Cognitive bottlenecks are where work still waits for scarce human context, synthesis or judgment even after AI has made much of that cognition easier to distribute. The tempting response is to remove layers. That is too crude. The more useful starting point is the named-expert queue: the place where work stops because one particular person carries the context, precedent or analytical capacity required to move it forward. AI can attack much of that waiting without pretending that authority, experience or accountability have become abundant.
Leadership Question
Where is a senior person still acting as the place cognition has to travel through—and how much of that queue can be redesigned without moving the accountability that legitimately belongs with them?
Start with the cognitive bottlenecks that have names
A practical AI operating-model review should not begin with the org chart. Listen instead for sentences such as “Ask her,” “Only he knows,” or “We cannot decide until they are back.” Those phrases reveal where organizational cognition is concentrated. They also expose cognitive bottlenecks that are easier to diagnose than a generic discussion about hierarchy.
The expertise queue describes the workflow-level problem: routine cognition waits for scarce people because knowledge, pattern recognition and context are concentrated in them. The organizational-cognition argument widens the lens: hierarchy has bundled cognition with authority, accountability, coordination and trust. AI changes the economics of the first of those much more directly than the others.
The executive implication is therefore not “flatten.” It is decompose the queue. Ask what the senior person is actually providing. If the answer is historical context, retrieval, synthesis, comparison with prior cases or preparation of options, much of the work may move closer to the point of need. If the answer is accepting liability, committing resources, resolving stakeholder conflict or making a judgment whose consequences they must own, the escalation may still be doing important work.
Redesign the cognitive preparation before you redesign the decision right.
Cognitive bottlenecks are easiest to remove around preparation
Consider a consequential approval. The final decision may still belong with an experienced leader because that person carries accountability a junior employee does not. But why should the accountable person also be the one who has to reconstruct the history, find comparable cases, gather evidence, challenge assumptions and discover missing context?
AI can compress that surrounding cognitive work. The same principle applies to workshops, approvals and decisions: preserve the consequential human event where it earns its place, while attacking the cognitive overhead around it. This is a lower-regret intervention than transferring authority, but it can still reduce cognitive bottlenecks, shorten decision latency and return scarce senior capacity.
Recent evidence reinforces the distinction. A field experiment across 66 firms and 7.137 knowledge workers found that generative AI reduced time spent on email among active users, but did not produce comparable changes in the overall quantity or composition of work. Individual cognitive leverage arrived more readily than organizational redesign. The NBER study is a useful warning against assuming that giving people AI automatically changes coordination-heavy work.
Do not let “risk” become a blanket defense of the old queue
Accountability is a real boundary. “Risk” is often less precise. Organizations can use the word to defend an approval chain without distinguishing the risk of the decision from the weakness of the preparation around it. That can leave cognitive bottlenecks untouched even when the underlying control objective could be met in a better way.
AI can itself become part of risk reduction: second-guess an analysis, surface contradictory evidence, test an option against prior failures, or create a control gate before a human commits the organization. That does not transfer accountability to the machine. It can improve what reaches the accountable human.
The boundary matters because access to stronger output is not the same as possessing stronger judgment. A September 2026 randomized field experiment with 133 practicing patent lawyers found that AI improved assisted drafting quality, while 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 argue that foundational expertise may be necessary to extract durable skill from AI-assisted practice.
This is why accountability should not be treated as another bottleneck to automate away. Experience still carries a history of consequences. The opportunity is to remove cognitive bottlenecks around that judgment, make stronger preparation available to more people and use the returned capacity to help others develop better judgment.
Spend the capacity dividend on broader cognition
The weakest AI productivity strategy is to remove cognitive work and quietly replace it with more volume. A better test is whether returned capacity expands the organization’s strategic and creative range: more time on difficult problems, more collaboration, and more diverse cognitive effort brought into consequential work.
This connects directly to the AI Capacity Dividend: efficiency only becomes transformation when recovered time, attention and decision capacity are deliberately reinvested rather than absorbed by the old operating model.
Meetings illustrate the point. New research using survey and administrative data from more than 9,000 workers in Norway estimates that meetings consume substantial work time and wage cost. Yet the same study links meetings to planning, problem solving, coordination, learning and knowledge transmission, with interactions involving senior colleagues associated with stronger wage growth. The evidence does not make every meeting valuable. It does make indiscriminate elimination look simplistic.
The AI-native question is different: how much preparation, context assembly and debriefing around a valuable human interaction can become dramatically cheaper? If that overhead falls, the organization can afford richer participation without proportionally increasing cognitive cost. The capacity dividend can fund broader contribution rather than simply fewer contributors.
Signals to Watch
- Named-expert latency: work repeatedly waits for the same people even when the requested contribution is mostly context, retrieval or synthesis.
- Preparation-to-decision ratio: senior time is consumed reconstructing context rather than exercising consequential judgment.
- Risk without decomposition: “risk” or “governance” is used to preserve a queue without specifying which control, liability or failure mode requires it.
- Capacity without reinvestment: AI saves time but strategic range, collaboration, creativity and stakeholder outcomes do not improve.
- Judgment erosion: output becomes faster while client satisfaction, employee fulfillment, decision quality or learning deteriorates.
Measure outcomes before celebrating efficiency
The baseline should be outcome quality and impact. If cognitive bottlenecks shrink while clients, employees and the company experience no deterioration in satisfaction, fulfillment or results, the redesign has already created a useful efficiency dividend. If those outcomes hold steady, leaders then have a choice about where to reinvest the capacity. If they deteriorate, the organization has probably removed more than cognitive overhead.
This is also why AI adoption statistics alone tell leaders very little about operating-model progress. Recent U.S. Census-based research finds that most adopting firms still use AI across a limited number of functions and tasks, with task augmentation much more common than employment reduction. The diffusion evidence suggests that deployment breadth and operating integration remain uneven. The important management question is not how many tools exist, but whether cognition is reaching decisions with less friction and whether the returned human capacity is being used better.
The executive principle
When cognition gets cheap, leaders should stop setting themselves up as bottlenecks, preserve the good judgment that comes from experience, and invest the returned capacity in democratizing that quality of judgment across the organization.
That is a more demanding objective than automation. It asks leaders to make their own context less scarce without pretending their accountability is transferable. It asks experts to turn experience into organizational learning rather than positional dependence. And it asks the enterprise to treat AI-generated capacity as something to invest, not merely harvest.
A useful first move is small: find the queue for a name. Separate preparation from judgment. Move the reusable cognition. Preserve the accountability that still earns its place. Then measure whether the people, customers and organization are actually better off.
