The Org Chart Implodes

The most interesting AI deployment story of the year did not happen in a lab or a data center. It happened in an org chart. Today, Reuters pulled the curtain back on Project OT — Meta’s internal plan to make the company “AI native” by letting AI agents perform “much of the daily work performed by thousands of human employees,” with small human teams supervising. The plan included cutting some teams by as much as 60 percent, in two waves. The first wave happened in May. The second was canceled. And the internal record, which Reuters reviewed across “scores of internal documents, posts, and recordings,” explains why: the agents were making “large-scale, disruptive actions that humans are unlikely to execute.”

The Plan on Paper

Project OT — “organization transformation” — was created in January and set in motion by Mark Zuckerberg, who directed executives to proceed with the management-structure changes. The scenario math was not subtle: one HR executive told Reuters that under the plan, Meta would have reduced headcount by roughly 25 percent or more. Some of the savings, two people familiar with the matter said, was going to be redirected to pay high-performing employees — especially people with AI engineering skills. The most efficient way to fund the AI future, in other words, was to fire the present one.

The internal definition of “AI native,” per a planning document Reuters reviewed, was a company where “AI-ready tools and agents interact, workflows are automated, [and] new builds are AI-first” — plus selling AI agents to third parties, which Meta has in fact started doing since June. The pilot had already restructured engineering teams, research teams, and at least eight other teams into smaller groups, following an internal “AI-Native Playbook” post that outlined removing middle management and using “agent-assisted analysis” to prioritize daily work. HR and “AI systems” were to help leaders decide promotions. Meta’s response to that particular claim is worth preserving: “Performance rating and promotion decisions were and are made by people, not AI.”

What the Machines Did

The rollout’s internal record reads like a slow-motion incident report. By March, internal warnings about agent reliability were already circulating. By April, an internal post described agents making “large-scale, disruptive actions that humans are unlikely to execute.” The aggregate numbers behind those posts: a 40 percent year-over-year increase in major technical and security incidents, and up to a 70 percent rise in the time employees spent fixing them. The agents were doing the work. The work included the failures, and the failures had a price.

The context makes the incident rate land harder. Meta had been tracking employees’ keyboard and mouse input to train its agents — a program it has since paused after an internal security breach — and in March an agent misinstruction leaked sensitive data to employees. A separate lawsuit, filed in July, claims Meta’s own layoff decisions were made by AI, not humans. The human-in-the-loop fiction, it turns out, had a long way to go in both directions.

The Honest Reading

Meta’s official statement calls Project OT a “scenario planning exercise” and adds, with the calm of a company that has already run one of the scenarios: “Ultimately, we didn’t move forward with every scenario from the exercise — and it was never assumed we would.” That is true and also beside the point. The May layoffs happened. Thousands of employees were moved to newly established teams. A scenario that has already executed a wave of layoffs is no longer a scenario — it is a project, and projects that get canceled get post-mortems. This is the first honest cost accounting of the “AI native” thesis, and the cost was not model capability. It was the incident rate.

We didn’t move forward with every scenario from the exercise — and it was never assumed we would.— Meta, on the plan that laid off thousands

Two days ago, this desk published a survey finding that 59.5 percent of senior enterprise leaders already run AI agents autonomously in production. Project OT is the counterweight to that number, and the difference between the two is the entire story. The autonomy majority is bounded: agents run unsupervised on workflows chosen precisely because failure is cheap, inside guardrails, on tasks where a mistake is a retry. Project OT was unbounded: agents doing open-ended daily work inside an organization whose failure mode is a 40 percent incident spike. The bounded deployments are working. The unbounded one imploded. That is not a contradiction. That is the whole finding.

What the Imploded Org Teaches

The org chart is not just a list of who does what. It is an interface — and interfaces include accountability. When a workflow fails, the org chart answers the question that no model can: who owns it? Project OT tried to replace the interface with the model and keep the rest of the building standing. The building did not stand. The next wave of enterprise agent news will not be about headcount. It will be about incident rates, audit trails, and who signs the post-mortem. The companies that can answer those questions will keep the machines. The rest will keep the lessons.

The org chart imploded. The machines did the work, and the work included the failures. The second wave of layoffs never came — not because the agents got better, but because the incident rate finally got counted.