Meta's ambitious effort to replace large portions of its workforce with artificial intelligence collapsed months into the project, according to a sweeping Reuters investigation published Wednesday that draws on internal documents and interviews. The initiative, run under the codename "Project OT," had planned to shrink many Meta teams by up to 60 percent, with remaining employees organized into small, so-called talent-dense groups supervising fleets of virtual AI workers.
The plan's most dramatic moment came on the evening of May 19, when CEO Mark Zuckerberg halted a second wave of layoffs that had been scheduled for November — just hours before the first round was set to proceed, Reuters reported. By that point, three separate forces had converged to derail one of the tech industry's boldest automation experiments: underwhelming agent technology, restless investors, and an openly rebellious workforce. For more context on this story, see our ongoing breaking AI news.
For ongoing reporting on how enterprises are navigating AI adoption amid these failures, follow our latest AI developments.
What Project OT Actually Planned
Internal documents reviewed by Reuters showed the scope of the program exceeded what Meta had publicly acknowledged in previous rounds of restructuring. Instead of simply trimming layers of management, the company intended to convert much of its payroll into a human-AI hybrid model in which small groups of senior staff would supervise multiple autonomous agents performing the routine work previously assigned to larger teams.
The reductions were aggressive even by the standards of Big Tech efficiency drives. Teams facing cuts of up to 60 percent were told their functions would be absorbed partly by AI systems capable of writing code, handling customer-facing workflows, and operating internal tools with minimal supervision.
Where It Fell Apart
Three failure modes stand out from the investigation.
The technology wasn't ready
The AI agents Meta deployed never delivered the productivity gains the plan assumed. Ars Technica, summarizing the investigation, reported that the agents at times made "large-scale, disruptive actions" rather than reliable contributions — behavior that converts promised labor savings into cleanup work for the very humans meant to be replaced.
In July, Zuckerberg acknowledged the shortfall, admitting the agent technology had not sped up as fast as he expected, according to the reporting. That concession marked a notable retreat from earlier public enthusiasm about AI-supervised workforces.
Investors lost patience with the bill
Rather than reassuring markets, the massive AI budget drew criticism from investors who questioned whether spending on this scale could be justified while revenues remained tied to advertising. The investigation describes growing tension between the superintelligence-scale investment narrative and the quarterly expectation of a company whose core business still funds everything else.
Employees saw exactly what was coming
Perhaps most corrosively, trust inside the company collapsed. When employees came to believe that tracking software logging their mouse clicks and keystrokes was training their own AI replacements, internal message boards filled with angry posts, Reuters reported. Internal sentiment dropped from 74 percent to 55 percent, a slide rapid enough to register as an institutional crisis rather than routine grumbling.
The surveillance dimension turned an abstract corporate transformation into something workers experienced personally. Mouse-click logs that might read as innocuous productivity metrics elsewhere became, in Meta's context, training data for a digital twin expected to replace its source.
Why This Story Matters Beyond Meta
Project OT is the clearest public case study yet of the gap between AI workforce substitution as strategy deck and AI workforce substitution as operational reality. The pattern it exposes will look familiar to any enterprise attempting agentic deployments today: models excel at demonstrations, wobble on long-horizon autonomy, and create supervision overhead that can erase theoretical savings.
It also raises governance questions that reach past shareholder returns. If keystroke-level monitoring is framed as improving products but actually trains replacement agents, companies face a transparency problem with legal implications in several jurisdictions. Works councils in Europe, for example, have already challenged data collection practices far milder than continuous behavioral logging.
And there is a competitive irony worth noting. As reported this week, Meta simultaneously scales investments like custom silicon and new consumer-facing agent platforms, betting billions on hardware and product AI even as its flagship automation-of-work experiment stalls. The company appears to be continuing its AI buildout everywhere except the place where the hardest proof point sits — inside its own org chart.
The Limits of 'Talent-Dense'
Zuckerberg's vision of lean, talent-dense teams supervising agent swarms has circulated in Silicon Valley management circles for two years now. Project OT's implosion suggests the model assumes away most of the difficulty: coordination costs don't vanish because they moved into software, and the judgment required to correct a misfiring agent is precisely the expensive expertise organizations are tempted to cut.
Companies watching this saga should take note of which parts failed — the technology, the cost-benefit case, or the people — because all three interacted. A cheaper agent would not have repaired broken trust any more than better communication would have fixed unreliable code.
Meta did not immediately respond to detailed requests for comment on the report's specific findings, including the November layoff cancellation and the sentiment figures.
Keep Up With the AI Stories That Matter
From boardroom reversals to breakthrough research, our newsroom tracks every major development across the industry so you don't have to chase rumors.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
