Insights
AI-Native Firms and the Loss of Middle Management
A generation of companies is now being built without the middle of the organisation, and almost no one is asking what the middle was doing.
A Harvard Business School working paper this year (26-090, “AI-Native Firms”) classified Y Combinator batches from 2020 to 2024 and US venture-backed startups, then linked them to workforce data on team size, function and seniority. Compared with non-AI startups in the same industry and cohort, AI-native firms are about 25% smaller. Their hierarchies are roughly half a seniority level flatter. They have about 15% fewer managers — and about 15% fewer entry-level workers. Valuations are comparable.
The first three findings are being celebrated. The fourth is being ignored.
There is a matching signal in the labour data. A US Census Bureau working paper (CES-26-27, April 2026), using matched employer-employee administrative records, found an immediate and persistent fall in hiring of 22-to-24-year-olds in the industry-state cells most exposed to AI: employment of early-career workers in the most exposed quintile fell about 12% over the ten quarters following ChatGPT’s release, while less exposed industries stayed flat.
Two independent datasets, one shape. The organisation is losing its bottom layer and its middle layer at the same time.
I want to be careful here, because the obvious reading is the wrong one. This is not mainly a story about young people losing jobs. It is a story about a function no one wrote down.
The hierarchy was never only a coordination device
It was also a manufacturing process.
Junior roles were how organisations produced judgement. Not through training programmes — through exposure. You made small decisions badly under supervision, and someone senior explained why they were bad. Middle management was where that explanation happened. The layer everyone described as bureaucratic overhead was, quietly, the place where people learned which problems were worth solving, how much uncertainty is normal, when to escalate, and what a good decision actually feels like from the inside.
Nobody costed this. It was a by-product of a structure built for other reasons, which is why it can be removed without anyone noticing they removed it.
So the honest question for an AI-native company is not “how flat can we go?” It is: where does judgement now come from?
Three answers seem to be forming, and only one
The first is to buy it. Hire only senior people. This works while a stock of experienced judgement exists in the market — a stock produced by the older, layered organisations these firms are outcompeting. It is extraction, not production.
The second is to assume the tools supply it. They do not. Models supply fluency, options and drafts. Judgement is the capacity to reject a plausible answer for a reason you can defend. That is developed by consequence, not by access.
The third is to build the developmental function deliberately, now that it is no longer free. That means naming it: which decisions must a person make unaided in their first year, who reviews the reasoning rather than the output, and what is the organisation willing to let someone get wrong at survivable cost. In a firm with three seniority levels instead of five, this has to be designed. It will not fall out of the org chart, because there is no longer enough org chart for it to fall out of.
I find this the most interesting design problem in front of us. Flatness is not the mistake. Flatness is mostly good — fewer people relaying information that no longer needs relaying. The mistake would be treating the layers we deleted as pure cost, when part of what they cost was the production of the next generation of people capable of contributing anything at all.
Every organisation removing its middle is making a bet about where judgement comes from. Most have not noticed they are making it.