education is. For the ancient Greeks, education was paideia—a holistic system of upbringing and cultural training, not just schooling but the complete formation of the human soul, mind, and body to produce a well-rounded, virtuous citizen who can participate in civic life. Modern education seems far-removed from this ideal—increasingly vocational and transactional, preposterously expensive, and vulnerable to competition from more efficient alternatives (including those powered by AI).

We don’t need no education—or do we?

An implication of the classical view is that students (and perhaps even parents) don’t know what they want; the educator’s job is to tell them. After all, education isn’t a commodity like toothpaste—we don’t want a system in which providers compete to offer consumers the education goods and services they want at the lowest prices. From this perspective, AI tools—however useful in teaching specific skills or tasks like car repair or sourdough baking—are not a threat to the conventional educational model.

Let me explain why I think this view is mistaken.

For starters, it’s true that education—like health care, privacy, and relationship-building—is a particular type of commodity. It is what economists call an experience good, exchanged in markets fraught with asymmetric information. It is governed by a byzantine set of regulations and legal restrictions supported by one the most powerful political lobby associations in most countries. Consuming education is very different from consuming toothpaste.

And yet, educational activities take place in the real world, involving the use of scarce resources (land, buildings, books, equipment, teacher and administrator labor, and more) to produce specific goods and services (books, articles, lectures, exams, certificates, degrees, etc.). Call these education goods. While there is no market for “education” as an abstract category, there absolutely are markets for (and non-market, state provision of) educational goods and services. Confusing the abstract category with the actual goods and services leads many people astray. (The same is true for similar categories like health care and consumer privacy.)

Thinking about education goods and services

Focusing on the production and consumption of education goods raises many interesting questions.

What is the production function? People don’t consume “education,” any more than they consume love or friendship or truth or beauty. Instead, people exchange specific, discrete units of educational goods and services (books, lectures, degree programs) which are produced using capital and labor. There is nothing magic about education in this sense. As a lover and collector (and occasional producer) of books, I regard them as special entities—books open your mind, they change the way you think, they transform your experience, and so on. But I also recognize that a book is produced using capital (paper, ink, binding materials, machinery, transportation infrastructure, digital networks and devices, the financial system) and labor (from authors, editors, publishers, marketers, etc.).

Who is the consumer? Critics of the student-as-consumer model are correct, at least for primary and secondary education. The consumer is the parent or some institution or entity playing the parental role. For tertiary education and “education” in the broad sense (e.g., buying a book, attending a paid lecture, subscribing to a podcast, hiring a consultant), the person receiving the education is, literally, the consumer—the person who decides whether or not to engage in the transaction. Do such consumers make “wise” choices? Who knows. But arguments for paternalism and other forms of regulation must proceed on the merits and do not follow from the general claim that education is somehow “different” from other goods and services.

How is production organized? While the institutions of higher education have changed dramatically over the centuries, in some sense the basic model has changed very little. Teachers are employed by schools, schools are members of larger networks (in the old days, typically governed and supported by religious bodies, now by secular ones), courses are arranged into curricula or degree programs which are certified by some higher authority, students are certified through writing and examinations, and so on. Some teachers are paid on a piece-rate basis but most are salaried (and in some cases, protected by de jure or de facto tenure which allows them to speak and write with limited outside interference). Most faculty engage in some combination of research, teaching, and service, though the allocation of effort varies widely across faculty categories and different types of institutions (e.g., research universities, liberal arts colleges, technical and vocational schools).

While we have many more subject areas, majors, certificates, and degree programs than our predecessors—let alone labs, startup incubators, technology commercialization programs, and in the U.S., semi-professional sports teams—the organization of the industry has remained remarkably stable over a long time. As it is sometimes put, who is the Henry Ford or Steve Jobs of higher education? Where are the disruptive innovators who radically alter the structure of the industry? Why haven’t big changes in technology, demographics, and funding led to a shakeup? For an industry that (in recent years, anyway) prioritizes diversity, why is there so little diversity?

I think regulation (and the resulting cartelization) is a large part of the answer, though there are many other explanations, as discussed in this post.

What are we actually exchanging? It often helps in these discussions to distinguish average from marginal effects. When we consume a complex, larger good or service, part of that transaction often includes our choice to delegate the details to the provider. If I book a tour of the Acropolis, part of what I’m paying for is the tour company’s judgment about what parts to visit, how long to stay in each section, when is the best time to go, what I should look for on each structure, and so on. I’m choosing “Teach me about the Acropolis” but asking the provider to use his or her specialized knowledge and expertise to say what specific activities that involves, on the margin.

This perspective challenges the view that only the teacher can decide what should be taught. In reality, teachers can offer their expertise, track record, and the like—including their authority to decide the specific contents of an educational program—to students (or parents or community members) who pick the teacher. Nothing about the teacher-as-expert or teacher-who-tells-you-what-you-should-want model implies that educational goods and services cannot be produced, priced, exchanged, and consumed in the same fundamental way as other goods and services. The market for toothpaste, it turns out, isn’t actually that different.

Technology and the future

Will AI be the revolutionary, disruptive technology that forever alters the industry? Will it encourage a greater variety of educational types, patterns, providers, units, categories, and so on? Will students still attend (mostly) in person, will they have majors and minors, will they earn degrees, will they seek networking and certification as well as learning? Will new players emerge as in Christensen’s famous model of industry disruption? Will the abstract notion of education be transformed as the production and distribution of specific, discrete quantities of educational goods and services evolves? Who knows, but I am eager to see what happens.

Demsetz’ 1982 paper is widely cited for its discussion of, well, entry barriers. This is somewhat unfortunate, because the paper is not just about entry barriers but about something, depending on your taste, more interesting.

Its real subject is a recurring tendency in industrial organization economics: the tendency to observe persistent firm advantages and then redescribe those advantages as evidence of monopoly. Whether the phenomenon is concentration, advertising, reputation, scale economies, or above-normal profits, the underlying inference is often the same. If competitors have difficulty replicating an advantage, economists become tempted to treat the advantage itself as suspicious. Demsetz was having none of it.

The paper’s central question is deceptively simple: Why should we assume that successful firms are easy to imitate? Once posed, the question proves remarkably difficult for much of the traditional barriers-to-entry literature to answer.

In this respect, the paper belongs to a broader intellectual project that runs through much of Demsetz’s work. The 1982 paper can be read as a companion to his earlier critique of the concentration-profit relationship (e.g., the famous 1973 paper mentioned above). In both cases, the target is not a particular theory but a style of inference. Economists observe persistent differences among firms and then too quickly conclude that competition has somehow failed. Demsetz repeatedly asks whether the same observations might instead reflect the outcomes of competition itself.

Seen in this light, the paper is not primarily about entry barriers. It is about how economists explain persistent differences among firms.

The postwar suspicion of profit

The intellectual backdrop matters. Postwar industrial organization was heavily shaped by the structure-conduct-performance paradigm and its suspicion of numerous real market phenomena and practices: persistent profits were suspicious: concentrated markets were suspicious; advertising was suspicious; product differentiation was suspicious. The implicit benchmark was a competitive equilibrium in which unimpeded entry would erode rents and firms converged toward some common condition of normal returns. When such convergence failed to occur, economists looked for barriers.

Demsetz’s paper is, in part, an attack on this mode of reasoning. More precisely, it is an attack on the tendency to infer monopoly power from persistent asymmetries without first explaining how those asymmetries arose.

Demsetz’s notes that writers such as Bain, Stigler, and Ferguson offer different definitions of barriers, but then point out that their definitions embody different theories of competition and therefore classify the same phenomena differently. Bain and Ferguson define barriers partly in terms of the outcomes they are meant to explain, such as persistent supra-normal profits or monopoly returns. Demsetz worries that this encourages economists to infer barriers from the very evidence that requires explanation. Stigler’s definition avoids some of these difficulties by focusing on cost asymmetries between incumbents and entrants. Even here, however, Demsetz insists that economists must ask how those asymmetries arose. A reputation, a brand name, or an organizational capability may disadvantage entrants, but that does not imply that competition has been impaired.

The central issue is not whether an asymmetry exists but whether it reflects anti-competitive restrictions or the ordinary consequences of investment, learning, innovation, and successful rivalry. That point may sound semantic, but it is not. Definitions matter because they shape the inferences economists draw from observed market outcomes. If one begins with the presumption that persistent profits are evidence of monopoly, barriers become almost unavoidable. If one begins by asking how firms created their advantages, however, a very different picture emerges.

Consider a firm such as Costco. Its advantages are substantial and difficult to replicate. Supplier relationships, organizational routines, customer trust, employee retention, purchasing scale, and membership economics have been accumulated over decades. Should these advantages be interpreted as evidence of barriers to entry? Or, should they be interpreted as evidence that the firm has been unusually successful at solving a complex set of organizational and logistical problems?

For Demsetz, the answer depends on how the advantage emerged. That question is prior to the question of whether the advantage should concern us.

Property rights are ”barriers” too

One of the most interesting parts of the paper is Demsetz’s observation that many institutions economists ordinarily defend are themselves barriers to entry.

A taxi medallion system limits entry, but so do patents, trademarks, property rights, and many forms of contract enforcement. The mere existence of a hurdle tells us almost nothing. One must ask what function the hurdle serves.

Take pharmaceutical patents. A patent prevents competitors from immediately copying a newly developed drug. In a literal sense, it is an artificial restriction on entry. But if generic firms could freely imitate the product the day after its invention, the original innovator might never have invested billions into R&D in the first place.

Or consider trademarks. Apple’s trademark prevents another electronics company from selling phones labeled ”iPhone.” That restriction reduces imitation and gives Apple pricing power over its branded products. But absent trademark protection, consumers would face a world flooded with counterfeit goods, confusion about quality, and weaker incentives for firms to invest in reputational capital.

Demsetz’s point isn’t t that all such protections are optimal, but rather that economists too casually equate unrestricted imitation with competition and too casually treat restrictions on imitation as anti-competitive. Indeed, he makes the striking point that the problem of defining property rights is ”precisely that of creating properly scaled legal barriers to entry.”

That is a much richer way to think about markets than the simplistic dichotomy between ”competition” and ”barriers.” In strategy, the resource-based view would eventually argue that firms sustain superior performance when they possess valuable assets that competitors cannot easily imitate. Demsetz can be read as providing part of the economic logic underlying that claim. Legal protections, reputational histories, and organizational learning are not incidental frictions sitting outside competition, but are among the mechanisms that make long-term investment and capability accumulation worth engaging in.

Reputation as an economic asset

The heart of the paper lies in Demsetz’s discussion of advertising, reputation, and information costs. The traditional industrial organization literature often treated advertising as an exclusionary device that insulated incumbents from competition. Demsetz instead begins from a more fundamental observation: information is costly.

Consumers do not perfectly know product quality, reliability, durability, or future performance. Firms therefore invest in reputation, brand names, and promotional activity because these economize on uncertainty. As he puts it, ”A reputable history is an asset to the firm possessing it and to the buyer who relies on it because information is not free.” This is arguably the paper’s central economic insight.

Imagine two restaurants opening across the street from one another. One is a new independent restaurant. The other is a new location of a well-known chain with decades of accumulated reputation. Even if the food quality were objectively identical, consumers would not treat the two establishments as equivalent. The chain benefits from the informational capital embedded in its history. Customers believe—quite rationally—that the probability of catastrophic disappointment is lower.

Or consider enterprise software. A Fortune 500 company deciding whether to buy cloud infrastructure from Microsoft or from a five-person startup is not simply comparing current technical specifications. It is also comparing reliability, continuity, support capacity, and survivability. The incumbent’s long history functions as an economic asset because it reduces uncertainty for buyers.

Demsetz’s point is that these asymmetries are not necessarily signs of market failure. They may instead reflect the normal consequences of competition unfolding through time under conditions of costly information.

This is also where one can see the beginnings of later strategy ideas about intangible assets. In much resource-based strategy research, reputation, brand equity, organizational routines, and accumulated know-how are treated as sources of sustained advantage because they are difficult to imitate. Demsetz effectively provides the microeconomic logic underneath that claim.

His answer rests on information costs, uncertainty, property rights, and historical accumulation. Reputation matters because information is costly. Brand capital matters because trust economizes on uncertainty. Organizational capabilities matter because learning is cumulative and often difficult to replicate. Time itself becomes economically productive.

Competition creates advantages that later look like barriers. A firm that consistently performs well accumulates customer trust, organizational learning, supplier relationships, distribution capabilities, and reputational capital. These advantages may subsequently make entry more difficult. But it does not follow that they were created through anti-competitive conduct.

Consider Toyota. Much of Toyota’s historical advantage did not reside in a single patent or machine. It resided in production routines, supplier coordination, quality processes, and tacit organizational knowledge accumulated over decades. From a static perspective, these advantages can appear as barriers. From a historical perspective, they are the result of sustained investment and learning. The same can be said of firms such as Costco: their advantages were not simply inherited; they were built.

What is striking in retrospect is how much of the barriers literature treated successful firms almost as a problem requiring explanation. Durable advantages were presumed suspicious. So were persistent profits were presumed suspicious and indeed difficult-to-imitate capabilities.

While Demsetz’s acknowledges that monopoly may exist, his problem is with the inference from persistence to monopoly. A firm may earn above-normal returns because rivals are excluded. But it may also earn above-normal returns because it has discovered better methods, developed superior capabilities, accumulated reputational capital, or made investments whose value only becomes visible over time. Observing persistence does not tell us which explanation is correct.

To say that a firm’s advantage emerged through superior performance is not to say that all subsequent uses of that advantage are socially benign. A firm may acquire dominance through innovation, reputation, or organizational excellence and subsequently engage in conduct that restricts competition. Demsetz’s contribution is not to deny this possibility. Rather, it is to insist that the origin of an advantage and the use of an advantage are analytically distinct questions. Much of the barriers literature blurred the distinction by treating the existence of a durable advantage as evidence that competition had already failed.

This point links the barriers paper directly to Demsetz’s earlier critique of the concentration-profitability literature. In both cases, the target is not a particular empirical finding but a particular style of reasoning. Economists observe differences among firms and then too quickly conclude that competition has somehow failed. Demsetz repeatedly asks whether the same evidence might instead reflect the outcomes of competition.

The predatory pricing problem

The paper’s long discussion of predatory pricing is fascinating because it shows how difficult it is to distinguish harmful from competitive behavior in practice.

The standard legal doctrine sought to identify predatory pricing by asking whether prices fell below marginal cost. Demsetz argues that this framework is much less straightforward than its proponents assume.

Firms frequently price below current cost for entirely legitimate reasons. A streaming service may offer a free trial. A grocery store may sell rotisserie chickens at a loss. A video game console may initially be sold below manufacturing cost in order to build an installed base.

These practices are ordinarily understood as competitive rather than predatory because the relevant economic calculation concerns future revenues generated by customer acquisition. Once this is recognized, however, the distinction between promotional pricing and predatory pricing becomes considerably more difficult to draw. In both cases, firms absorb losses in the short run in anticipation of future gains.

Demsetz’s point is not that predation never occurs; it is that economists and regulators often underestimate the difficulty of distinguishing predation from aggressive competition.

That insight remains highly relevant in debates over platform businesses, cloud computing, artificial intelligence, and digital ecosystems. When a firm prices aggressively, is it attempting to eliminate rivals? Or is it competing vigorously in pursuit of scale economies, learning effects, or future efficiencies?The answer is often less obvious than antitrust discussions assume.

Demsetz and the evolution of industrial organization

One of the interesting ironies of the paper is that much of later industrial organization accepted Demsetz’s criticism while simultaneously moving in a different direction. I vividly remember a conversation towards the end of the 1990s with a CBS colleague who had spent time at UCLA as visiting faculty, and, still awestruck, reported that the “young game theorists were laughing at Demsetz.” It was clear that the colleague sided with the “young game theorists” (and the remark was intended as a put-down of my praise of Demsetz). Another colleague who was eavesdropping interjected: “Then let them laugh.” The latter colleague was an industrial organization economist. And he was right.

In fact, the game-theoretic revolution largely abandoned the crude structure-conduct-performance reasoning that Demsetz criticized. Modern industrial organization is far more attentive to mechanisms than its predecessors, and economists no longer naïvely infer monopoly power from concentration alone (antitrust authorities still often do this, however). Instead, they ask how particular strategic interactions generate durable advantages.

In this respect, Demsetz won. However, a deeper difference remains.

Demsetz’s instinct is fundamentally comparative-institutional. His question is always: Compared to what? Compared to what feasible alternative arrangement? Compared to what realistic institutional structure? Compared to what competing explanation?

The game theorist often asks a different question: Is there an equilibrium in which this behavior occurs? Can advertising deter entry? Can reputation support predation? Can excess capacity discourage rivals? Can network effects generate durable market power? The answer is often yes. It is fundamentally a possibility exercise.

From a Demsetzian perspective, however, that answer is only the beginning of the inquiry. The existence of a theoretical possibility does not establish that it provides the best explanation of observed outcomes.

One might therefore say that Chicago and post-Chicago approaches differ less in their conclusions than in their standards of evidence. The post-Chicago theorist demonstrates that a strategic mechanism can exist. The Demsetzian asks whether there is good reason to believe that this mechanism is actually driving the observed outcome rather than some competing explanation rooted in efficiency, information costs, learning, or institutional arrangements.

This is not a disagreement about theory. It is a disagreement about explanation.

The broader lesson

Demsetz’ paper ultimately becomes a meditation on frictions in economic systems. Economists often speak as though the ideal market is frictionless and any impediment to instantaneous competition represents a distortion. Demsetz argues instead that many frictions are economically productive.

Property rights create friction, reputation creates frictions, consumer uncertainty creates frictions, and organizational history creates frictions. But, these same frictions also support innovation, investment, trust, and long-term cooperation. (See my earlier post on G B Richardson).

The challenge, therefore, is not to eliminate all barriers, but rather to distinguish socially productive frictions from socially harmful ones. That turns is perhaps also a more difficult intellectual problem than much of industrial organization economics perhaps has acknowledged (and admittedly at a somewhat different level of analysis).

This is why the paper still feels fresh more than forty years later. It is not merely an article about barriers to entry. It is a challenge to a recurring temptation in economic analysis: the temptation to observe a durable advantage and immediately treat it as evidence that competition has somehow failed.

Demsetz’s response is characteristically simple. Before concluding that an advantage reflects monopoly, one should first ask where it came from. That question remains as relevant today as it was in 1982.

Subscribe now

Classic economic theories of the firm

I wrote a short piece on this in 2012, inspired by a presentation by Demsetz at the ISNIE (now SIOE) conference in a session honoring Yoram Barzel on his 80th birthday (I took the photo below). Demsetz pointed out that Coase (1937) defines the firm in terms of the employment relation. A one-person operation, in this definition, is not a firm, and vertical integration deals with the question of adding producers of intermediate products to the firm’s employment roll. Demsetz thinks independent contractors are firms, and hence it makes little sense to speak of “firm” and “market” as alternatives, as Coase does. (Oliver Williamson, during an earlier session, noted that Coase expressed more interest in intermediate product markets in his 1988 article than in “The Nature of the Firm.”)

For Knight, Williamson, Hart, and Foss and Klein, in contrast, the firm is defined not by the employment relationship, but by the ownership of alienable assets. In this approach, the question is who owns what, not who is employed by whom. Of course, even in the Knightian approach, to get from the one-person firm to the multi-person firm requires some theory about the relative transaction costs of employment versus independent contracting, a theory Nicolai and I try to provide in chapter 8 of our 2012 book, focusing on the conditions under which the entrepreneur can delegate judgment to subordinates.

The firm as a nexus of contracts

Cheung, as Nicolai points out, holds a position closer to what we now call the nexus-of-contracts approach. For Alchian and Demsetz and (even more so) for Jensen and Meckling, the “firm” is a label we assign to a set of individuals creating joint value via team production. The team includes people legally classified as employees and owners, but also suppliers, distributers, and collaborators who are independent contractors or employees of other firms. (In a famous rhetorical flourish, Jensen and Meckling—ironically, the bêtes noires of many contemporary stakeholder theorists—write that “most organizations are simply legal fictions which serve as a nexus for a set of contracting relationships among individuals.”) In this approach, questions of firm boundaries—who is “inside” or “outside” the firm—are largely irrelevant. The interesting issues involve the nature of the relationships among team members, whatever the formal legal status of those relationships.

A feature of Alchian and Demsetz’s (1972) approach is that owners (as residual claimants) are simply part of the team, with no special status other than their role as monitors who identify and limit shirking on the part of other team members. One notable implication is that bosses don’t necessarily hire workers; workers may just as easily hire bosses. Recall Cheung’s (1983, p. 8) famous illustration: “My own favorite example is riverboat pulling in China before the communist regime, when a large group of workers marched along the shore towing a good-sized wooden boat. The unique interest of this example is that the collaborators actually agreed to the hiring of a monitor to whip them.” In Alchian and Demsetz’s example, the employee can “fire” his employer by quitting, just as I can “fire” my grocer by shopping at a different store.

Ownership, authority, and residual control

A problem with the nexus-of-contracts approach is that, as Nicolai and I have argued in various works, ownership is in fact a distinct economic function, the key element of which is not residual claims (which can be assigned to non-owners) but residual control. Ownership is the right to decide what to do in conditions not specified by prior agreement. Loosely speaking, ownership is not just another factor of production, but the “controlling factor” that designs and implements the procedures and practices—the organizational “rule of the game”—within which the other team members operate. The ownership function can also be exercised with greater or lesser capability or ownership competence. As Ludwig Lachmann (1956) put it, in the context of owners and hired managers, even those with substantial day-to-day decision rights: “We might . . . distinguish between the [owner] and the [manager]. The only significant difference between the two lies in that the specifying and modifying decisions of the manager presuppose and are consequent upon the decisions of the [owner]. If we like, we may say that the latter’s decisions are of a ‘higher order.’”

These ideas are rooted in Knight (1921) who foregrounds the idea of resource or asset ownership as the defining characteristic of a firm: the firm is defined as an entrepreneur plus the alienable assets she owns and controls. The entrepreneur may or may not have employees—the key is ownership, not the nature of the contractual relationship with other stakeholders. The Grossman and Hart property rights view comes out of this tradition as well.

Williamson also distinguishes sharply between employees and non-employees (or intra-firm and inter-firm transactions). He invokes the notion of “forbearance,” the idea that the law treats the firm itself as the arbiter of intra-firm disputes, while the courts are more likely to intervene in disputes between firms. (See the discussion in his 1991 article, especially pp. 98-100 of the version that appears in The Mechanisms of Governance.)

The firm as a reference point

Hart’s more recent work takes a different tack, treating the firm as a “reference point” for organizing interactions among stakeholders (see here and here). Drawing on concepts from behavioral economics, Hart argues that parties enter transactions with some notion of fairness, and renege on commitments or shirk on performance to the extent that they feel aggrieved, or shortchanged relative to this reference point. This explains why contracts are incomplete (a response to criticisms from Eric Maskin and Jean Tirole of the standard explanations for incompleteness)—incomplete contracts leave room for the ex post renegotiation that is needed as new circumstances arise. In other words, the contracts that constitute the firm are written not to specify future interactions, but to provide a general framework within which bargaining and negotiation can take place.

The bottom line

So, what is a firm? The answer probably depends on the problem one wants to solve. In other words, perhaps the better question is, What are the important research questions that can be answered when the firm is defined as X?

While Cheung may be underappreciated by the profession, he is certainly not lacking in self-esteem (at least in his public self-presentation) and extraordinary confidence in his own intellectual abilities and results. In autobiographical reminiscences published in 2016, Cheung states things like: “if one does not mind the tedious task of going through the Chinese Internet, one would find that I am credited as the founder of contract economics, Coase the founder of transaction costs, and Alchian the founder of property rights.”

Cheung has indeed written several brilliant papers that focus on contracts, property rights, and transaction costs. In “The Theory of Share Tenancy “(1969), he argued that sharecropping can be an efficient contractual arrangement rather than an inherently exploitative one. In “The Fable of the Bees” (1973), he examined externalities and showed how clearly defined property rights can facilitate private solutions. In “A Theory of Price Control” (1974), he analyzed how price ceilings generate non-price forms of competition and resource allocation. Across these works, Cheung emphasized how institutions and contractual arrangements shape economic outcomes in the presence of transaction costs.

His 1983 paper “The Contractual Nature of the Firm” is perhaps the clearest window into both his intellectual style and his place within what might loosely be called the UCLA property-rights tradition. The paper is ostensibly about Ronald Coase’s famous 1937 article “The Nature of the Firm.” But, in actuality it is doing something more subtle and radical. Cheung is simultaneously interpreting Coase, extending Coase, correcting Coase, and dramatically radicalizing Coase. While I don’t agree with Cheung, I think the result is one of the most elegant and provocative essays ever written about the theory of the firm.

The UCLA counterrevolution

To understand Cheung, however, one first has to understand the intellectual environment from which he emerged. The modern economics curriculum often presents neoclassical economics as if it evolved smoothly from marginalism to general equilibrium to game theory. In reality, a major intellectual rupture occurred in postwar microeconomics. During the 1950s and 1960s, a small cluster of economists associated with UCLA and Chicago began systematically questioning one of the deepest assumptions embedded in standard price theory: the assumption that exchange itself was frictionless.

Ronald Coase opened the door in “The Nature of the Firm” (1937), where he posed an extraordinarily simple but radical question: if markets are so efficient, why do firms exist at all? Why are so many economic activities coordinated internally through managerial authority rather than through prices? His answer was transaction costs. Using markets is costly. Discovering prices is costly. Negotiating contracts is costly. Monitoring performance is costly. Firms arise because, under some conditions, administrative coordination is cheaper than market contracting.

But, Coase’s original paper was more suggestive than fully developed. In fact, part of the fascination of reading Cheung’s 1983 paper is seeing how much interpretive uncertainty still surrounded Coase’s argument almost fifty years later. Cheung explicitly notes that economists had produced wildly different readings of Coase’s thesis and that “we do not exactly know what the firm is—nor is it vital to know.”

That sentence captures something essential about Cheung’s style. He was deeply theoretical but profoundly anti-essentialist (if that is the right word). He distrusted abstract categories detached from observable contractual arrangements. Again and again in his work, he dissolved familiar economic entities into structures of contracts, measurement problems, and property-right allocations.

Alchian, property rights, and institutional realism

This placed him intellectually close to Armen Alchian and Harold Demsetz, perhaps the most underrated economic theorists of the postwar period (scandalously, neither received the Nobel Prize). Both shared Coase’s skepticism toward frictionless textbook models but added an evolutionary and property-rights perspective that profoundly influenced modern organizational economics. Alchian and Demsetz viewed economic institutions less as idealized solutions than as adaptive responses to constraints on information and measurement. Economic organization emerged not from optimization in the abstract but from trial-and-error adjustments to real-world transaction costs.

Cheung absorbed this orientation completely. Indeed, one of the striking features of “The Contractual Nature of the Firm” is how highly concrete it is. Unlike much later work in organizational economics, which increasingly drifted toward abstract formalization (although it is also fair to say that org econ has very much come back to reality over the last couple of decades or so), Cheung remained obsessed with actual contractual practices.

The paper is filled with examples involving hardwood-floor subcontracting in Hong Kong, shirt exporters, bead-stringing homeworkers, and piece-rate manufacturing systems. These examples are not illustrations added after the fact. There is a sense in which they are partly are the theory, laying out mechanisms at a lower level of abstraction.

Cheung’s central move in the paper is deceptively simple. He argues that firms should not be understood primarily as sharply bounded entities replacing markets. Instead, firms are particular forms of contractual arrangements that emerge when direct pricing of every contribution becomes too costly (an idea anticipated in Alchian and Demsetz’ 1972 paper). Referring to Coase’s 1937 paper, Cheung says that it not quite correct to say that a ‘firm’ supersedes ‘the market,’” he writes. “Rather, one type of contract supersedes another type.”

The measurement-cost revolution

The conceptual implications are enormous. In standard textbook economics, markets and firms appear as fundamentally different coordinating mechanisms. Markets use prices; firms use hierarchy. But Cheung insists that this distinction is far less clear than economists imagine. The world is filled with hybrid arrangements: subcontracting, franchising, leasing, sharecropping, piece-rate labor, licensing, partnerships, vertical agreements. The economy is not neatly divided into “markets” and “firms.” It is a continuous spectrum of contractual forms.

This insight places Cheung very close to another great economist in this tradition: Yoram Barzel. Barzel’s work on measurement costs and property rights, clearly initially heavily inspired by Cheung, pushed transaction-cost economics in an even more radical direction. His core argument was that economic organization is fundamentally shaped by what can and cannot be measured economically. Property rights are never perfectly specified because measuring attributes is costly. Contracts therefore emerge as imperfect attempts to allocate rights under conditions of costly measurement. An early classic statement of this is Barzel’s 1982 paper, “Measurement Cost and the Organization of Markets.”

Cheung’s 1983 paper is saturated with this logic. Indeed, the entire essay can almost be read as an extended meditation on measurement problems. He repeatedly argues that firms arise because directly measuring and pricing every contribution is prohibitively costly. Piece-rate contracts occupy an especially important role in the paper because they sit “squarely in between the market and what Coase called the firm.”

Under pure market exchange, every contribution would theoretically be separately measured and priced. Under a pure wage contract, contributions are not directly priced at all; workers are instead paid according to proxy measures such as hours worked. Piece rates sit between these extremes because they directly price some dimensions of output while leaving other dimensions under managerial control. The degree of supervision, hierarchy, and delegated authority depends on how costly direct measurement becomes. In this framework, hierarchy itself becomes an economizing response to costly measurement.

This is quintessential UCLA economics. Instead of beginning with abstract categories like “authority” or “organization,” economists begin with the “primitives” -- property rights, information costs, and measurement problems. Institutions emerge from attempts to economize on transaction costs under imperfect measurement.

Sharecropping, contracts, and the real world

Cheung pushed this logic much further than Coase himself had. Indeed, one could argue that Cheung’s real contribution was to make transaction-cost economics operational (even before Oliver Williamson started doing that also (and in a very different way)). Coase’s original insights were brilliant but elusive. Cheung turned them into a concrete research program centered on contracts, measurement, and observable organizational forms.

His earlier work on share tenancy already demonstrated this orientation. At the time, many economists viewed sharecropping as an inefficient feudal remnant. Cheung attacked this interpretation. Sharecropping was not irrational backwardness; it was an efficient contractual response to monitoring costs, risk-sharing, and information problems. This work became foundational for modern contract theory (also in a “negative” way, as Joseph Stiglitz explicitly credited Steven Cheung’s work on sharecropping, with which he strongly disagreed, with motivating his research program in information economics).

Similarly, Cheung’s work on property rights and common resources challenged simplistic “tragedy of the commons” narratives by emphasizing the actual contractual and institutional arrangements governing resource use. Throughout his career, he displayed extraordinary skepticism toward theories built on abstract assumptions detached from real contractual practices.

The paper’s concluding pages contain a revealing intellectual memoir in miniature. Cheung recalls sitting on a wooden box in Hong Kong while two shoeshine boys spontaneously split the polishing of his shoes. This mundane episode triggered his return to Coase and eventually his research program on contracts.That anecdote perfectly captures the older price-theory tradition at its best: economics begins not with equations but with puzzling observations about ordinary life.

When transaction costs entered economics

The paper also provides a fascinating snapshot of the intellectual revolution occurring in economics during the 1960s and 1970s. Cheung describes a period when concepts such as transaction costs, property rights, and contractual arrangements were still intellectually marginal. Today it is difficult to appreciate how radical these ideas once seemed. At the time, mainstream economics was dominated by highly abstract equilibrium models in which institutions often disappeared entirely.

Cheung and his intellectual allies were pushing economics back toward the real world. And they succeeded spectacularly. The intellectual lineage connecting Coase, Alchian, Demsetz, Cheung, and Barzel (as well as Reuben Kessel, Henry Manne, …) transformed large parts of economics, law, organization theory, and political economy. Modern theories of vertical integration, contracting, corporate governance, franchising, relational contracts, incomplete contracts, institutional economics, and organizational design all bear their imprint.

Why Cheung still feels fresh

Reading Cheung today is striking because of how empirically grounded and institutionally concrete his work feels compared to much work in formal organizational economics. Contract economics often became increasingly formalized and abstract. Cheung remained deeply suspicious of theoretical elegance detached from observable contractual structures.

He was also unusually willing to question categories other economists took for granted. One of the most provocative sections of “The Contractual Nature of the Firm” concerns the ambiguity of firm boundaries. Cheung argues that trying to define firm boundaries precisely is often futile because real economies consist of overlapping contractual arrangements rather than sharply bounded organizations. Subcontractors, franchisees, licensors, contractors, suppliers, and platform participants blur any clean distinction between firms and markets.

In many ways, this argument feels even more relevant today than when he wrote it. Modern digital ecosystems, platform economies, gig work, cloud services, and global supply chains increasingly dissolve traditional organizational boundaries. The old textbook image of the vertically integrated corporation coordinating clearly identifiable employees inside rigid boundaries looks increasingly archaic. Cheung saw this conceptual instability decades earlier.

And perhaps this is why his work still feels intellectually fresh. He approached economics less as a system of settled categories than as an ongoing attempt to explain how real people organize exchange under costly information and imperfect measurement.

The economist who dissolved the firm?

At the end of the paper, Cheung makes a striking prediction. He suggests that transaction costs and contracting may someday rival marginalism itself as a foundational basis for economic analysis. In some ways, he was right. Much of modern economics now revolves around incentive problems, contractual structures, information asymmetries, and institutional design.

Cheung’s work definitely helped shift attention toward contracts, transaction costs, and property rights. Later theories of the firm developed these themes in different directions, focusing on ownership, asset specificity, and incomplete contracts as additional determinants of organizational structure.

On the basis of these later lines of research, it may be questioned whether Cheung’s contractual view of the firm fully captures the importance of ownership and legal organization. He emphasized contracts and transaction costs, arguing that the distinction between firms and markets is often overstated. Later theories of the firm, particularly those focused on asset specificity, incomplete contracts, and residual control rights, placed greater emphasis on ownership and firm boundaries. Corporate law likewise treats firms as distinct legal entities rather than merely collections of contracts. While these developments were not usually framed as direct critiques of Cheung, they suggest that contractual arrangements alone may not explain all aspects of economic organization.

Still, reading Cheung today also reveals how much of his deeper institutional realism has been diluted as these ideas became absorbed into mainstream theory. What made the Coase-Alchian-Cheung-Barzel tradition so powerful was not simply its focus on transaction costs. It was its insistence that economics must begin from real institutional arrangements, actual contractual practices, and observable organizational problems.

Cheung never lost sight of that. And that is why “The Contractual Nature of the Firm” remains such a great read. It is not merely an interpretation of Coase. It is a demonstration of a distinctive way of doing economics: concrete rather than formalistic, contractual rather than categorical. And, unfortunately, a way of doing economics that increasingly seems to be becoming extinct.

The piece is interesting per se: it is always interesting to learn what one Nobel laureate thinks of the work of another laureate. But, there is also something quite revealing about Paul Samuelson’s late-life remembrance of Friedrich Hayek. Let me explain.

On the surface, Samuelson is generous. He calls Hayek’s Nobel “worthy.” He credits Hayek with an important contribution to “information economics.” He even concedes that Hayek “was actually the debate’s winner” in the socialist calculation debate.

But beneath the compliments lies a remarkably patronizing essay, one that treats Hayek as a brilliant but fundamentally misguided eccentric whose main ideas were either confused, disproven, or psychologically driven.

Samuelson’s piece reads less like an intellectual engagement with Hayek and more like the victor’s retrospective indulgence toward a defeated rival. To be sure, Hayek is granted originality, even genius in spots, but only after Samuelson carefully confines him to the margins of “serious” economics. But, the striking thing is not merely the condescension. It is how much Samuelson misunderstands what Hayek was actually trying to do.

“Mumbo-Jumbo” and the Keynesian blind spot

The tone is set early and firmly! Samuelson outright dismisses Hayek’s 1931 contribution, Prices and Production as “mumbo-jumbo about the period of production” that “grossly misdiagnosed” the macroeconomics of the Depression. This is classic Samuelson: (very) rhetorically confident, historically triumphant, and conceptually inattentive.

Hayek’s theory was never primarily a short-run theory of aggregate demand management. Samuelson reads Hayek as though he were simply offering a bad Keynesian model before Keynes perfected the genre. But Hayek’s central concern was different: the intertemporal coordination of production under distorted price signals generated by credit expansion. And, as I explained in an earlier post, John Hicks interpreted Hayek’s theory as a theory of growth and how growth could be harmed by monetary policy.

That is not “mumbo-jumbo.” It is a theory about how monetary disturbances alter the structure of relative prices and thereby misalign investment plans across time. It is an important theme in the history of economics, part of what Axel Leijonhufvud called “the Wicksell connection.”

Hayek’s “period of production” language admittedly became cumbersome and overextended. And Prices and Production is a hard read. Even sympathetic readers concede that. But Samuelson caricatures the argument rather than engaging it. The Austrian claim was not that recessions are caused by “too much employment” or that recovery policies are necessarily evil. It was that artificially cheap credit can generate patterns of investment that later prove unsustainable once underlying consumer preferences and resource scarcities reassert themselves.

One may disagree with that theory. But Samuelson’s presentation reduces it to a cartoon. His line that “if Hayek believes that the spending of newly printed currency on employment and consumption will worsen our current terrible depression, then Hayek is a nut” is revealing precisely because it assumes the Keynesian framework as self-evidently exhaustive.

Hayek’s question, however, was never merely whether stimulus can raise employment in the short run. It was whether monetary expansion distorts the informational and coordinating role of prices in ways that create deeper structural maladjustments later. Samuelson never really enters that argument.

Hayek as failed economist, successful philosopher?

Samuelson repeatedly suggests that Hayek effectively failed as an economist and escaped into philosophy after losing to Keynes. Hayek’s economics, he writes, fell into a “black hole.” Hayek’s 1941 book, The Pure Theory of Capital was “a pebble thrown into the pool of economic science that seemingly left nary a ripple.”

Again, the tone matters. Samuelson writes as though economics had decisively settled the questions Hayek raised. However, in actuality much of twentieth-century macroeconomics eventually moved toward concerns Hayek had emphasized: imperfect knowledge, dispersed information, heterogeneous capital, expectations, coordination problems, and the limits of centralized optimization.

Ironically, Samuelson’s own highly formalized equilibrium economics often abstracted away precisely the epistemic problems Hayek insisted were fundamental.

Indeed, Samuelson praises Hayek’s work on information while subtly domesticating it into the mainstream neoclassical framework. Hayek’s “greatest important contribution,” he says, was “information economics.”

But Hayek’s argument was far more radical than Samuelson acknowledges. Hayek was not merely adding “information” as another variable into equilibrium theory. He was challenging the very assumptions that made equilibrium models tractable in the first place. The central planner problem was not computational in the narrow sense. It was epistemic.

Knowledge is dispersed, tacit, context-specific, and often inarticulable. Market prices are not simply outputs of a known optimization problem; they are discovery mechanisms through which individuals coordinate plans they themselves only partially understand.

Samuelson turns Hayek into a precursor of modern information economics while stripping away the deeper challenge Hayek posed to the equilibrium imagination itself.

The Scandinavian straw man

The essay becomes especially weak when discussing The Road to Serfdom. Samuelson treats the book as a prediction that welfare states like Sweden or Denmark would become Nazi-style dictatorships. Since Sweden and Denmark did not become totalitarian, Samuelson concludes Hayek’s fears were refuted.

But this is a totally crude reading of Hayek. Hayek’s argument was not that every welfare state inevitably culminates in concentration camps. His concern was with the political logic of comprehensive economic planning—especially the way coercive powers expand when governments attempt to direct economic life according to unified collective purposes.

Hayek repeatedly distinguished between limited welfare measures compatible with liberalism and centralized planning incompatible with it. His 1960 book, The Constitution of Liberty essentially endorses a number of of welfare state measures. Samuelson collapses those distinctions.

This matters because Samuelson interprets Hayek entirely through the lens of postwar American liberalism: if mixed economies survived without fascism, then Hayek must have been hysterical!

But Hayek’s argument was institutional and constitutional, not prophetic in the crude sense Samuelson imputes to him. Hayek worried about cumulative erosions of spontaneous order, rule-bound governance, and decentralized decision-making. One can debate whether those worries were justified without pretending that Hayek predicted Copenhagen or Stockholm would become Stalinist.

Samuelson prefers the easier target …

The psychological reduction

Perhaps the most revealing passages come near the end, where Samuelson subtly psychologizes Hayek’s intellectual disagreements. Hayek becomes “slightly depressive.” His ideological persistence is framed as irritation in old age. Samuelson even compares him to elderly scholars whose personalities deteriorated because of failing arteries.

The implication is unmistakable: Hayek’s continued disagreements with Samuelson are not really substantive intellectual disputes but manifestations of temperament, aging, or personality. This is deeply patronizing.

Samuelson never seriously entertains the possibility that Hayek’s criticisms of Keynesianism, planning, inflationism, or technocracy might have enduring force. Instead, Hayek is transformed into an interesting but flawed character in the grand Samuelsonian narrative of enlightened centrist progress.

Even Samuelson’s praise has this quality. Hayek is admired the way one admires a gifted dissenter who contributed usefully before history moved on. A smart loser, essentially.

The irony

The deepest irony is that Samuelson’s essay unintentionally vindicates Hayek.

Hayek spent his career warning against intellectual overconfidence—especially among experts who believed they possessed sufficient knowledge to manage complex social orders from above. He titled his Nobel Prize speech “The Pretence of Knowledge.”

Samuelson’s remembrance radiates exactly that confidence. He writes as the representative of a victorious postwar technocratic consensus: Keynesian demand management, welfare-state liberalism, and mathematically formalized economics. Hayek appears as the colorful dissenter who misunderstood where history was going. But history did not end there.

Inflation crises, stagflation, knowledge problems in planning, public choice critiques of government, and renewed attention to institutions and incentives all complicated the Samuelsonian post-war consensus. Meanwhile, Hayek’s work on dispersed knowledge became increasingly central to economics, political theory, and even cognitive science.

Samuelson wanted to preserve Hayek as a partially redeemed eccentric: wrong about macroeconomics, wrong about politics, but useful on information. The problem is that Hayek’s insights cannot be compartmentalized so neatly. His theory of knowledge was inseparable from his skepticism toward centralized economic control and technocratic ambition.

That is exactly what Samuelson never fully grasped.

Subscribe now

In the years after his Nobel prize, Phelps turned his attention to what he called “dynamic capitalism”: not just an economy that grows, but one that generates a continuous stream of novel commercial ideas. His book Mass Flourishing argued that what distinguishes the Anglo-American economies from continental European corporatism is not their high savings rates or even their institutions in a static sense, but their openness to entrepreneurial experimentation and the varieties of entrepreneurial judgment that makes such experimentation viable. This is not the macroeconomics of the graduate textbook but something closer to what Austrian economists have been saying for decades.

What struck me when I read Phelps’s Kauffman Foundation piece in 2005—I blogged about it at the time—was his account of the capital market as a matching process. In his telling, entrepreneurs cannot fully articulate their vision; they operate under what Knight would call genuine uncertainty. Financiers, similarly, cannot objectively rank projects. What they can do is identify entrepreneurs whose “model of the world” seems compatible with their own. The result is not a textbook auction of capital to highest-expected-value projects but a decentralized sorting of worldviews. Phelps explicitly credited this insight to Hayek and Michael Polanyi—the idea that everyone in a market economy carries around some piece of personal knowledge that no central authority can aggregate or replicate.

This is a significant concession from a mainstream macroeconomist, and I don’t think it was fully appreciated at the time. The standard approach to entrepreneurship in economics—when it appears at all—treats the entrepreneur as a residual claimant who responds to price signals. Phelps was saying something different: that the heterogeneity of beliefs among both entrepreneurs and financiers is not a friction to be minimized but the very engine of dynamism. A world in which all financiers agreed on which projects were worth funding would be a world of fast convergence and slow discovery. The disagreement is the point.

This framing connects directly to work Nicolai Foss, Samuele Murtinu, and I have been developing on entrepreneur-investor matching under Knightian uncertainty. The basic problem is this: what makes entrepreneurial projects difficult to finance is not primarily incentive misalignment—the moral hazard and adverse selection that dominate the academic finance literature—but something more fundamental. Entrepreneurs and potential funders often cannot agree on how to classify the relevant future events. They hold different subjective beliefs about what the venture is, what it could become, and whether the underlying resources are the right ones. We call this intersubjective uncertainty, and it is a direct consequence of the Knightian conditions Phelps was describing: when entrepreneurial imagination is genuinely novel, there is no shared category into which it fits, and therefore no easy basis for agreement.

The matching that results is not arbitrary, but discriminating in ways the standard model misses. Projects that are highly novel in their value proposition but relatively transparent in their resource structure call for different financing arrangements—and different persuasion strategies—than projects where the product is familiar but the underlying resource combination is opaque and hard to evaluate. What looks like capital market inefficiency is often the system working as it should: financiers applying their own theories of value, entrepreneurs finding the funders whose worldview is close enough to make a deal, and both parties gradually developing shared cognitive ground through experimentation and communication. Phelps’s insight was that this pluralism of financier judgment is not a bug. Our work tries to show, in some detail, why this is and how the matching actually unfolds.

I’m also working on a paper that pushes deeper into the epistemology of that process. Mises observed that the entrepreneur and the historian are doing essentially the same thing in opposite temporal directions: both are trying to make sense of human action—the motives, valuations, and choices of situated persons—through what Mises called thymology, the interpretive understanding of human conduct. The historian asks what an actor was doing and why; the entrepreneur asks what consumers, competitors, and partners will do and why. The methods are similar. This matters for the matching problem because successful enrollment of a financier requires the entrepreneur to exercise exactly this kind of prospective interpretation, to inhabit the financier’s model of the world well enough to communicate across the gap in beliefs. And it suggests that what separates better from worse entrepreneurial judgment is not superior data or a more rigorous hypothesis-testing protocol, but something closer to what a skilled historian possesses: source discipline, motivated reconstruction, and the imaginative capacity to think from within another’s worldview.

Phelps worried, in his later work, that this dynamism was eroding—that corporatist drift, risk aversion, and a cultural retreat from the spirit of enterprise were slowing the generativity of Western economies. Whether or not he is right on the cultural causes, the underlying point seems right: an economy that loses the pluralism of its financiers, or that centralizes the selection of entrepreneurial experiments in the hands of government or consensus-seeking industry bodies, will run fewer and less diverse experiments. It will get less discovery, not more efficiency. The continental model he contrasted with Anglo-American capitalism, where experts convene to set product standards before any version reaches the market, is not just slower but weaker.

Economists who know Hayek’s work will find most of this familiar. What made Phelps worth reading was that he arrived at these conclusions from within the mainstream, using its tools and taking its constraints seriously, and that he was willing to say in the Wall Street Journal and at Nobel lectures that capitalism’s great virtue is not equilibrium but productive disequilibrium. That’s a harder sell than it sounds from inside a tradition that built its prestige on models of market clearing.

Richardson was quite a remarkable figure. He worked in military intelligence during WWII, was (Brian Loasby told me) the youngest person ever to become a Fellow at Oxford U; produced a string of profound contributions to econ well into the 1970s; held a number of government advisory roles; then checked out of research in frustration that his work had, he thought, little impact; still did a stint as pro-vice-chancellor at Oxford U; became the CEO of Oxford University Press; a college warden, Chairman or Director (I forgot) of the Ashmolean museum and much else. The first time I met George, he said: “You know, Nicolai, I had lunch with your queen yesterday” (Queen Margrethe had been visiting the Ashmolean Museum). In spite of his accomplishments and position in Britain, he was a very humble person.

Anyway, here is why Richardson is relevant and interesting in the context of entrepreneurship and AI.

Economies work because of “imperfections”

Modern economics often presents competition as a process that works best when markets are frictionless, information is perfect, and participants behave like interchangeable agents. In that world, imperfections are distortions: asymmetries of information, coordination failures, uncertainty, and uneven capabilities are viewed as problems to be minimized. Yet Richardson offered a profoundly different perspective. He argued that the competitive system functions not in spite of imperfections, but because of them.

Richardson’s insight is especially relevant in the age of artificial intelligence. Today, entrepreneurs increasingly rely on the same large language models, the same data sources, the same software infrastructure, and even the same startup playbooks. A growing concern is that AI will homogenize entrepreneurship (e.g., “convergence effects”). If every founder uses ChatGPT, Claude, Gemini, or similar systems to analyze markets and generate ideas, will entrepreneurship become a race toward identical projects or as some like to say, “opportunities”?

Richardson’s theory suggests the opposite. The more powerful and widespread AI becomes, the more valuable the remaining imperfections become. The future of entrepreneurship may depend precisely on the irreducible differences between people, organizations, networks, timing, judgment, and coordination.

Richardson’s critique of perfect competition

Richardson, especially in his 1960 book, Information and Investment and later writings such as “The Organisation of Industry” (1972), criticized the overly abstract nature of neoclassical competition theory. Richardson argued that the textbook conception of perfect competition (as well as more foundational GE models) overlooked the fact that firms possess different capabilities and must coordinate complementary activities under conditions of uncertainty.

Richardson believed this view ignored how actual economies function. Real economic systems are characterized by uncertainty, incomplete knowledge, diverse capabilities, institutional relationships, and evolving expectations. Firms do not merely react to prices; they interpret uncertain futures, build relationships, coordinate activities, and develop specialized competencies.

For Richardson, competition is productive precisely because knowledge is dispersed and imperfect. If everyone possessed the same information, the same expectations, and the same capabilities, there would be little room for entrepreneurial discovery or strategic differentiation. Economic dynamism emerges because actors perceive the world differently. Imperfections are therefore not accidental blemishes on capitalism. They are the conditions that make experimentation, innovation, and entrepreneurship possible.

This also explains why Richardson was unusually sympathetic, especially by modern neoclassical standards, to certain “restraints of competition.” In “The Organisation of Industry,” he argued that many forms of vertical and horizontal coordination that economists and antitrust authorities often regarded as anti-competitive could in fact improve economic performance by facilitating the coordination of complementary activities. Long-term supplier relationships, collaborative agreements, industry standards, and various forms of inter-firm cooperation were not necessarily distortions of competition. They were often mechanisms through which firms managed uncertainty, synchronized investments, shared knowledge, and stabilized expectations.

This was an unfashionable position because much postwar economics tended to interpret restraints of competition primarily as departures from the ideal of perfect competition. Richardson instead saw many such arrangements as functional responses to the realities of coordination under imperfect knowledge (work in the new institutional economics would interpret these as transaction cost minimizing arrangements a couple of decades later).

A key theme in his work is that the competitive model (at least the version without perfect contingent forward markets) does not tell us how coordination happens. This is a problem in the long run. Because of assumptions of atomism, firms cannot coordinate complementary or competitive investments. In effect, he argues, the long-run industry supply curve isn’t defined under competitive (Marshallian) conditions. Real-world “constraints on competition,” both vertical and horizontal constraints exist partly to assist the coordination complementary and competitive investments, respectively. In fact, Richardson is one of very few economists who have been at least favorable-in-principle to even horizontal price agreements.

Entrepreneurship, opportunities, and judgment

Richardson’s diagnoses of a fundamental issue of coordination at the heart of the competitive model relates directly to entrepreneurship. Seen from the view of the still-dominant opportunity discovery perspective in eship research, entrepreneurship manifests because opportunities are not obvious. If profit opportunities were fully visible and universally recognized, they would not simply disappear in a mechanical sense. The Richardsonian point is that under full information the supposed opportunity would no longer constitute an exploitable opening. Once all relevant actors fully understand and act upon the same opportunity, attempts to exploit it become self-defeating. Entrepreneurship matters because markets are imperfectly understood and because economic action depends on uneven knowledge, divergent expectations, and imperfect coordination.

But entrepreneurship is not only about discovery. It is also about judgment under uncertainty. As the judgment-based view associated with Frank Knight and Ludwig von Mises and later developed by Peter Klein and me (and an increasing number of younger economists and management scholars), entrepreneurs must commit resources in situations where outcomes cannot be known probabilistically in advance. The entrepreneur is therefore not merely someone who notices an opportunity, but someone willing to exercise judgment about an uncertain future and bear the consequences of that judgment.

This perspective complements Richardson in important ways. Richardson stressed the coordination of complementary capabilities under uncertainty. Entrepreneurs do not simply identify price gaps; they organize activities, align expectations, and create workable patterns of cooperation in environments characterized by incomplete knowledge.

From a judgment-based perspective, this coordinating role is inseparable from entrepreneurial decision-making. Entrepreneurs must decide which capabilities to assemble, which markets to enter, which technologies to back, and which relationships to cultivate before outcomes are known. These are not decisions that can be mechanically derived from available data. They involve interpretation, imagination, and commitment.

This becomes especially important in industries where activities are highly interdependent. Launching a startup requires more than identifying customer demand. It involves assembling teams, securing financing, developing products, forming partnerships, navigating regulations, building trust, and coordinating knowledge that no single individual fully possesses. The entrepreneur must exercise judgment not only about the product itself, but about the organization of complementary activities and the sequencing of investments across time.

In Richardson’s framework, firms emerge because markets alone cannot coordinate all these complementary activities efficiently. Organizations exist partly to handle the imperfections that pure market exchange cannot solve. From the judgment-based perspective, firms also emerge because entrepreneurs need structures through which they can exercise and implement judgment under uncertainty. This insight becomes relevant when we consider AI.

AI Does Not Eliminate Imperfections

A common concern about generative AI is that it will standardize entrepreneurial thinking. Founders increasingly use AI tools for market research, competitive analysis, business model ideation, coding, customer support, product design, financial modeling, marketing copy, and investor pitches. While this would seem to make sense, there is at least the possibility that entrepreneurs may converge on the same ideas. AI models are trained on similar public information and optimized to produce statistically plausible responses. Some fear that entrepreneurship becomes derivative, as discovery appears “commoditized” and entrepreneurial ideas therefore becomes “convergent.”

However, this misunderstands entrepreneurship. While AI expands access to information, it does not eliminate uncertainty and it doesn’t eliminate the need for entrepreneurial judgment.

Information is not the same thing as judgment. AI can summarize existing patterns, identify trends, and generate hypotheses. But, entrepreneurship is about acting under conditions where data is incomplete, futures are unknowable, and — the Richardsonian angle — coordination is difficult.

Even if two entrepreneurs discover the same apparent opportunity using AI, they will still differ in interpretation, timing, risk tolerance, execution ability, networks, organizational culture, access to capital, strategic patience, credibility, team composition, and much else. These differences are not marginal imperfections. As Richardson would argue, they are the substance of competitive processes. Indeed, along Richardsonian lines, one might argue that as AI lowers the cost of information processing, competitive advantage may shift even more strongly toward areas where imperfection remains irreducible.

The New Scarcity: Coordination and Judgment

But, the increased use of AI in the entrepreneurial process surely matters. When analytical tools become widely accessible, the bottleneck moves elsewhere. Historically, information itself may have been scarce. Today, judgment and coordination are becoming scarcer (this also why universities will survive the AI revolution; they offer high-level judgment and connections).

Suppose two entrepreneurs use AI to identify an opportunity in elder-care automation. Both receive similar analyses showing aging demographics, labor shortages, and unmet demand. One founder builds a technically sophisticated but socially tone-deaf product. Another founder spends months building relationships with care providers, understanding emotional resistance among staff, and designing workflows that complement rather than replace caregivers. The AI-generated insight was identical, but the entrepreneurial outcome was not.

Richardson’s framework helps explain why. Economic activity depends on coordinating complementary capabilities under uncertainty. This process cannot be fully automated because it involves tacit knowledge, trust, institutional understanding, and evolving social relationships.

AI can help identify possibilities, but it cannot fully determine which possibilities are feasible, legitimate, or scalable within specific institutional contexts. Here, judgment is required.

AI may increase competition -- and diversity

Paradoxically, AI may intensify the very imperfections Richardson considered essential. By lowering barriers to entry, AI allows more people to experiment entrepreneurially. Individuals who previously lacked technical skills can now prototype software, conduct research, and automate operations.

This democratization increases the number of market participants. And as the number of participants rises, diversity of interpretation may also increase. The reason is that more people use the same tools within different contexts.

A founder in Copenhagen, Nairobi, São Paulo, and Bangalore may all access the same language model, yet they will apply it differently based on local constraints, social networks, institutional environments, and lived experience. The result is not necessarily convergence; it could instead be an explosion of variation.

Richardson’s theory suggests that competition thrives when actors combine shared market mechanisms with differentiated capabilities and expectations. AI may provide common infrastructure while leaving entrepreneurial imagination deeply decentralized.

In this sense, AI could resemble earlier general-purpose technologies such as electricity or the internet. The internet did not eliminate entrepreneurial difference simply because everyone gained access to the same protocols. Instead, it enabled an enormous expansion of experimentation. AI is very likely to do the same.

The importance of organizational capabilities

Drawing on Penrose’s thinking, Richardson emphasized that firms differ not only because they possess different information, but because they develop distinct capabilities. An organization learns through experience, routines, culture, and specialization. These features cannot be instantly copied.

This basic insight, key to resource-based and capabilities theories, matters enormously in AI-driven entrepreneurship. Many startup ideas may become easier to imitate at the conceptual level. But execution capabilities may become more decisive.

In a world where everyone can generate business ideas with AI, investors and customers may care less about the originality of the concept and more about the organizational ability to execute. The premium shifts from ideation alone toward speed of adaptation, quality of coordination, organizational learning, customer integration, trustworthiness, and long-term resilience

Richardson anticipated this logic decades ago. Competition is not simply rivalry between isolated firms responding to prices. It is a process involving coordination among specialized organizations with different competencies. AI does not erase this structure, but, if anything, it may deepen it.

Why identical opportunities still produce different outcomes

Suppose AI identifies the same opportunity for 500 entrepreneurs. Does that eliminate entrepreneurial profit? As Peter Klein and I have argued, if you must use the notion of “opportunities,” one thing they are not are objective entities waiting passively to be discovered. Rather, they are enacted through organizational processes.

An opportunity only becomes economically meaningful when someone successfully mobilizes resources around it. This is why the judgment-based approach stresses the combination of heterogenous capital in the pursuit of profit as an integral part of the entrepreneurial process. The implication, furthermore, is that “the same opportunity” is often not truly the same.

Two founders entering the same market may pursue radically different interpretations. One may target premium customers while another pursues mass adoption. One may prioritize speed while another emphasizes reliability. One may build vertically integrated infrastructure while another relies on partnerships. One may optimize for acquisition while another aims for long-term independence.

As noted earlier, AI may converge idea generation but divergence reappears through strategy, execution, and organizational evolution. Richardson’s central point survives intact: competition works because economic actors are not identical.

His fundamental critique of the received Coasian theory of the firm comes to mind. So does his 1982 paper on entry barriers. (Stay tuned for future posts on both Demsetz’ specialization theory of the firm and his view of entry barriers). Or his work on competitive bidding for exclusive franchises (“Demsetz competition”). Or … There are so many other examples. Here is a wide-ranging interview with Demsetz that gives you an idea of the scope of his interests (and lots of anecdotes, to boot).

While I have long considered Demsetz to be the GOAT (and thought it a scandal that he never received the Nobel Prize), I only him once, namely at the founding 1997 conference in St. Louis of the International Society for New Institutional Economics. He was sharp as a razor and surprisingly verbally aggressive for a man of his age and professional stature (he was very kind in private conversation). But, his many interventions during the conference were absolutely to the point and as far as I can remember completely justified (only Ronald Coase impressed me more back then).

Demsetz’ 1973 paper

Perhaps the famous example of Demsetzian “fundamental” thinking is his critique of the structure–conduct–performance paradigm and his emphasis on efficiency over market power. Demsetz pushed economists to ask whether observed outcomes could reflect underlying differences in firm capability rather than monopolistic distortions. His parallel work on property rights, information, and firm heterogeneity helped lay intellectual groundwork that strategy scholars would later build on. Demsetz not only profoundly shifted the conversation toward a more nuanced, evidence-driven view of competition, he may also be the true originator of the resource-based view of strategy (RBV), which is still strategy’s dominant perspective.

Demsetz’s 1973 paper, “Industry Structure, Market Rivalry, and Public Policy,” is conventionally read as a critique of the structure–conduct–performance (SCP) paradigm. That’s correct and SCP was indeed Demsetz’ explicilt target. But, there is more to the paper: the first two pages of this paper contain, in remarkably compressed form, the core logic of what would later become the RBV.

In a post a week ago I covered Lippman and Rumelt’s equally seminal work. If Lippman and Rumelt provide the formal machinery for RBV analysis, Demsetz provides the conceptual spark. He reframes the entire question of competition by shifting attention away from market structure and toward firm-level differences in efficiency. In doing so, he effectively plants the seed for RBV, years before the term existed.

The target: The SCP paradigm

At the time Demsetz was writing, industrial organization was dominated by the SCP framework. (When I studied econ, our industrial organization course (1985) was still 100% based on the SCP framework). The logic was straightforward: concentrated industries were presumed to exhibit less competition, which led to higher profits. The empirical task, therefore, was to identify monopoly power by looking at concentration ratios and profit rates

Demsetz opens by attacking this orientation head-on. He notes that the field has been overly focused on “searching for monopoly” largely because of policy concerns and the availability of concentration data . This is not just a critique of method; rather, it is a critique of how economists think about competition.

His alternative is deceptively simple: what if concentration is not a cause of profitability, but a result of it?

The core move: “Concentration through competition”

The key conceptual move comes immediately in Section II, titled “Concentration Through Competition.” This is where Demsetz essentially writes the first two pages of the RBV.

He argues that, under competitive conditions, concentration can arise because some firms are simply better:

“An industry will become more concentrated… only if a differential advantage in expanding output develops in some firms.”

This is the pivot.

Instead of asking whether concentration enables firms to earn profits, Demsetz asks whether profits and concentration both reflect underlying differences in firm capability. Firms grow not because they collude, but because they are more efficient, whether through lower costs, better products, or superior organization.

In one stroke, he flips the causal arrow that had dominated industrial organization.

The essence of RBV in two pages

What’s striking is how much of RBV is already present in this early argument.

Demsetz explicitly centers firm heterogeneity. Firms are not identical units operating in a market—they differ in their ability to produce and compete. These differences are not incidental; they are the driving force behind industry structure.

He also emphasizes that these advantages can take many forms, including cost reductions, better products, or organizational efficiencies. This anticipates the RBV idea that competitive advantage can stem from a wide range of firm-specific resources and capabilities.

Most importantly, he highlights the difficulty of imitation. On the second page of the article, he makes an argument that reads like a direct precursor to causal ambiguity:

It may be “very difficult for these firms to understand the reasons for this difference in performance or to know to which inputs to attribute the performance of the successful firm.”

This is pure RBV logic: even when success is observable, its causes may not be. Competitors cannot easily replicate superior performance because they cannot isolate the relevant factors. The firm is a complex, integrated system—a team—whose productivity cannot be disassembled and copied piece by piece. Here is the second page of the article where these ideas are launched:

The firm as a bundle of non-transferable capabilities

Demsetz goes even further by suggesting that superior performance may be tied to the firm as a whole, not to individual inputs: Superior performance may be “unique to the firm, viewed as a team, and unobtainable to others except by purchasing the firm itself.” This is a powerful statement (suggesting a link to his work with Armen Alchian published the year before)

It anticipates several core RBV ideas at once. Capabilities are firm-specific and not easily tradable in factor markets. They are embedded in organizational relationships, routines, and histories. In some cases, they can only be transferred through acquisition of the entire firm.

He also points to sources of advantage that later RBV work would emphasize, such as reputation and goodwill, which are difficult to separate from the firm, and team-specific knowledge arising from interactions among employees. These insights foreshadow the later emphasis on tacit knowledge and social complexity.

Profits without monopoly

Another crucial insight is Demsetz’s reinterpretation of profits. He argues that high returns need not reflect monopoly or collusion:

“Profit does not arise because the firm creates ‘artificial scarcity’… Superior performance can be attributed to the combination of great uncertainty plus luck or atypical insight.”

This is a direct rejection of the SCP interpretation of profits.

In Demsetz’s framework, profits are rewards for superior performance under uncertainty. Firms experiment, make decisions, and sometimes get it right. When they do, they earn returns not because they restrict output, but because they deliver more value or lower cost.

This is exactly the logic that RBV later formalizes: rents arise from valuable and difficult-to-replicate capabilities.

Why imitation doesn’t eliminate advantage

Demsetz also addresses a key question: if firms are more efficient, why don’t competitors quickly catch up?

His answer anticipates later RBV mechanisms. Information is costly, outcomes are uncertain, and the mapping from inputs to performance is unclear. As a result, even if competitors try to imitate, the process is slow and imperfect. By the time the value of certain inputs or strategies is recognized, “they are beyond acquisition by other firms at the same historic cost.”.

This introduces a temporal dimension to advantage. Firms that discover valuable opportunities first can enjoy superior returns for a period, even in competitive markets.

Demsetz vs. Penrose: Competition vs. growth

It is useful to contrast Demsetz with Edith Penrose, who also articulated ideas that resemble RBV. In her 1959 book, The Theory of the Growth of the Firm, Penrose emphasized the internal resources of the firm, how different “services” may be derived from resources over time as a result of management learning, and how learning and resource services shape growth, particularly through diversification.

But Penrose was not primarily concerned with competition. Her focus was on how firms grow through related diversification, not on how they outperform rivals in equilibrium.

Demsetz, by contrast, embeds heterogeneity directly into a competitive setting. His argument is not about how firms grow, but about how competition itself generates differences in size, profitability, and structure. That is what makes his contribution foundational for strategy. He connects firm-level differences to market outcomes and competitive dynamics.

The policy implication: A warning

Demsetz’s argument is not just theoretical; it has sharp policy implications. If concentration reflects efficiency rather than monopoly, then policies aimed at breaking up large firms may be misguided.

He warns that deconcentration policies risk reducing efficiency by penalizing successful firms and shifting output to less efficient producers. The empirical analysis later in the papershows that larger firms often earn higher returns, without evidence that smaller firms benefit from concentration. This suggests that observed profitability differences may reflect efficiency rather than collusion.

The deeper point is that misdiagnosing the source of profits leads to bad policy. If profits are due to superior capabilities, not collusion, then attacking concentration may undermine the very forces that drive innovation and efficiency.

Conclusion: The first clear statement of RBV logic

What makes Demsetz (1973) a cornerstone of the RBV is the conceptual leap it makes. He reframes competition as a process in which firms differ in their capabilities, those differences drive both profits and market structure, and imitation is limited by information costs and complexity.

There is no formal model, no explicit theory of resources, and no later RBV terminology. But the logic is unmistakable. In just a few pages, Demsetz replaces a structural view of competition with a resource-based one.

If Penrose planted early seeds of resource-based thinking within the theory of the firm, Demsetz is the one who brings those ideas into the analysis of competition.

The first two pages of “Industry Structure, Market Rivalry, and Public Policy” contain the essence of the RBV. Firms are heterogeneous, some are more efficient than others, their advantages are difficult to understand and imitate, and these differences shape both profits and industry structure.

Everything that follows in the RBV literature can be read as elaboration, formalization, or extension of this core insight. That is why Demsetz (1973) deserves to be seen not just as a critique of industrial organization, but as one of the true foundational texts of modern strategy.

I will make the case that it is more than a worthy and influential contribution. Lippman and Rumelt (1982) is a “killer paper,” meaning a paper that has decisively reshaped how we (certainly those of us who are strategy scholars) think about competition, firm performance, and the origins of advantage.

What makes the paper so powerful is not just the idea it introduces, but the way it challenges a deeply entrenched assumption in economics: that competition and free entry should eliminate differences in firm performance.

By challenging that assumption, the paper lays down key intellectual foundations for what would about a decade later become the resource-based view (RBV) of strategy (there is another, earlier precursor, namely Demsetz [1973] which I will discuss in a future post). In fact, Margaret Peteraf’s 1993 canonical statement of the RBV is essentially Lippman and Rumelt in words and with diagrams instead of equations.

The Puzzle: Why Don’t Profits Get Competed Away?

To understand the significance of the paper, it helps to start with the puzzle it addresses. In standard economic theory, persistent differences in profitability across firms are difficult to justify. If some firms are earning above-normal returns, competitors should imitate them or enter the market, driving profits down. If that doesn’t happen, the typical explanation points to some form of market imperfection, such as barriers to entry, monopoly power, or regulatory protection.

But this story has always sat uneasily with empirical reality. Even in competitive industries with no obvious barriers, firms display striking and persistent differences in performance. Some firms are simply better, and they stay that way.

Lippman and Rumelt take this empirical regularity seriously and ask whether it can be reconciled with competitive equilibrium rather than explained away as an anomaly. Their answer is yes, but only if we rethink a critical assumption, namely that firms can understand and replicate the sources of superior performance.

The Core Insight: Causal Ambiguity

The central concept they introduce—uncertain imitability—captures the idea that even when firms observe that others are more successful, they may not be able to figure out why. The issue is not just incomplete information, but something deeper: causal ambiguity. As they put it, the factors responsible for superior performance may resist “precise identification,” meaning that even with observation and effort, competitors cannot unambiguously determine what to copy.

This is the conceptual pivot that makes the paper so important. In standard models, imitation is assumed to be feasible once information is available. But if the causal links between actions and outcomes are fundamentally ambiguous, then imitation becomes unreliable. Firms can try to replicate success, but they cannot be sure which elements matter or how they interact. As a result, imitation is not just costly—it is uncertain in a deep, irreducible way.

From Intuition to Formal Theory

What elevates the paper from an interesting idea to a “killer” contribution is that Lippman and Rumelt formalize this intuition in a rigorous model. They do not simply assert that ambiguity matters; they build it into the structure of firm entry and competition.

In their framework, entering a market is equivalent to “producing” a production function. Each entrant draws a cost structure from a known probability distribution, but cannot know in advance which draw it will receive. Some firms end up with efficient production processes, others with inferior ones, and this heterogeneity arises endogenously from the entry process itself.

Crucially, the same uncertainty applies to imitation. A firm attempting to replicate a successful competitor faces the same probabilistic process as a new entrant. It cannot simply copy the superior production function, because the underlying causal structure is not fully understood. The result is a world in which firms are rational, markets are competitive, and entry is free—and yet persistent differences in efficiency and profitability emerge as equilibrium outcomes.

The Modeling Achievement Itself

It’s worth pausing here, because part of what makes this paper “killer” is not just what it argues, but how it argues it. The modeling effort is doing heavy conceptual lifting.

First, Lippman and Rumelt take something that had largely been treated as a vague, almost philosophical idea—i.e., “you can’t always copy what you don’t understand”—and translate it into a tractable economic object. They operationalize causal ambiguity as ex ante uncertainty over production functions. That move seems simple, but it’s incredibly powerful. It allows them to stay within a fully rational, equilibrium framework while still capturing something that looks, at first glance, like bounded rationality or organizational complexity.

Second, they insist on preserving the discipline of neoclassical modeling. Firms are profit-maximizing. Entry decisions are rational. Markets clear. There is no appeal to behavioral quirks or ad hoc frictions. In fact, they explicitly contrast their approach with earlier work that relied on bounded rationality or simulations. Their goal is to show that persistent heterogeneity and rents can emerge even when you give the standard model every advantage.

This is just …. cool. It is what gives the result its bite. If rents survive under these stripped-down, fully rational conditions, then you can’t dismiss them as artifacts of irrationality or market failure.

Third, the paper builds the argument progressively across multiple model structures, including atomistic firms, fixed-capacity firms, and scale-based uncertainty. This layering matters. The simplest case shows the core intuition: even with price-taking firms, free entry does not eliminate profit dispersion. More complex cases then show that the result is robust to different assumptions about firm size, industry structure, and dynamics. The modeling is not just illustrative; it is deliberately designed to demonstrate generality.

Fourth, the authors are unusually careful about equilibrium logic. Entry stops not when profits are zero, but when the expected value of entry becomes negative. That subtle shift—i.e., from realized profits to expected returns under uncertainty—is doing crucial work. It explains how observed profits can remain positive even though entry is rationally deterred. The equilibrium concept itself is quietly redefined.

Finally, there is a kind of intellectual elegance to the whole construction. Rather than adding frictions, constraints, or institutional details, the model hinges on a single, disciplined modification: uncertainty in the creation of capabilities. From that one change, a cascade of non-obvious results follows: persistent heterogeneity, rents without market power, and entry deterrence without barriers.

In that sense, the modeling effort is not just a vehicle for the idea. It is the contribution. It shows that you can get RBV-style conclusions without abandoning economic rigor.

The Radical Result: Rents Without Market Power

One of the most striking implications of the model is that positive profits can persist even in perfectly competitive settings. The authors show that, under uncertain imitability, industries can exhibit “stable interfirm differences in profitability” along with “an above-normal industry rate of return” despite the absence of market power .

This is a direct challenge to the standard link between rents and monopoly. In their framework, rents do not arise from restricting competition, but from the inherent difficulty of reproducing successful capabilities.

Even more counterintuitive is the relationship between profitability and entry. In conventional theory, high profits attract entrants. But in this model, high profits can signal the presence of firms with exceptionally efficient—and difficult to imitate—production processes. Observing such success does not necessarily encourage entry; it may deter it. Potential entrants recognize that the probability of replicating that success is low, and the cost of failure is real. As a result, entry may cease even when profits remain above normal levels.

Barriers to Entry Become Barriers to Understanding

This inversion of the usual logic leads to one of the paper’s deepest insights. What looks like an “entry barrier” may not be a structural feature of the market at all, but a consequence of uncertainty and ambiguity. The barrier is not physical or institutional—it is cognitive.

Firms cannot enter successfully because they cannot understand what they would need to do.

This idea maps directly onto what would later become the core of the resource-based view. RBV emphasizes that firms possess resources and capabilities that are valuable, rare, and difficult to imitate. But early RBV work often treated inimitability as a given. Lippman and Rumelt provide a mechanism. They show how inimitability can arise endogenously from the nature of knowledge and the process of capability creation. If the link between actions and outcomes is ambiguous, then even observable success does not translate into replicable knowledge.

Competition as Selection, Not Convergence

The paper also anticipates another central RBV theme: the role of selection. In their model, the industry evolves through a process that resembles natural selection. Firms enter, draw capabilities, and either survive or fail depending on their efficiency. Over time, resources flow toward the more efficient firms, while less efficient ones exit.

But unlike in standard models, this process does not lead to convergence. Because new entrants continue to face uncertainty, the distribution of firm performance remains dispersed. The industry settles into an equilibrium characterized by persistent heterogeneity rather than uniformity.

This emphasis on selection rather than convergence is critical. It shifts the focus of analysis from equilibrium conditions to the processes that generate and sustain differences. Firms are not simply optimizing given known production functions; they are experimenting, discovering, and sometimes stumbling into superior ways of operating.

Factor Markets and the Limits of Knowledge

Another important contribution of the paper is its treatment of factor markets. In standard theory, efficiency differences should be eroded as factors of production move to their highest-valued uses. But Lippman and Rumelt argue that factor mobility itself depends on the ability to identify which factors are valuable.

If the sources of success are ambiguous, then factors cannot be correctly priced or reallocated. As they note, “factors of production cannot become mobile unless they are known.”

This insight connects causal ambiguity directly to the persistence of advantage, reinforcing the idea that knowledge limitations, and not just market structure, shape competitive outcomes.

Laying the Foundations for the Resource-Based View

The broader implication is that competition is not merely a process of price adjustment and entry, but, echoing a key Austrian theme, a process of discovery under uncertainty. Firms differ not only in what they have, but in what they understand. Some firms possess capabilities that are not only valuable but also difficult to articulate and transfer. These capabilities generate rents not because competitors are excluded, but because they cannot reliably reproduce them.

Seen in this light, the paper marks a turning point in the intellectual development of strategy. It challenges the dominance of the structure-conduct-performance paradigm, which locates the sources of profitability in industry structure. Instead, it suggests that structure and performance may both be outcomes of deeper processes related to uncertainty and capability development.

The observed relationships between concentration, market share, and profitability need not reflect market power; they may instead reflect the distribution of underlying capabilities.

The Deeper Shift: Strategy in a World of Unknowns

This shift has far-reaching consequences. It opens the door to a view of strategy centered on firm-specific resources rather than industry positioning. It provides a theoretical basis for sustained competitive advantage in the absence of barriers to entry. And it reframes the role of managers, not as optimizers within a known environment, but as actors navigating uncertainty and shaping the capabilities of their organizations.

At its core, the paper makes the point that the limits of knowledge—not just the structure of markets—shape competitive outcomes. Firms are not just optimizing known production functions; they are grappling with systems they do not fully understand. Thus, the equilibrium logic does not preclude links to Schumpeter, Knight and Kirzner.

Conclusion: A Quiet Revolution

What makes Uncertain Imitability a “killer paper,” then, is not just that it introduces a new concept, but that it reorients the entire conversation. It shows that persistent performance differences are not anomalies to be explained away, but natural outcomes of an, if you like “Austrian,” world in which knowledge is imperfect and causality is ambiguous.

It demonstrates that rents can arise without market power, that competition can coexist with heterogeneity, and that the limits of imitation are central to understanding strategy.

Perhaps most importantly, it changes the question we ask. Instead of wondering why profits do not get competed away, we should ask why they cannot be. The answer, in Lippman and Rumelt’s framework, lies in the fundamental uncertainty surrounding the creation and replication of capabilities.

That insight—simple, elegant, and deeply consequential—is one of the reasons the paper is a true cornerstone of modern strategy theory.

Artificial intelligence is very often framed as a general-purpose technology on par with electricity or the internet. But its most immediate and profound impact may be organizational rather than purely technological. As I have argued in another post on this substack, AI is a productivity shock—but at least as much it is an organization design shock.

Specifically, AI is poised to reshape the internal structure of firms—their hierarchies, decision rights, and coordination mechanisms—in ways that challenge long-standing assumptions in organizational economics and management theory. The pic below provides a rough illustration of the argument.

The Classical View: Hierarchies as Solutions to Transaction Costs

Ronald Coase’s foundational insight in his 1937 paper was that firms exist because markets are costly to use and firms may have a cost advantage relative to markets for some transactions. Oliver Williamson extended this logic, emphasizing asset specificity and opportunism as drivers of internal organization (but didn’t say much about internal organization).

Within firms, hierarchies themselves are solutions to coordination and control problems. Herbert Simon, in Administrative Behavior (1947), described organizations as systems of boundedly rational agents, where authority substitutes for complete contracting. Roy Radner later formalized the problem of information processing in organizations, showing how communication constraints shape optimal structures. Across these traditions, hierarchies economize on scarce cognition and costly communication

Luis Garicano’s 2000 paper, “Hierarchies and the Organization of Knowledge in Production,” provides a particularly sharp microfoundation for this view. In his model, organizations are systems for solving problems under constraints of limited knowledge and costly communication. Workers at lower levels handle routine problems; more complex or exceptional cases are escalated upward to agents with greater expertise. Hierarchies arise as an efficient solution to allocating problem-solving capacity when it is too costly for everyone to know everything.

Garicano’s key comparative statics hinge on two costs: the cost of acquiring knowledge and the cost of communicating problems. When communication is expensive, firms rely on specialization and multiple layers of escalation. When communication becomes cheaper, hierarchies can flatten because problems can be routed more efficiently to those best able to solve them.

AI can be understood as a shock to precisely these parameters.

AI as On-Demand Expertise

From a Garicano-style perspective, modern AI systems function as widely accessible, on-demand expertise embedded throughout the organization. Instead of escalating problems up a chain of command, employees can query systems that approximate the knowledge of more senior managers. The traditional ladder of escalation is partially replaced by horizontal access to codified intelligence.

At a more general level, AI dramatically lowers the cost of producing, processing, and transmitting information. Large language models, predictive systems, and decision-support tools compress the gap between data and action. Tasks that once required specialized intermediaries can now be performed instantly and at scale.

This can be read as a reduction in internal transaction costs in the Coasian sense (I here abstract from the new transaction costs that AI may introduce). Garicano’s framework makes the mechanism more precise. The cost of acquiring relevant knowledge falls because it can be accessed through AI. The cost of communicating problems falls because they can be formalized and processed by machines. As both margins shift, the optimal span of control expands and the need for layered escalation diminishes.

Radner’s insight that organizational form reflects limits on information processing also becomes newly salient. AI effectively augments those limits, allowing organizations to process more signals with fewer human intermediaries. Simon’s bounded rationality is not eliminated, but it is partially externalized into technical systems.

When Information Stops Being Scarce

Yet the Garicano model also suggests why flattening is not the whole story. If communication becomes cheap and knowledge widely accessible, the bottleneck shifts. What becomes scarce is not information but attention and prioritization.

Here, the logic begins to intersect with Aghion and Tirole’s distinction between formal and real authority . Even if decision rights are formally delegated, real authority depends on who effectively shapes decisions. AI expands the set of feasible actions and available insights, but someone must still decide which problems matter.

As a result, senior leadership may become more central in setting direction, even as operational layers thin out. Strategy shifts from gathering information to selecting among proliferating options. In Garicano’s terms, the problem is no longer routing questions to experts, but determining which questions are worth asking.

At the same time, AI systems themselves introduce new coordination problems. Decisions about model design, data governance, and acceptable risk create a layer of meta-organization. These functions are likely to be centralized, both for efficiency and for control, echoing Williamson’s emphasis on governance structures that mitigate hazards.

The emerging pattern is therefore bifurcated: operational hierarchies flatten as knowledge becomes more accessible, while strategic and governance layers may thicken around the control of AI systems.

Who Decides When Everyone Knows?

Garicano’s framework implies that decision rights should follow knowledge (a point made earlier by Hayek, Jensen & Meckling and others). When knowledge is concentrated at the top, authority is centralized; when knowledge is dispersed, authority can be delegated. AI complicates this mapping by simultaneously democratizing knowledge and enhancing monitoring.

On the one hand, employees across the organization gain access to tools that approximate expert reasoning. This compresses the knowledge gradient that justified hierarchical escalation. On the other hand, AI enables more precise and continuous monitoring of decisions and outcomes, reducing agency costs.

This dual effect aligns closely with Aghion and Tirole’s insight that formal delegation does not necessarily imply real autonomy. Firms can delegate decisions while retaining the ability to intervene or override, creating a more fluid and contingent allocation of authority.

But this perspective remains incomplete without incorporating the notion of judgment emphasized by Peter Klein and me.

In our view, the essence of firm organization is not merely the allocation of information or decision rights, but the exercise of judgment under uncertainty—decisions that cannot be fully specified ex ante or reduced to algorithmic rules.

AI, for all its strengths, primarily operates in the domain of patterned inference rather than genuine uncertainty-bearing judgment. It can extend the reach of knowledge and improve prediction, but it does not bear responsibility in the Knightian sense.

This creates a sharp distinction between decisions that can be delegated to machines or lower levels and those that require committed judgment by individuals or leadership.

From this angle, AI does not eliminate hierarchy but clarifies its core function. If Garicano explains how hierarchies economize on knowledge and communication, Peter and I seek to explain why some decisions must remain centralized: they require ownership-backed judgment that cannot be decomposed or fully delegated.

The result is a reconfiguration of authority. Routine and even moderately complex decisions may move downward or be automated, while high-uncertainty, high-consequence decisions become more explicitly concentrated. Hierarchy becomes less about knowing more and more about deciding when it matters most.

The Disappearing Middle—or Its Reinvention?

Garicano’s model is particularly illuminating when thinking about middle management. In a knowledge hierarchy, middle managers exist partly because they possess intermediate levels of expertise and partly because they economize on communication between the top and the bottom.

AI erodes both justifications. If knowledge can be accessed directly and communication costs fall, the need for intermediate layers as conduits diminishes. This is the mechanism behind the often-predicted hollowing out of middle management.

However, neither Simon nor Garicano would, I surmise, predict their disappearance. Both emphasize that not all knowledge can be codified and not all problems can be anticipated. Middle managers may therefore shift toward roles that involve contextual judgment, integration across domains, and the handling of non-routine exceptions, as also implied by my work on the managerial hierarchy with Peter Klein.

In this sense, their function becomes less about relaying information and more about interpreting it. They act as buffers between the formal logic of algorithmic systems and the messy reality of organizational life.

What Stays Inside the Firm?

The implications extend beyond internal hierarchy to the boundary of the firm. Coase and Williamson frame this boundary as a trade-off between internal and market transaction costs. AI shifts this trade-off by making it easier to specify, monitor, and integrate tasks performed outside the firm.

Garicano’s framework complements this by suggesting that when knowledge can be codified and transmitted cheaply, it becomes easier to separate problem-solving from organizational membership. Tasks that were once internal because they required close supervision may be externalized if AI can replicate the necessary knowledge and monitoring.

At the same time, the economics of AI itself—particularly scale effects in data and model training—may encourage internalization of key capabilities. Firms may become more modular at the periphery while becoming more integrated at the core, especially around proprietary data and models.

The New Politics of Information

None of these changes are purely technical. As the behavioral theory of the firm argues, organizations are coalitions with conflicting interests. Garicano’s model is largely efficiency-driven, but in practice, changes in knowledge distribution alter power structures.

AI redistributes informational advantages. Individuals who can effectively leverage AI gain influence, while those whose roles are tied to now-automated tasks may lose it. At the same time, control over AI systems—data access, model design, and deployment—becomes a new source of authority.

This creates potential tensions between technical units and traditional business functions, and between centralization for control and decentralization for responsiveness. The formal structure of the firm may change more slowly than the underlying distribution of real authority.

From Hierarchies to Systems

Taken together, these perspectives suggest that AI is not simply making organizations more efficient along existing dimensions. There is a sense in which it is shifting the underlying problem that organizations are designed to solve.

Firms are becoming less about allocating scarce expert knowledge through hierarchical escalation and more about orchestrating abundant, machine-augmented knowledge across contexts. In Simon’s terms, bounded rationality is being partially reconfigured rather than eliminated. In Aghion and Tirole’s terms, the relationship between formal and real authority is becoming more fluid.

While the resulting organizations may appear flatter in their formal charts (but perhaps more complex in their actual functioning), hierarchy does not disappear. It is still there to set direction, governing systems, and resolving ambiguity in environments where information is no longer the primary constraint.

Redesigning Hierarchy Itself

In other words, AI does not eliminate the need for hierarchy; it changes its rationale.

Garicano’s theory suggests that as the costs of knowledge acquisition and communication fall, the structure of hierarchies must adjust. The early evidence from AI adoption is consistent with this prediction, but it also reveals new constraints—attention, governance, and politics—that were less central in earlier models.

Judgment, stressed in my work with Peter Klein, seems to me to add the crucial final layer. As knowledge becomes cheap and widely accessible, judgment becomes the scarcest resource in the organization. The purpose of hierarchy shifts accordingly: not to concentrate information, but to concentrate responsibility for decisions that cannot be reduced to rules.

The firms that navigate this transition successfully will be those that understand these theories not as abstract frameworks, but as design tools. They will recognize that AI is not just a productivity technology, but a reconfiguration of the informational and decisional foundations of the firm.

The question is no longer how to manage within a given hierarchy, but how to redesign hierarchy itself in a world where knowledge is cheap, communication is instantaneous, and judgment remains irreducibly human.

It started when the activist and journalist Chris Rufo attacked my colleague Elizabeth Corey who, during a podcast interview, expressed some reservations about Rufo’s methods while acknowledging his accomplishments. I defended Elizabeth, which led to a torrent of responses by Rufo and his admirers, whose larger point seemed to be that conservative and libertarian professors are weak, ineffective, lazy, and cowardly types who prefer to play within the rules at the expense of “winning.” Rufo is described by his admirers (and by himself) as the most effective higher-education reformer of our time, maybe of all time, and anything less than adoration is motivated by jealousy or self-protection. (As a friend of mine quipped, paraphrasing Ibram X. Kendi, “It’s not enough to be pro-Rufo; you must be actively anti-anti Rufo.”)

Elizabeth subsequently outlined her position in more detail in a WSJ article (which did not satisfy Rufo and his team). Others including Kevin Vallier entered the fray. I mostly stayed out, while jumping in with a bon mot here and there. But I now that the furor has died down, I’d like to share a few more substantive thoughts about what is going on and what we might do to improve things. (My discussion is focused on the U.S. but the general ideas should apply to other settings as well.)

The conformity problem

Start with the most striking feature of higher education: its extraordinary resistance to change. Despite superficial variation, most universities look remarkably alike. Four-year degrees. The same menu of majors and minors. Lectures, discussions, papers, exams. Professors with PhDs. The differences between an R1 research university, a liberal arts college, and a community college are real but surprisingly modest, especially given how much the surrounding economy has transformed over the same period. There’s been no Henry Ford or Steve Jobs of higher education—no one who fundamentally reimagined what the product is or how it should be delivered. (I once thought it might be Apple, which suggested a paraphrase of the famous quip about Tolstoy: “Steve Jobs is the Steve Jobs of higher education.”)

Education, like health care, isn’t a homogeneous blob but a discrete set of goods and services that can be bundled or unbundled, priced creatively, and offered in various shapes and sizes. Yet we treat it as if it’s categorically unique and immune to the normal forces of competition and innovation (perhaps because it generates positive externalities?). Hollis Robbins has a good recent piece arguing that institutional sameness—enforced by the accreditation cartel and state funding requirements—is the real enemy, one that has reduced pedagogy to box-checking rather than genuine teaching. Professors across disciplines are increasingly alienated from their own profession as a result.

The structural and organizational explanation

Part of the explanation for this conformity is selection. Hayek’s 1949 essay on the intellectuals described the tendency of academia to attract people who prefer designing society to acting in the marketplace. This generates a steady, structural pressure toward anti-market attitudes—not as a conspiracy but as an emergent property of who self-selects into the profession. Schumpeter added that is the “absence of direct responsibility for practical affairs” that distinguishes the academic intellectual from others “who wield the power of the spoken and the written word.” The critical attitude of the intellectual arises, says Schumpeter, “no less from the intellectual’s situation as an onlooker—in most cases also as an outsider—than from the fact that his main chance of asserting himself lies in his actual or potential nuisance value.” Combine selection bias with insulation from accountability, and you have a recipe for ideological drift.

The modern university has also undergone a massive expansion in scale, funded largely by the GI Bill, Title IV loans, NIH and NSF grants, and the like. This growth has produced the familiar pathologies of large, weakly governed organizations: bloat, inefficiency, diffuse accountability, and severe information asymmetries. Crucially, no one holds residual control or cash flow rights in the way that shareholders or owner-managers do in a firm. Nicolai Foss and I have argued that corporate wokeness is driven by middle managers—agents who face few internal constraints and can use their discretionary authority to pursue ideological agendas. The same dynamic applies in universities, where DEI bureaucracies and administrative bloat are rarely driven by faculty, students, or senior leadership but by mid-level administrators in HR and student affairs who have found a niche and are filling it. This is not mainly a story about the wrong people with the wrong ideas being in power. It’s a story about incentive structures and the absence of the constraints that, in other organizational contexts, keep such tendencies in check.

A new paper by David Cutler and Ed Glaeser argues that the long-term survival of universities reflects structural advantages of combining faculty autonomy with external control over capital—a model that supports adaptation within the existing form. Scott Masten has related work exploring why universities, like legislatures, are not organized as firms. Both are worth reading. Neither, however, explains the ideological drift of the 20th and 21st centuries—which suggests that organizational resilience and ideological capture are partly independent phenomena. The university form has survived for good functional reasons; what’s been loaded into it lately is a separate question, with a separate answer.

What is to be done

This brings me to what I’d call my middle position on the Great Awokening. There is certainly an ideological element. But the underlying problems are long-term and structural, not primarily the result of bad actors who need to be purged and replaced. The activist model, whether it comes from politicians threatening funding or campus organizers demanding loyalty oaths, tends to leave the underlying structure intact while fighting over who controls it. Battles over speech on state campuses are ultimately fights over state property—political contests like any other, not principled free-speech battles. Winning those fights doesn’t fix the institution; it just changes the faction running it.

The remedy isn’t to capture the existing apparatus but to change the incentive structure that makes it worth capturing. That means promoting private ownership, competition, and entrepreneurship in education. In this sense, I applaud genuinely different models like the University of Austin and the Hamilton School (though their effectiveness remains to be seen), purely private schools like Grove City College, vocational and for-profit institutions, organizations like the Mises Institute, online platforms, and AI-assisted learning. Even chatbots are pretty good at teaching certain subjects! Let a thousand flowers bloom.

We want education entrepreneurs to experiment and discover what works, not a one-size-fits-all solution imposed from above—whether that imposition comes from accreditors, state legislatures, or the federal government. We’re already seeing that top-down reforms can be more symbolic than substantive, something the activists have not recognized or grappled with. We need more bottom-up experimentation, driven by a variety of actors.

What prevents this experimentation is state funding, regulation, and the accreditation cartel. Make institutions sit on their own bottoms, have them compete on multiple margins, and let faculty and students sort into the models they prefer. That’s the reform agenda worth pursuing.

That framing matters. Because if AI is a continuation rather than a rupture, then the question is not whether it will reshape society—it will—but whether the forces shaping it are compatible with human flourishing.

Klowden and Tao are cautiously skeptical. They emphasize competitive pressures, warning of a “Faustian bargain” between efficiency and ethical risk, and call for a more human-centered approach to AI development.

But there is another way to read the same dynamics: not as a drift away from human-centered outcomes, but as an uneven, market-driven process that may already be moving—imperfectly, but materially—in that direction.

AI as a phase change

There’s something quietly persuasive about the central claim in Klowden and Tao’s paper: that artificial intelligence is not an alien rupture, a discontinuous “intelligence explosion,” a “hard takeoff,” etc., but a continuation of humanity’s long tradition of building tools to extend thought.

AI, they argue, is less a threat than a phase change—like writing, printing, or computation before it. The question then is how it can be integrated in a way that remains “fundamentally human-centered.”

That framing is appealing not just philosophically, but practically. It resists both techno-utopianism and apocalyptic pessimism (Roko’s Basilisk and all that).

But when you look at the same problem through the lens of economics (and management theory), the picture becomes richer—not because Klowden and Tao are wrong, but because the forces shaping AI development may be more aligned with their goals than is often assumed.

A prisoners’ dilemma? Or beneficial experimentation?

The tension runs through the entire paper, sometimes explicitly, sometimes just beneath the surface. It appears most clearly in their description of AI adoption as a kind of “Faustian bargain,” driven by competitive pressure. Firms and individuals adopt quickly not because they’ve resolved the ethical questions, but because they can’t afford not to. This is framed as a cautionary tale—specifically, a prisoners’ dilemma pushing society toward premature deployment.

But from an economic perspective, that same dynamic is doing something else as well: it is accelerating experimentation.

As Nobel laureate Friedrich von Hayek argued, markets are “discovery procedures” that enable ongoing experimentation, revealing knowledge that we never realized was available, as well as incentivizing the production of new knowledge.

Competitive markets are extraordinarily efficient discovery mechanisms. They force rapid iteration, expose weak ideas, and reward systems that actually deliver value to users. The “frenzied pace” Klowden and Tao worry about is also the mechanism by which AI capabilities have improved so quickly, and, crucially, spread beyond a narrow research elite.

The real issue, then, is not speed per se. It’s whether the feedback loops embedded in these markets are aligned with broader human benefit or narrowly with private gain—that is, the extent of externality. And here the paper gestures at an answer, calling for human-centered outcomes. But it may understate the extent to which markets themselves can help generate those outcomes.

Markets and human-centered outcomes

Consider their discussion of the Industrial Revolution. The analogy is familiar: automation displaces workers in the short term but increases overall prosperity in the long run. Klowden and Tao rightly emphasize that these transitions are not painless, invoking the Luddites not as reactionaries but as rational actors facing genuine economic threat.

What’s striking, though, is how well this historical pattern fits the current moment—not just in its costs, but in its adaptive capacity. Over time, markets have repeatedly reallocated labor, created new industries, and lowered costs for consumers. The same forces that displace jobs also create demand for new skills, new services, and new forms of organization.

That doesn’t mean the transition is automatically just or smooth. But it does suggest that market systems are not inherently at odds with human-centered outcomes—they are incomplete without complementary institutions.

The question is less “markets versus humanity” and more “which institutions and policies produce the best distribution of gains?”

Innovation, concentration, and diffusion

This becomes particularly important when the paper turns to the concentration of power in AI. The comparison to the transcontinental railroad and the Gilded Age is apt: large fixed costs, massive upfront investment, and the emergence of powerful firms controlling critical infrastructure.

It’s easy to read this as a story of inevitable monopolization. But management theory on innovation offers a more nuanced view. While scale advantages are real, they are often counterbalanced by forces that fragment markets over time: technological diffusion, open-source ecosystems, and shifting consumer preferences.

There are early signs of this in AI, such as the rise of smaller, specialized models, the growing importance of fine-tuning, and the ability of firms to build differentiated products on top of shared foundational technologies.

Klowden and Tao hint at this possibility when discussing “local models.” But “local models” are not just altruistic projects; they can also be competitive strategies. Decentralization, in many cases, is not imposed from outside—it emerges from the incentives of firms and users who prefer flexibility, cost efficiency, and control over their data.

Judgment as the binding constraint

One of the most interesting discussions in the paper is of mathematics as a “sandbox” for AI. The idea that AI could flood a field with technically correct but low-value results is not just a philosophical concern; it’s a classic case of supply overwhelming attention.

In economic terms, we are moving from a world where knowledge production is scarce to one where it is abundant. That shift doesn’t eliminate value; it relocates it.

When production becomes cheap, curation becomes expensive. When answers are easy to generate, judgment becomes the scarce resource.

This is not necessarily a decline. It’s a reallocation. Markets are already adapting to this shift by elevating new forms of signaling—reputation, trust, editorial standards, and brand. The risk, as Klowden and Tao note, is that low-quality output could drown out meaningful work. But the countervailing force is that users, institutions, and platforms have strong incentives to filter aggressively. Attention is too valuable not to.

A similar dynamic appears in their discussion of environmental and resource costs. The paper highlights the significant energy and water demands of AI infrastructure and the risk that these costs are not adequately accounted for.

This is indeed a genuine concern. But markets are not blind to it. Energy is a cost. Efficiency improvements—whether in hardware, algorithms, or data center design—are strongly incentivized.

Firms that can deliver the same performance with lower resource consumption gain a direct competitive advantage. The history of computing is, in large part, a story of relentless efficiency gains driven by exactly such pressures.

That doesn’t eliminate a role for policy—especially around environmental externalities—but it does suggest that market forces can be part of the solution rather than purely the problem.

Is AI like vanilla extract?

Perhaps the most delicate part of the paper is its treatment of the human–AI interface. The “vanilla extract” metaphor is charming and insightful: AI as a small enhancement to human workflows, useful in moderation but overwhelming if overused.

In the short term, this feels right. But over time, organizations tend to scale what works. If AI consistently improves productivity, it will move from the margins to the core. This is not a failure of judgment; it is the logic of optimization. Firms that use AI more effectively will outcompete those that don’t.

Yet this doesn’t necessarily lead to the dystopian outcome Klowden and Tao worry about. Management theory suggests that as tools become more powerful, the nature of work changes rather than disappears. Tasks are re-bundled. New roles emerge. Human effort shifts toward areas where judgment, creativity, and coordination remain valuable.

The paper’s proposal to use AI primarily in a “red team” capacity, that is, verification rather than creation, is a useful heuristic. But this is likely transitional. Over time, the boundary between creation and verification will blur. The more interesting question is how human cognition and machine cognition co-evolve as complementary systems.

A human-centered AI?

This brings us back to the paper’s central aspiration: a human-centered AI. Klowden and Tao frame this largely as a matter of values—what we choose to prioritize, what we consider beneficial, how we weigh costs and benefits.

That’s important. But economics suggests that values alone are not enough. What matters just as much is how those values are embedded in incentives.

The encouraging part is that markets are not static. They can be shaped. Competition can be steered. Incentives can be aligned—through pricing, regulation, standards, and institutional design.

The same forces that currently drive rapid, uneven adoption can, under the right conditions, also drive broader access, greater efficiency, and more widely shared benefits.

In that sense, the paper may be more right than it realizes.

A negotiated future

If AI is a phase change, then its trajectory is not preordained. It will be shaped—messily, unevenly—by the interaction of innovation, competition, and policy.

That process will produce frictions. It will also produce adaptation.

The future of AI will not be dictated from a single lab, company, or regulator. It will be negotiated—through markets, through institutions, and through the choices of the people who build and use these systems.

That is not a guarantee of good outcomes.

But it is a system that learns.

And that, more than anything, is why there are stronger grounds for optimism than the current discourse often allows.

The latest entry in this genre is a widely circulated piece co-authored by Jack Dorsey and Sequoia’s Roelof Botha, arguing that Block is replacing hierarchical management with an AI “intelligence layer”—and that this represents a fundamental break with two thousand years of organizational design. It’s a well-written piece and the historical sweep—Roman legions to Prussian general staffs to American railroads to the modern matrix organization—is genuinely illuminating. The ambition is admirable. But the core claim is less radical than it looks, and the problems it skips over are exactly the ones organizational scholars have been working on for decades. Old wine, exceptionally well-bottled.

A story as old as the Roman army

The piece tells a clean story: all organizational history has been built around a single constraint—a leader can manage only three to eight people directly, so you add layers, and layers slow everything down. AI breaks that constraint. Instead of humans routing information through management hierarchies, a continuously updated “world model” does the coordination work. Middle management isn’t reformed; it’s replaced.

There’s something to this. The information aggregation function of hierarchy—knowing what’s happening across the organization, relaying context up and down the chain—is in many ways being automated. Anyone who has watched a capable AI system synthesize project statuses across a hundred Slack threads and a thousand GitHub commits knows this isn’t science fiction. As Nicolai and I have written, middle management plays essential buffering and brokerage functions, some of which are being algorithmized at breakneck speed.

But the piece conflates two things that organization theory keeps separate: coordination and judgment. Hierarchy exists not just to route and shape information but to make resource allocation decisions under genuine uncertainty—the kind of uncertainty where the right answer isn’t in the data, because the data don’t exist. That’s the Knightian problem at the heart of Why Managers Matter. A world model that knows everything that has happened doesn’t tell you what to do next when you’re facing a genuinely novel situation. The intelligence layer “composes” capabilities in response to patterns it has seen before. What happens when the pattern is new? What happens when the situation calls not for the best available composition of existing capabilities, but for a judgment that those capabilities are the wrong ones entirely?

Meet the new manager, same as the old manager

Dorsey and Botha don’t just overlook this issue, they paper over it deliberately (with a cool label). The “Directly Responsible Individual” (DRI) holds full authority to pull resources across teams and own outcomes for 90-day sprints. The “player-coach” develops people and makes judgment calls the model shouldn’t make alone. These are not algorithmic outputs. They are managers with different job titles. The hierarchy hasn’t been eliminated; it’s been clarified. The layers that existed mainly to pass information up and down, without discretion, get stripped out—a genuine improvement—while the layers that exist to exercise judgment remain. That’s not a revolution in organizational design. It’s a version of what every serious restructuring effort tries to accomplish, including many of the ones that failed.

The piece’s own historical narrative makes this point almost explicitly, though without drawing the conclusion. The reason Spotify reverted, Zappos attrited, and Valve struggled to scale isn’t that these companies lacked good software. It’s that they misdiagnosed the problem. The issue was never that hierarchy is inherently bad. The issue is that poorly designed hierarchies mix functions that should be kept distinct, such as information routing (which technology can handle) and judgment under uncertainty (which requires human authority and legitimacy). The companies that flattened indiscriminately eliminated both at once. Block, to its credit, is trying to eliminate only the first. But that means it is not actually building a post-hierarchical organization. It is building a leaner, more deliberate hierarchy—which is a worthwhile goal, and perhaps genuinely novel in its execution, but not the paradigm shift the framing implies.

What AI actually changes—and what it doesn’t

My own view, developed at length in Why Managers Matter and in related work with Nicolai, is that this is roughly the right prediction for what AI will do to organizations: automate the codifiable, concentrate the judgment. What AI won’t do is eliminate the need for human authority at the points where decisions are genuinely consequential. The Block piece is more honest about this than most flat-hierarchy advocacy—it explicitly carves out ethical decisions, novel situations, and high-stakes moments for humans. But those carve-outs are large. They are, in fact, exactly what managers are for.

Block may well be building something genuinely interesting. The idea of capabilities as primitives, composed dynamically by an intelligence layer rather than assembled into fixed products by product managers, could turn out to be a real organizational innovation. The customer world model, built from both sides of millions of transactions, sounds like a VRIN resource that compounds in ways most firms’ data cannot. These are worth watching closely, and the piece deserves engagement rather than dismissal.

But the framing—“from hierarchy to intelligence”—overstates the novelty. The Romans needed a decanus not just because they lacked Slack. They needed one because someone had to decide, under fire, what to do next. Two thousand years later, Block still needs someone to do that. They’re calling that person a DRI.

The wine is the same. The bottle is quite nice.

If strategy is ultimately about managerial judgment under uncertainty, what role should structure and formal processes really play?

These questions were prompted by my reading of a new paper in Management Science by Mu-Jeung Yang, Michael Christensen, Nicholas Bloom, Raffaella Sadun and Jan Rivkin.

These are some of my favorite authors. They have been doing consistently impressive empirical work quantifying management practices, showing how structured processes and managerial quality shape productivity, performance, and differences across firms. The paper was handled by Alfonso Gambardella at Management Science, which is another reassuring signal of quality.

The paper

The authors examine how CEOs actually approach strategy in practice and whether differences in approach matter for firm performance. Using survey data from 262 CEOs across countries and industries, they construct what they call a Strategy Process Index. This index captures how structured, participatory, and disciplined a firm’s strategy making process is.

The central finding is straightforward but powerful. CEOs differ widely in how they make strategy. Some follow highly structured processes, with regular planning meetings, formal documentation, clear communication of priorities, and systematic tracking of execution. Others rely on more informal and ad hoc approaches, where strategy is less explicitly defined, more centralized in the CEO, and less consistently monitored.

These differences are not small. They are large, persistent, and observable even among otherwise similar firms.

More importantly, the study shows that these differences matter. Firms with more structured strategy processes tend to perform better. They exhibit higher productivity, faster growth, and greater profitability. These relationships hold even after accounting for firm size, industry, and country effects.

The authors also show that these differences are partly driven by the CEOs themselves. Leaders with formal business training, such as MBAs, or prior experience in well managed firms are more likely to implement structured strategy processes. This reinforces the idea that managerial background and individual leadership style shape firm outcomes in systematic ways.

Taken together, the paper reframes strategy as an ongoing organizational process rather than a one time decision. Effective CEOs do not just set direction. They build systems for diagnosing problems, formulating choices, communicating priorities, and monitoring execution.

Overall, the key takeaway is that disciplined and well managed strategy processes are an important, and often underappreciated, source of competitive advantage.

Of course, the paper is not without limitations. The findings are largely correlational, so causality remains an open question. It is possible that higher performing firms are simply more likely to adopt structured processes, rather than those processes directly driving performance. The Strategy Process Index itself focuses on formalization, communication, and participation, but it does not directly capture the quality of strategic decisions. That leaves open the possibility that the measure blends effective management with effective strategy.

There are also concerns related to the data. The reliance on CEO self reports may introduce bias, for example a tendency to present practices in a more favorable light. In addition, the emphasis on structured approaches may not fully capture contexts where flexibility and improvisation are valuable, such as startups or highly uncertain environments. Still, these are fairly standard limitations, and the paper provides a strong empirical foundation while leaving room for future work on context, judgment, and strategic content.

Relations to mainstream strategy research

One of the most interesting aspects of the paper is how it connects to earlier management research. It shifts attention away from what strategies firms choose and toward how they develop and execute them. In teaching, it is often difficult to find good material on strategy processes, so there is a tendency to emphasize content and treat process and implementation as secondary. This paper helps correct that imbalance.

It also complements classic strategy frameworks, particularly the dominant content oriented approaches such as Porter’s positioning framework and the resource based view. Yang and coauthors show that even well conceived strategies may underperform if they are not developed and implemented through disciplined processes. That insight is directly useful in the classroom.

At the same time, the paper engages with Mintzberg’s long standing critique of formal planning. Mintzberg emphasized emergence and warned against rigid planning systems. The evidence here points in a slightly different direction. More structured processes are associated with better outcomes, but this should not be read as a blanket endorsement of bureaucracy. A more balanced interpretation is that effective firms combine structure with adaptability. In that sense, the paper helps bridge the divide between deliberate and emergent strategy.

The findings also resonate strongly with Upper Echelons Theory. The idea that organizational outcomes reflect the characteristics of top executives receives direct support here. CEO background, such as education and prior experience, shapes how strategy is made. What is especially valuable is that the paper moves beyond simple demographic correlations and identifies concrete behavioral channels through which CEOs influence performance.

There is, of course, also a clear link to the management practices literature, especially the work by Bloom and Van Reenen. That literature has shown how practices like monitoring and target setting relate to productivity. By focusing on strategy making, this paper extends that line of research into a core domain of executive activity. It shows that strategy itself can be understood as a measurable management practice.

Finally, the paper connects to behavioral strategy. Rather than focusing narrowly on cognitive biases, it highlights how organizational processes can structure and discipline strategic thinking. In doing so, it reduces reliance on intuition alone and embeds decision making in a broader system.

Connection to Foss and Klein

The paper also fits nicely with the argument Peter Klein and I have made in Why Managers Matter (and elsewhere). Our perspective, grounded in organizational economics and Austrian economics, emphasizes that managers are essential because they exercise judgment under uncertainty. Firms exist because not all future contingencies can be specified or priced in markets, so managers must make forward looking decisions about resource allocation. Strategy, in this view, is fundamentally about entrepreneurial judgment.

At first glance, Yang and coauthors seem to emphasize something different, namely structured processes. But the two perspectives are better seen as complements rather than substitutes. Their results show that judgment does not operate in a vacuum. Structured processes such as regular meetings, systematic use of data, clear communication, and follow up mechanisms help organize and support decision making.

These processes improve the information available to decision makers, allow for more distributed input, and help ensure consistent execution. In that sense, they do not replace judgment. They discipline it and make it more effective.

The paper’s finding that CEO characteristics shape strategy processes also aligns closely with the idea of heterogeneous managerial capabilities which is central to my work with Peter. Both perspectives (reject the notion that managers are interchangeable. Instead, they are a key source of firm level differences in performance.

A useful way to think about the relationship is that our work explains why managers matter, because judgment is unavoidable under uncertainty, while Yang and coauthors explain how managers matter in practice, through the design and use of strategy making processes.

At the same time, from our perspective the paper arguably underplays the role of radical uncertainty and judgment. This matters because in highly uncertain environments, structure may not always help and may even slow down adaptation. Relatedly, the Strategy Process Index is clever, but it captures formalization and communication rather than the quality of strategic insight (i.e., the quality of judgment). A firm can score highly on process and still pursue a poor strategy.

From a broader microfoundational perspective, the paper focuses on process architecture but says less about the underlying mechanisms. It does not directly address how disagreements are resolved, how power shapes decisions, or how politics and incentives influence strategy making inside firms.

These are, however, relatively minor reservations. The paper is well executed, conceptually interesting, and empirically rich. It is definitely worth your time.

That framing is understandable, but it is also misleading.

AI is not primarily a substitute for scholars. Instead, it changes the way knowledge is produced, coordinated, and developed inside research organizations.

It is a productivity shock, to be sure, but it is also an organization design shock. The basic framework is depicted below.

To understand why, it helps to combine a transaction cost economics (TCE) perspective with a knowledge‑based view of the firm (KBV).

TCE generally highlights the costs of coordinating activities across markets, hybrids, and hierarchies. The KBV adds that economic organization (here, internal structure) is shaped by knowledge, or more precisely the costs of retrieving, articulating, building, transferring, codifying, and contracting over knowledge transactions. The KBV has typically highlighted knowledge that is costly to articulate and transfer, i.e., so-called “tacit knowledge.”

Seen through this combined lens, AI is not simply a productivity tool. It changes which parts of research can be codified and automated, and which parts still depend on human expertise embedded in scholars and research communities.

Research as a Bundle of Knowledge Activities

A research project is not a single task. It is a bundle of knowledge activities, such as searching and synthesizing literature, framing theoretical arguments, collecting, cleaning, and structuring data, coding interviews or text, running statistical models, drafting manuscripts, submitting manuscripts, responding to reviewers, etc. etc.

Each of these activities involves the creation, interpretation, and integration of knowledge. But they differ greatly in one crucial dimension: how codifiable the underlying knowledge is.

Some tasks rely primarily on codified rules and established procedures. Others depend on tacit judgment, domain expertise, and contextual understanding.

This distinction lies at the heart of the knowledge‑based view of organizations.

Organizations—such as universities—exist partly because some knowledge cannot easily be transmitted through markets or formal contracts. Instead, it must be embedded in communities of experts who share norms, training, and interpretive frameworks.

AI changes the boundary between these two types of activities.

AI and the Expansion of Codifiable Knowledge Work

Many components of research rely on relatively standardized procedures, such as formatting references, summarizing articles, generating literature overviews, writing boilerplate code for statistical models, and cleaning structured datasets

These activities involve knowledge that is largely codified and rule‑based.

Historically, universities organized such tasks through hierarchical arrangements: faculty supervised PhD students, research assistants, and postdocs who performed much of this work.

AI systems now perform many of these tasks quickly and cheaply. (I am still amazed, though, at how bad ChatGPT is at handling references …).

In effect, AI dramatically lowers the cost of accessing standardized cognitive capabilities. Researchers can draw on these capabilities on demand rather than coordinating them through internal labor.

From a transaction cost perspective, AI reduces the costs of accessing certain types of knowledge services—search, drafting, and routine analysis. Activities that once required internal coordination can now be executed through tools embedded in the research workflow.

But this mainly applies to tasks whose underlying knowledge can be codified at low cost.

Tacit Knowledge and the Limits of AI

Many core elements of research rely on knowledge that is difficult to articulate or formalize.

Think of recognizing which research questions are genuinely interesting; diagnosing subtle methodological problems; interpreting ambiguous qualitative evidence; designing credible empirical identification strategies; and integrating disparate theoretical perspectives

These activities depend heavily on experiential, tacit knowledge: skills, intuitions, and interpretive frameworks developed through training and experience.

Tacit knowledge is difficult to encode in explicit rules. It is learned through practice, mentorship, and participation in research communities.

This is where the knowledge‑based view becomes particularly important. Organizations are not merely coordination devices—they are repositories and integrators of distributed expertise.

Universities bring together scholars with specialized knowledge who collectively generate new insights by combining their expertise.

AI systems can assist with codified tasks, but they do not possess the deeply contextual understanding that underlies frontier research.

A Simple Framework: Codifiability and Expertise

Enter Friedrich von Hayek. In his 1945 paper, “The Use of Knowledge in Society,” Hayek introduces a key distinction between

• Scientific (or explicit) knowledge—which is general, abstract, and systematic; can be formulated as rules, theories, or data; and is easily communicated and centralized (e.g., in books, statistics, or reports)

  and

• Local (or tacit, particular) knowledge—which is context-specific and dispersed among individuals; often tacit (costly to articulate and transfer), and based on time, place, and circumstance.

We can use Hayek’s distinction to clarify the organizational implications of AI in universities. Distinguish between two dimensions:

1. Codifiability of knowledge (low to high)

2. Degree of domain expertise required (general to highly specialized)

Tasks that rely on highly codifiable knowledge and relatively general expertise—such as summarizing articles or generating standard code—are the most susceptible to AI substitution.

Tasks that require low codifiability and deep domain expertise—such as developing new theoretical frameworks or interpreting complex empirical findings—remain predominantly human activities.

Between these extremes lies a large intermediate zone where AI can assist researchers but not replace their judgment.

The effect of AI is therefore not to eliminate research work, but to shift the composition of work toward activities that rely more heavily on tacit expertise and creative integration.

Verification and the Governance of AI Outputs

AI introduces another organizational challenge: reliability.

AI systems can produce outputs that appear fluent and plausible but contain errors or fabricated citations. As a result, researchers must invest effort in verifying and validating AI-generated material.

From a transaction cost perspective, AI reduces some coordination costs but at least at the current stage of AI development also involves monitoring costs: Researchers must check sources, confirm results, and ensure methodological correctness.

In other words, AI changes not only production costs but also the governance problems associated with knowledge production.

Human Capital Formation and Tacit Skill Development

In the university, there is also a longer-term skill-related dynamic that is easy to overlook, but which I think is crucially important.

Academic training historically involved performing many difficult tasks manually: reading extensively, coding data, building models from scratch, and drafting early versions of papers.

These activities served not only to produce output but also to develop tacit expertise.

If AI substitutes for some of these learning-intensive activities, junior researchers may produce output more quickly but develop less deep expertise.

From a knowledge-based perspective, this matters because universities rely heavily on the accumulation of tacit human capital. The ability to generate new knowledge depends on scholars who possess deep, experience-based understanding of their fields.

If training processes change, the long-run knowledge base of the institution may change as well.

The Changing Organization of Research Work

Taken together, these dynamics suggest that AI will reshape the internal organization of research.

Routine, codifiable components of research will increasingly be automated or tool-based. That is, of course, happening right now. Researchers will spend less time on mechanical tasks and more time on activities that require interpretation, judgment, and synthesis.

At the same time, the value of deep expertise may increase. Scholars who possess strong domain knowledge can use AI as a complement, leveraging it to explore ideas, draft materials, and test alternative approaches.

Researchers with weaker domain expertise may rely more heavily on AI-generated structure, potentially increasing output quantity without necessarily increasing conceptual depth.

This dynamic resembles skill-biased technological change: technology amplifies the productivity of individuals who already possess valuable expertise. An AI-induced Matthew Effect.

Evaluation and Escalating Research Standards

AI is likely to also affect the broader research ecosystem.

If routine components of research become easier to produce, the total volume of submissions to journals may increase. Lots of LinkedIn posts (etc.) have talked about how the publishing system will be overloaded. It will probably will.

However, it is more overlooked that that evaluation standards (in a broad sense) are likely to shift.

Elements that were once costly—comprehensive literature reviews, extensive robustness checks, polished drafts—may become baseline expectations rather than differentiators. The bar will be raised.

Papers may (at least temporarily) be less reliable signals on capability. Capability will increasingly be probed through seminar presentations.

Distinctive contribution may increasingly depend on a high level of originality, conceptual integration, and theoretical insight. Merely “competent” research will not serve to sufficiently differentiate scholars. Thus, judgment and its tacit knowledge components will become increasingly important.

AI as a Fundamental Knowledge Infrastructure

The bottomline arguably is that AI is best understood not as a substitute for scholars but as a new knowledge infrastructure.

It expands access to codified knowledge and automates certain cognitive routines. At the same time, it increases the relative importance of expertise that cannot easily be codified.

Universities therefore face a strategic challenge. They must decide how to integrate AI tools into research workflows while preserving the processes that generate deep expertise (while also dealing with the “distributional” consequences of AI among faculty).

The future of research will depend not only on what AI can do, but on how institutions cultivate and govern the human knowledge that remains difficult to replicate.

In that sense, AI is less (perhaps much less) a replacement for scholars than a catalyst for the reorganization of academic knowledge production.

Under conditions of probabilistic risk, it should be possible for entrepreneurs to share the objective, codifiable information they use to construct their forecasts. They may prefer to keep some private information confidential for strategic reasons (which explains why large firms often rely on internal capital markets rather than outside investors). If they can test their theories of value creation through experimentation and Bayesian updating, then even subjective prior beliefs can be made “objective” for sharing with potential partners. (I share Nicolai’s reservations about this understanding of entrepreneurial experimentation—will post more on that soon.)

What if the entrepreneur’s judgments about the future are tacit, subjective, and difficult (or impossible) to codify or parameterize? Is the entrepreneur then forced to go it alone?

Intersubjective uncertainty

In a new paper forthcoming in Small Business Economics, Nicolai, Samuele Murtinu, and I explore the techniques used by entrepreneurs to convince others to participate in their projects. The core problem is well understood by practitioners: entrepreneurs and stakeholders often hold different visions of the future. These differences are not always reducible to missing information or poor signaling. Rather, they often reflect genuinely incompatible ways of framing a project’s potential. We call this phenomenon intersubjective uncertainty—a lack of shared belief about what the relevant future states are and how a venture’s activities map onto those states.

In fact, one could even define Knightian uncertainty as the condition under which parties can’t agree about possible futures and their likelihoods. Rather than treat uncertainty as an individual-level epistemic condition—Does the decision-maker believe she is maximizing expected utility under a given probability distribution, or does she believe she is making an intuitive judgment?—we could treat it as a group-level condition. If the parties agree on future possibilities, and can reach agreements to act in certain ways based on these beliefs, then we can say they’re operating under conditions of risk. If they cannot agree, then we can say they are operating under conditions of uncertainty, regardless of what they think they are doing. (The journal reviewers wisely encouraged us to cut this longer discussion from the paper, but I intend to return to it at some point.)

Novelty, opacity, and why some ventures are hard to evaluate

Back to the paper: Two characteristics of entrepreneurial projects help explain why intersubjective uncertainty is so persistent. One is novelty—how far a proposed product or business model departs from familiar categories. The other is opacity—how difficult it is for an outside observer to see the causal linkages between resources, activities, and outcomes.

Novelty and opacity are not the same. A highly novel idea might still be transparent if its logic is easy to grasp; a familiar domain can be opaque if the mechanisms generating value are complex. When either dimension is pronounced, stakeholders may struggle not just to price risk but to interpret what the entrepreneur is actually proposing.

Beyond contracts: Persuasion, framing, and cognitive alignment

Traditional models of entrepreneurial finance emphasize incentives, information asymmetries, and contractual design. These are certainly important. But under deep uncertainty—where probabilities are unknown and mental models differ—contracts alone cannot generate cooperation. What also matters is whether stakeholders can develop sufficiently compatible interpretations of the imagined opportunity.

This helps explain why entrepreneurs devote so much effort to activities that seem peripheral from a purely financial perspective: crafting narratives, staging demonstrations, building prototypes, signaling commitment, and designing governance structures that convey interpretive cues. These practices don’t eliminate uncertainty; they help create enough shared understanding for coordinated action to occur.

Discriminating alignment and the social side of entrepreneurial finance

Our framework, in tradition of Williamsonian discriminating alignment, links project characteristics to the enrollment strategies and financing arrangements most likely to work. Projects that are relatively familiar and transparent can often rely on standard debt or equity financing because stakeholders already share relevant cognitive categories. Others require extended cycles of persuasion and belief revision before external commitments become feasible. In some cases, entrepreneurs begin with self-funding precisely because outsiders cannot yet make sense of the imagined opportunity.

Seen this way, entrepreneurial finance is as much a cognitive and communicative process as a contractual one. Entrepreneurship under uncertainty is not just about making bold decisions; it is also about making the future intelligible enough that others will commit resources. Appreciating this intersubjective dimension helps explain both why promising ventures sometimes struggle to attract support and why successful entrepreneurs often function as interpreters of uncertain possibilities, not merely decision-makers.

The content and sequencing of entrepreneurial persuasion

In a companion paper, forthcoming at the Review of Austrian Economics, we explore a related but distinct aspect of this enrollment problem—namely, the content and sequencing of entrepreneurial persuasion. While the SBE paper focuses on how characteristics of the project shape stakeholder alignment, the companion piece looks more closely at how entrepreneurs structure their communication efforts over time. The key idea is that persuasion under Knightian uncertainty is rarely a one-shot event. Entrepreneurs can often reduce cognitive barriers sequentially, addressing some dimensions of uncertainty before others rather than trying to resolve everything simultaneously.

In particular, we suggest that entrepreneurs may alternate between emphasizing outputs (the envisioned product, market, or value proposition) and inputs (the resources, capabilities, and organizational arrangements needed to realize that vision). Sometimes it makes sense to lead with the ends—clarifying the imagined product or service first and worrying about means later. In other situations, especially when the imagined opportunity itself is hard to categorize, starting with familiar resources or capabilities can create a foothold for shared understanding.

The two papers thus speak to different but complementary questions: one about how project characteristics shape stakeholder alignment, the other about how entrepreneurs actively manage that alignment through the timing and framing of persuasion.