Eric Schmidt used to spend about 12 hours a day in meetings as Google’s chief executive. He mentions the figure now not as a war story but as a cost estimate, a measure of how much of a large company’s payroll goes to moving information between people who already work there.
Speaking on the In Good Company podcast with Nicolai Tangen, the head of Norges Bank Investment Management, Schmidt argued that AI agents are about to make a large share of that function unnecessary. His number: about 20% of the workforce at a big company could disappear and the place would still run.
He sketched the case with a hypothetical employer of 10,000 people. A few thousand of them, he suggested, are not strictly needed. The roles he had in mind are not the engineers or the salespeople. They are the planners, the people who assemble and reformat data, the administrators who keep compliance calendars and chase approvals. Schmidt’s word for them is the glue — the staff whose job is coordination rather than production.
The claim is less radical than it sounds and more uncomfortable than it reads. Large organizations have known for decades that coordination consumes a fixed share of white-collar hours; what has changed is that a software agent can now hold a project timeline, draft the status update, reconcile two spreadsheets and book the meeting that resolves the discrepancy. That is not a reasoning breakthrough. It is a labor substitution in the least glamorous layer of the org chart, and it is happening at companies that are under no immediate financial pressure to try it.
Schmidt pointed to a friend’s company that cut roughly 20% of its staff while still growing, a description that appeared to point to Cloudflare. That company has said AI compressed the preparation of its earnings documents from about two weeks to three minutes. The anecdote matters because of the ratio, not the technology: the work did not get faster and the team did not get smaller as separate decisions. One caused the other.
He was careful about the politics of it. Companies that remove roles, he said, have an obligation to handle the transition legally and humanely — severance, notice, help finding other work. He also conceded a limitation that cuts against his own forecast: established companies are often reluctant to reorganize, because reorganizing means someone has to decide which managers lose their teams. The technology may arrive faster than the willingness to use it.
The prediction market Kalshi puts roughly 94% odds that layoffs in the U.S. information sector will exceed 2025’s total in 2026. That is a market reading, not a measurement, but it captures the direction of the consensus. Technology companies have spent the past two years trimming in small, repeated rounds rather than in single dramatic cuts, which makes the aggregate easy to miss and the pattern easy to deny.
The distinction Schmidt draws is worth holding onto. The argument is not that AI eliminates work. It is that AI eliminates the coordination overhead that large companies accumulated because coordination used to be cheap to add and expensive to remove. Middle management grew for a reason that no longer applies with the same force: a person was the cheapest available router.
If that is right, the effects will show up first in hiring rather than in headcount. Junior roles exist largely as an apprenticeship into coordination work — you learn the business by assembling the deck, chasing the numbers, sitting in the room where the decision gets made. Remove the deck and the room, and the entry-level ladder loses its first rungs. Firms can cut 20% of staff and still operate. Whether they can still produce the people who will run the place in a decade is a different question, and not one a quarterly cost ratio answers.
Analysts said the arithmetic is harder to test than it looks. Coordination work rarely appears as its own line item; it is distributed across departments and bundled into titles that describe something else. A company can therefore reduce headcount by a fifth and report productivity gains without ever having measured what the removed layer was doing, which makes the early evidence for Schmidt’s thesis suggestive rather than conclusive.
The transition cost also lands unevenly. The employees most exposed are the ones with the most institutional knowledge and the least portable title, and they are typically the least equipped to argue that their work is not automatable, because the work is hard to describe. A plant manager’s output is visible. A program coordinator’s output is the absence of problems.
Schmidt’s own framing suggests he knows the estimate is a provocation rather than a projection. Twenty percent is a round number, offered on a podcast, defended by anecdote and a hypothetical. The interesting question is not whether it is precise.
It is whether 20% is the ceiling of what companies will cut, or the number they will be embarrassed to have called ambitious.
