Why Meta Scrapped Its Plan to Replace Staff With AI

Picture a Meta engineer this spring, noticing that the tracking software logging her mouse clicks and keystrokes is not just measuring productivity. She and her colleagues come to believe it is training the AI agents meant to replace them. The internal network fills with furious posts, the satisfaction survey craters, and somewhere upstairs a plan with a codename starts to wobble. That plan, "Project OT," is the subject of a Reuters special report published August 26, built on internal documents Meta has not publicly commented on.

The blueprint: 60 percent cuts, two waves

According to the planning documents Reuters reviewed, Project OT would have shrunk many teams by up to 60 percent, leaving small, "talent-dense" groups of humans supervising virtual AI workers. People would move from producing the work to auditing it. Execution was designed in two waves: a first purge in May and a second restructuring in November. The first wave happened; Meta laid off 10 percent of its workforce the next day. But on the evening of May 19, hours before those first cuts landed, Zuckerberg killed the November wave, and per Reuters, planning for it never resumed.

The numbers that broke the plan

The internal data tells a clean story of a gap between promise and delivery. Employees were generating far more code with AI tools, yet the surge was not translating proportionally into products reaching users. Meanwhile the cost side of the ledger grew: major technical and security incidents, outages and possible data leaks among them, rose 40 percent year over year, and the time employees spent firefighting those problems jumped 70 percent. The assembly line sped up, and so did the repair shop, quietly eroding the productivity assumption the whole restructuring rested on. Add investors publicly questioning the AI budget and an internal satisfaction score sliding from 74 to 55 percent, and the plan was taking fire from three directions at once: the technology underdelivered, the market doubted the spend, and the workforce refused to train its own replacements.

The lesson smaller companies get for free

One of the deepest-pocketed companies on earth, with elite engineers and effectively unlimited GPUs, tried to hand entire roles to agents and walked it back. That matches what we see in far smaller deployments: agents absorb tasks well and roles badly. Defined, bounded work, invoice matching, first-draft replies, report generation, delivers real gains. Handing over a whole role, with its context, exceptions, and accountability, comes back as outages and security workload, which is exactly what Meta's own incident numbers show. A ten-person firm that automates the reconciliation emails but keeps a human owning the ledger gets the upside; one that replaces two staff outright inherits Meta's problem at a scale that cannot absorb it.

Sources: The Decoder, Engadget, Reuters (The Globe and Mail)