The Enterprise Was Built to Ration Cognition

Organizational cognition has always been unevenly distributed. Every company I have worked in has had some version of the same person: the one whose name appears when a difficult problem stops moving. Someone says, “Ask him.” Or: “She’ll know.” Eventually the joke becomes darker: if that person gets hit by a bus tomorrow, we can close shop.

We usually treat that as a knowledge-management problem. I think it reveals something larger about how the enterprise was designed.

Modern organizations were partly shaped around the fact that human cognition is scarce, unevenly distributed and slow to transfer. Expertise accumulates in people. Context accumulates with seniority. Judgment improves through situations going right and wrong. So companies built mechanisms for routing work toward the people who knew more: specialist functions, management layers, escalation paths, reviews and approval chains.

Those structures did more than ration cognition. They also distributed authority, established accountability, coordinated action and created trust. AI does not make those functions disappear. But it changes one of the constraints underneath them: the cost of making useful cognition available where work happens.

When cognition becomes cheaper to distribute, the organization should stop making decisions travel toward context when context can travel toward the decision.

Hierarchy was also an information technology

There is a long intellectual history behind the idea that organizations are responses to constraints rather than natural laws. Ronald Coase explained the existence and boundaries of firms partly through transaction costs: markets have costs, administrative coordination has costs, and firms emerge where one arrangement can coordinate activity more economically than another. Coase’s Nobel lecture makes the broader point useful here: organizational arrangements have an economic logic.

Herbert Simon gets even closer to the cognition problem. In his Nobel lecture on decision-making in organizations, he described organizations as machinery for coping with limits in human ability to comprehend and compute under complexity and uncertainty. One response is to divide decision-making among specialists and coordinate their work through communication and authority. Simon’s formulation of bounded rationality gives us a useful historical anchor for thinking about organizational cognition today.

That does not mean hierarchy exists because people are not smart enough. It means organizations had to solve a routing problem. Relevant information, experience and decision capacity could not exist everywhere at once. Hierarchy, specialization and management became part of the infrastructure for moving limited attention and knowledge through the firm.

AI does not repeal Coase or Simon. Transaction costs remain. Human rationality remains bounded. The interesting change is narrower and more practical: some forms of analysis, retrieval, synthesis and pattern matching can now be replicated and distributed at dramatically lower marginal cost. That means one important input into organizational design is moving.

I experienced the old constraint directly while leading a cross-border product organization. Part of my role was legitimately a decision layer. I had a say in the product vision. But the more important objective was to make that vision shared intellectual property: to mentor teams until they could make good decisions inside it without treating the vision as something that belonged only to me.

That took years. New teams first had to learn how I thought about the product, the customers and the trade-offs. A team working several time zones away made the transfer even slower because much of the relationship had to be built asynchronously. The problem was not simply whether I had delegated a decision. The person best positioned to make the decision and the person carrying enough context to make it well were often not the same person.

Organizational cognition changes what empowerment requires

Organizations have spent decades talking about empowerment and pushing decisions closer to the people doing the work. The principle is sound. But a decision right is not enough. Decision rights without context are theatre.

If the person closest to the problem lacks the historical context, comparable cases, policy knowledge, customer evidence or understanding of dependencies required to act, the decision travels upward anyway. Often it lands with a senior officer because seniority has accumulated context. That officer may know more about the organization while knowing less about the specific domain than the person who originally encountered the problem.

We then mistake the routing pattern for evidence that rank produces better decisions. In reality, rank may simply be correlated with access to organizational cognition.

This is why AI changes the operating-model conversation. The interesting question is not only whether AI can make the senior person faster. It is whether more of the context that made the senior person necessary can become available where the decision originates. That is a different form of democratization: not handing out authority indiscriminately, but reducing the information asymmetry that forces ordinary decisions upward.

The expertise queue describes this at workflow level. Work waits because the useful cognition lives somewhere else. The Macro implication is larger: if organizational cognition can move, some of the structures built to route work toward it deserve to be re-examined.

I have watched scarce knowledge become less scarce. It was slow.

One software team I worked with had a developer who had reindexed the product database and knew its roughly 80 tables almost by heart. That knowledge made him unusually valuable. It also made the organization dependent on him.

A surprisingly important intervention was stronger code review. Other developers reviewed his work even when most reviews found nothing to change. The point was not catching errors. The process exposed other people to the decisions, structures and knowledge that had accumulated in one person.

Over time, the specialist became less singular because the team became more capable. That broader change was one of several important elements in a period in which development output increased by more than 300 percent. I would not attribute that increase to code review alone. What the experience demonstrated much more safely was that distributing scarce knowledge changed the capacity of the system.

But there is a catch: knowledge transfer took time. People had to encounter the work. Reviews had to happen. Relationships had to form. Lessons had to survive long enough to become habits. Organizational cognition spread through repeated human contact.

Generative AI changes the economics of that transfer. In a large customer-support deployment studied by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, access to an AI assistant increased productivity by 14 percent on average and by 34 percent among novice and lower-skilled workers, with minimal impact on the most experienced workers. The researchers found suggestive evidence that the system disseminated practices associated with more capable workers. The study covered 5.179 customer-support agents.

Customer support is not every knowledge job, and a field experiment does not prove a universal operating model. The mechanism is what matters: patterns that once moved primarily through apprenticeship, proximity and repeated exposure can also be made available at the moment another worker needs them.

That is a fundamentally different distribution technology from waiting for documentation, the next review, the next training session or the next meeting with the person who knows. It does not eliminate learning. It changes how quickly the organization can reuse what it has already learned.

The Context–Authority Split: six functions hiding inside hierarchy

This is the operating-model proposition I think matters. Many organizations have bundled cognition and authority into the same hierarchy. Senior people have more decision rights, but they also tend to have more organizational history, broader networks and greater access to information. When a decision travels upward, it can be difficult to tell which of those things made the escalation necessary.

AI gives us a reason to separate them. Before preserving or removing an organizational layer, ask what scarce function it is actually providing.

AI can dramatically change the cost and availability of the first category. It can augment the second. It does not automatically inherit the remaining four. This is why “AI will flatten the organization” is too crude.

A layer that exists mainly because information has to travel through it is exposed to redesign. A layer carrying genuine authority, accountability, coordination or trust may remain essential even when everybody can access the same analysis. The future organization may be flatter in some places, but flatness is not the objective. Better alignment between context and decision rights is.

This extends an argument I made in The Org Chart May Be Legacy Infrastructure for the AI-Native Organization. The point is not that structure disappears. The point is that we can become much more explicit about which structure exists for which reason.

There is another implication. If organizational cognition becomes easier to distribute, management can no longer justify every escalation with “I need visibility.” Visibility and decision ownership can be separated. A manager may still need to set objectives, resolve conflicts, allocate scarce resources or accept risk. But those are clearer jobs than acting as a human router because context cannot otherwise reach the team.

Do not spend the capacity dividend by deleting the bottom of the pyramid

Professional services makes the tension especially visible. The traditional leverage model has a simple economic logic: bring in junior talent as comparatively inexpensive labour, place expensive experience above it, and generate margin by distributing work through the pyramid. If AI can perform more of the research, synthesis, first-pass analysis and production work, the immediate spreadsheet answer is obvious: employ fewer junior people.

I think that is an unimaginative use of the capacity dividend.

Young professionals should not exist to be cheap versions of experienced professionals. They should bring fresh ideas, challenge patterns that have become assumptions and become the future source of organizational renewal. Experienced professionals contribute something different: pattern recognition built from seeing situations succeed and fail.

An AI-native professional-services model can use machines to reduce cognitive overhead across every layer while creating more—not less—interaction between fresh perspective and accumulated judgment. Instead of paying juniors to manufacture slides that senior people already know how to interpret, the firm can expose them earlier to problems, customers, experiments and consequences. Instead of asking seniors to repeatedly reconstruct familiar analyses, it can use their attention on exceptions, mentoring and judgment.

That is a very different talent model. The capacity dividend is not merely a headcount dividend. It is the opportunity to change what humans at every level spend their scarce attention doing.

This sharpens the argument from The Expertise Premium Is Moving. If AI makes reusable knowledge easier to distribute, seniority cannot justify itself merely by being the place where organizational memory lives.

Seniority after knowledge scarcity

More experienced people have not necessarily learned more at university. They do not always have more credentials. What they reliably have is a longer history of consequences.

They have seen the transformation that looked perfect on paper and failed because nobody trusted it. They have seen the technically weaker option win because it fit the organization. They have watched a risk materialize after everybody agreed it was unlikely. They have learned which patterns repeat and which superficial similarities are dangerous.

AI cannot give a person that biography. But it can help prevent the lesson from remaining a one-off event trapped in that person’s memory. Organizational cognition improves when individual experience becomes reusable without pretending that the system itself lived through the consequences.

That changes the Human Premium. The senior employee who derives status from being the only person who knows should expect some of that status to erode. The senior employee who can recognize exceptions, mentor judgment, build trust, accept accountability and turn experience into better organizational learning becomes more valuable.

There is a warning here. A September 2026 randomized field experiment with 133 practicing patent lawyers found that AI improved AI-assisted drafting quality, but durable gains on a later unassisted professional-judgment task were concentrated among senior lawyers. Junior lawyers showed no average gain on that later task. The researchers conclude that foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.

That should make leaders cautious about confusing access to expert-like output with the development of expertise itself. If AI removes the work through which novices once learned, organizations need a deliberate replacement for the learning—not a nostalgic preservation of low-value work, but better apprenticeship.

The question for talent systems is therefore not “Which junior tasks can AI eliminate?” It is “Which experiences must humans still accumulate to become the people whose judgment we will need later?” That is a harder question, but it is also a much more consequential one.

Some scarcity should survive the Cognitive Revolution

Regulation is a useful boundary case. Regulatory knowledge can be unusually suitable for AI-supported retrieval because much of it is documented and changes on identifiable cycles. Making that context broadly accessible can reduce unnecessary dependence on the person who remembers which rule applies.

But democratizing regulatory context does not democratize regulatory accountability. The organization may still need formal interpretation, independent review, documented controls and a person who owns the risk. The AI system can help bring the rule to the decision. It cannot simply declare itself the legitimate bearer of responsibility.

This is precisely why the distinction matters. Cheap cognition does not make capital cheap. It does not make trust abundant. It does not eliminate legal responsibility. It does not remove physical constraints or resolve stakeholder conflict. And it does not make every decision reversible.

Nor should we assume that cheaper organizational cognition automatically changes collective behaviour. A large cross-industry field experiment involving 7.137 knowledge workers across 66 firms found that generative AI reduced time spent on activities such as email, but did not produce comparable changes in coordination-heavy activities such as meetings. The result is a useful reminder that individual cognitive leverage can arrive before the organization redesigns the coordination around it.

The future organization is therefore not necessarily flatter. It is less dependent on hierarchy as a knowledge-distribution system. That is a narrower claim than the disappearance of management, but I think it is much more useful.

The organizational cognition diagnostic: find the queue for a name

If I were looking for the first place to redesign, I would not begin with the org chart. I would listen for names.

Where does work stop until a particular person becomes available? Where do colleagues say, “Only she knows how this works”? Where does a manager repeatedly become the human router between a problem and somebody else’s expertise? Where does the same senior person reconstruct context for the hundredth time?

Then ask why the queue exists. If people are waiting because the named person carries legitimate authority, accountability, trust or a genuinely exceptional judgment call, the queue may be expensive but necessary.

If they are waiting because useful knowledge, historical context, known patterns or routine analysis exist primarily inside that person’s head, you have found organizational cognition being rationed by the operating model.

This is where AI-native redesign should start. Not by removing the person. By making the organization less dependent on routing familiar cognition through them and giving that person more capacity for the judgment, mentorship, trust and exceptions that made their experience valuable in the first place.There is an interesting analogy to make here: AI cannot replace meetings, but the collateral effects that make people avoid meetings, such as preparation and debriefing, can be almost eliminated by AI. You can also trust a much more diversified agenda if the meeting’s outcome is no longer dependent on the limitations of human cognition. The same applies to the decision queue: is it blocked by decision itself or rather by the work around it?

It also changes what leaders should measure. A healthy organization should not merely count how many AI tools are deployed. It should be able to see whether decision latency is falling, whether named-expert queues are shrinking, whether teams can access prior learning without escalation, and whether senior attention is moving toward genuinely exceptional work.

That is the deeper promise of organizational cognition as a design lens. AI is not only another capability inside the existing enterprise. It changes the economics of how context can be assembled, remembered and delivered. Once that happens, some old routing mechanisms stop being neutral infrastructure and become a choice.

The enterprise was built, in part, to ration cognition. The Cognitive Revolution gives us the opportunity to stop rationing what no longer needs to be scarce.