Insights
AI, Management Layers, and the Future of Hierarchy
Every org chart is an answer to a question almost nobody asks anymore: what does it cost to move knowledge?
Twenty-five years ago the economist Luis Garicano offered the cleanest explanation of hierarchy we have. Organizations build layers because knowledge is expensive. You cannot afford to have everyone know everything, so you let most people handle the routine cases and route the exceptions upward to someone who knows more. A layer is not a status symbol. It is a compression device for scarce expertise.
Hierarchy has always been a technology
Which means hierarchy has always been a technology, and technologies have assumptions. Garicano’s assumption was that knowledge is costly to acquire and costly to transmit. That assumption is now visibly loosening. The evidence is stranger than the slogans suggest.
A 2025 NBER working paper reconstructed the internal hierarchies of more than 2,500 U.S. public firms from the résumés of 16 million employees. The average firm turned out to have about ten layers. Companies added layers after demand shocks — exactly as the theory predicts — and, notably, flattened them after adopting AI. Cheaper knowledge, fewer routing stations.
But a longitudinal study of Italian firms found something close to the opposite: firms investing in emerging information technologies got deeper, with narrower spans of control and more middle management. A recent Harvard Business School working paper predicts flattening, narrowing and winnowing all at once. A modelling paper on generative AI in the knowledge economy finds that as AI becomes more capable, spans may actually narrow, because someone has to supervise output that is plausible whether or not it is correct.
So: does AI flatten organizations or thicken them? The research disagrees, and I think the disagreement is the finding.
Layers are not disappearing
Layers are not disappearing. They are losing their original job and looking for a new one.
When routing knowledge was the work, a manager’s value was proximity to answers. You escalated to them because they knew. Remove that, and a layer has two futures. It can become a control layer — approvals, reporting, verification, the endless checking of machine output. Or it can become something we have never seriously designed: a layer whose product is other people’s thinking.
Most organizations will drift toward control, because control is legible. It generates documents. It survives budget reviews. And there will be real work there — someone must own the judgment that AI cannot own. But a company that converts its entire managerial capacity into a quality-assurance function has quietly decided that its people are a risk to be managed rather than an intelligence to be assembled.
The alternative is not flatter
The alternative is not flatter. It is differently loaded.
Hamel and Zanini once estimated that excess bureaucracy costs the U.S. economy over three trillion dollars a year, roughly 17% of GDP, and that the country carried a bureaucratic class of some 24 million people — about one manager or administrator for every five workers. Their conclusion was that bureaucracy is a tax on human potential. I would put it slightly differently. Bureaucracy is what a coordination layer becomes when its coordination problem has been solved and nobody has told it.
We are about to solve a great deal of the coordination problem. The question is what we ask those layers to do next.
My own guess is that the most valuable manager in an AI-native organization will not be the one who knows most, or who checks most, but the one whose team produces the best questions. Not a router of knowledge. An architect of conditions.
That is a genuinely new job, and no org chart in the world currently has a box for it.