Why Meta Shelved Project OT, Its Plan to Replace Work With AI Agents

Project OT was Meta’s attempt to make parts of its workforce “AI native”: AI agents would take on more day-to-day work while smaller groups of employees supervised the systems. Internal planning explored reducing some teams by as much as 60% in two waves. But Meta ultimately pulled back from the plan’s most aggressive near-term form. 117

The clearest explanation is practical rather than philosophical. The company’s desired operating model required AI agents to perform consistently enough that fewer people could safely oversee complex work. Reporting indicates that the technology, operational results, and employee response did not meet that test. 117

What Project OT was designed to do

Project OT—short for Organization Transformation—envisioned AI taking over substantial portions of work previously done by thousands of employees. The remaining human organization would be smaller and more concentrated, with “talent-dense” teams overseeing virtual workers. Meta confirmed to Reuters that it had explored scenarios involving reductions of up to 60% on some teams and two rounds of layoffs; it said the plan included reassignments and would not affect all teams. 117

That distinction matters. The reported 60% figure described scenarios for particular teams, not a confirmed company-wide target.

Why the plan lost momentum

Agents had not demonstrated dependable production performance

The workforce case depended on agents being more than useful copilots. They needed to be predictable enough to take on meaningful work without creating more review, remediation, and risk for the smaller human teams left behind.

Reuters’ reporting described internal concerns about agent performance and disruptive actions. Separate reporting based on that investigation said agents were capable of “large-scale, disruptive actions.” 18

For an organization responsible for large consumer platforms, that is a serious constraint: productivity gains are not enough if they are offset by unreliable execution or a heavier operational burden on the people who remain.

Staff resistance made the change harder to execute

A transformation framed around replacing much of employees’ work with AI naturally created anxiety about jobs, accountability, and workload. Reporting on Project OT describes significant internal resistance, and Reuters reported that the proposal failed to launch in its most sweeping form. 117

That resistance was not separate from the technical problem. When employees are asked to trust AI with consequential work, confidence depends on evidence that the systems improve results—not simply on ambitious headcount models.

Leadership acknowledged problems with timing and communication

According to reporting on a Reddit discussion by Reuters correspondent Katie Paul, Mark Zuckerberg said he had gotten the timing wrong, while CTO Andrew Bosworth said Meta had done an “atrocious job” explaining the vision internally. 7

Those comments suggest that Project OT’s failure was not only about whether AI could eventually improve productivity. It was also about rolling out a high-stakes workforce transformation before the evidence, operating model, and internal explanation were strong enough.

Why tokenmaxxing was a weak measure of AI adoption

“Tokenmaxxing” turned AI consumption into a visible signal of participation. At Meta, an internal leaderboard known as Claudeonomics ranked employees by token use and awarded labels such as “Token Legend.” But the leaderboard was created as a side project by an employee, rather than being established as Meta’s universal performance system, and it was later taken down at the creator’s discretion. Meta said it had a separate AI Insights dashboard that tracked usage more broadly than token counts. 3

The core measurement problem is straightforward: tokens measure model consumption, not business value. A high token count does not establish that a team shipped a better product, resolved an incident faster, reduced rework, or made a process safer.

A more durable AI-adoption standard is outcome-based:

• quality of the work produced;
• speed from idea to shipped result;
• complexity and scope of tasks completed; and
• the reliability and review burden created by AI-assisted workflows.

Those measures are harder to collect than usage totals, but they better answer the question executives and employees actually need answered: did AI make the organization more effective?

Has Meta abandoned AI-driven workforce redesign?

Not based on the available evidence. Meta stepped back from Project OT’s aggressive timetable and scope, but its leadership has not abandoned the goal of making the company more AI native. Zuckerberg said agentic AI development had been progressing more slowly than he hoped, while predicting more meaningful benefits from Meta’s AI investments within three to six months. 17

That forecast should be treated as a projection, not proof that agents will soon be ready to replace teams. The next test is whether AI can produce sustained improvements in output and reliability under real operating conditions—and whether those gains reduce, rather than shift, the human work required to supervise, correct, and recover from automated systems.

The broader lesson

Project OT illustrates the gap between adopting AI tools and reorganizing a company around AI. Agents can be valuable without being dependable substitutes for a large share of a workforce. Before tying staffing decisions to automation, leaders need evidence of consistent outcomes, robust safeguards, and an operating model employees can understand and trust. 117