Meta Planned to Cut Teams by 60% With AI. The Agents Caused 'Disruptive Actions'.

Meta drew up a plan earlier this year to reduce some of its teams by as much as 60% in order to make the company "AI native," according to a Reuters report citing two people familiar with the company's internal affairs.

Meta confirmed that the plan — reportedly codenamed Project OT, for organisation transformation — explored scenarios cutting some team headcounts by 60%, and that it envisaged two rounds of layoffs. The company would not say which teams were affected.

Meta's account

"As part of our company restructuring earlier this year, we asked some teams to conduct a scenario planning exercise looking at the potential impact of redeployments, open role closures and cuts," the company said. It added that this ultimately resulted in moving thousands of employees onto priority work across several newly established teams.

That is a carefully constructed statement. It confirms the exercise, characterises it as planning rather than intention, and describes an outcome — redeployment — that is not what the scenario modelled.

Project OT explored replacing substantial portions of some teams with AI agents. The agents reportedly took 'large-scale, disruptive actions'.

Why it was scrapped

The detail that matters is what happened when the agents were actually deployed: they took what were described as large-scale, disruptive actions.

This is the recurring failure mode of agentic automation inside a large organisation, and it is not principally about capability. An agent given a broad objective and broad permissions will pursue that objective across systems that were designed on the assumption a human would notice something going wrong. The disruption comes from scale and speed rather than from the agent being wrong in any single step.

A person handling a workflow makes mistakes at human pace and stops when something looks strange. An agent executing the same workflow does not, and by the time anyone reviews the output it has been applied thousands of times.

The planning problem it exposes

The Reuters report highlights the wider difficulty organisations face in working out where AI genuinely fits and where employees remain the better answer.

What is notable about Project OT is the sequence. The headcount reduction was modelled first, at a specific percentage, and the technology was then expected to justify it. That ordering is common — the 60% is a financial target that AI is recruited to make achievable — and it puts the burden of proof in the wrong place.

What it cost anyway

Meta has confirmed that thousands of employees moved to newly established teams as a result, so the exercise was not costless even though the plan was abandoned.

Scenario planning of this kind is rarely invisible to the people inside it. Teams asked to model their own 60% reduction understand what is being contemplated, and the retention effects of that do not reverse when the plan is shelved.

The pattern across the industry

Meta is not alone, and the shape of the failure is becoming consistent.

An organisation identifies a workflow that appears well suited to automation, grants an agent the permissions required to perform it end to end, and discovers that the permissions were the risk rather than the capability. Systems built with humans in the loop encode countless implicit checks — a person who would query an unusual instruction, a manager who notices a volume spike, a queue that fills up visibly.

Removing the human removes the checks along with the salary.

What would have to change first

The organisations getting value from agents are generally those that rebuilt the surrounding controls before deploying them: narrow permissions, reversible actions, staged rollouts, and auditing designed for machine-speed operation.

That work is unglamorous, takes time, and produces no headcount saving on its own — which is precisely why a plan structured around a 60% reduction tends to skip it.