Project OT: Meta's Secret Plan to Cut Teams 60% With AI — and the Internal Data That Killed It

The most consequential AI story of the week isn’t a model release. It’s an autopsy.

On August 26, Reuters published a special report drawing on internal Meta documents and interviews, reconstructing how Mark Zuckerberg’s boldest workforce gamble — a plan codenamed Project OT, short for Organization Transformation — was drawn up, defended, and ultimately abandoned. The plan’s ambition was simple and staggering: make Meta “AI native” by slashing the size of many teams across the company by as much as 60%, in two waves, with AI agents absorbing the routine work of the departed humans.

It didn’t survive contact with reality. And the internal data explaining why it failed is the clearest quantitative picture yet of what happens when a giant tech company tries to convert AI coding hype directly into headcount cuts.

The plan: delete the org chart, not just trim it

According to Reuters, Project OT was hatched at Zuckerberg’s Hawaii compound in January 2026. The idea went beyond giving every Meta employee an AI assistant. Internal documents envisioned smaller “AI-assisted pods,” layers of middle management deleted, and product development increasingly run with agent labor. In February, Zuckerberg justified an initial 40% cut to some staff by saying AI “fundamentally changes what it means to build and run a company.”

The structural half of the plan was the more radical part. As the Times of India’s summary put it, Project OT “was designed to delete Meta’s org chart, not just trim headcount.” Scenarios modeled some teams shrinking by up to 60% across two waves, with the remaining work parceled out to AI agents and a smaller human core.

What actually happened is now well documented. In May 2026, Meta laid off 8,000 employees — about 10% of its workforce — and transferred another 7,000 staff into AI-related roles, together affecting roughly 20% of employees. Zuckerberg initially framed the cuts as a consequence of heavy capital expenditures, with AI spending racing toward a reported $145 billion. But the far deeper second wave, the one that would have taken many teams toward that 60% figure, never came.

The data: more code, fewer features, more fires

The heart of the Reuters report is a set of internal metrics that Meta gathered after its AI push began. The numbers deserve to be quoted precisely:

• Code changes to internal platforms and infrastructure were up 220% year-over-year. AI tools were generating vastly more code churn.
• But changes resulting in new or improved features reaching users rose only 36%. Output volume and user-facing value diverged dramatically.
• Major incidents spiked 40%. More changes meant more things breaking.
• Time spent firefighting those incidents ballooned. Engineers were increasingly consumed by cleanup rather than creation.

Ars Technica, citing the report, highlighted another detail: internal posts reportedly pointed to AI agents making “large-scale disruptive actions” — autonomous changes significant enough that employees flagged them internally as destabilizing.

This is the productivity paradox of AI coding agents rendered in production telemetry. The 220% surge in code changes is exactly what an aggressive agent rollout looks like from the inside: agents iterate fast, generate diffs continuously, and touch systems relentlessly. But diffs are not features. When feature delivery grows at 36% while change volume grows at 220%, roughly five-sixths of the additional churn never reaches users as value. And when major incidents rise 40% in the same window, a chunk of the remaining human capacity gets diverted from building to repairing.

The revolt: engineers aren’t passive infrastructure

Data wasn’t the only force that killed the second wave. Reuters describes a workforce in open revolt. Employees, convinced the AI transformation initiatives were partly designed to replace them, pushed back hard — The Pragmatic Engineer’s newsletter characterized it as employees “in open revolt” against transformation tasks. Morale had already been described as collapsing through the spring: Reddit threads from April documented “4-6 weeks of zero morale” between layoff announcements and executions, with employees describing the period as “28 days of hell.”

Zuckerberg’s own posture shifted. In June, according to a separate Reuters report, he told employees in an internal memo that the company had made “mistakes” in its AI workforce shift. By August, the second wave of cuts had been cancelled, with reports indicating Zuckerberg halted planned November layoffs after both the employee rebellion and the failure of AI agents to deliver the anticipated productivity gains.

There is a bitter irony in the sequencing. The agents failed the productivity test at the very moment they were producing their most impressive-looking raw statistics. If Meta had measured success by lines changed, Project OT would have looked like a triumph. It measured what mattered instead — features shipped, incidents, engineer time — and the plan collapsed under the evidence.

Why this matters beyond Meta

Three lessons generalize.

First, “AI native” is an organizational claim, not a tooling claim. Meta didn’t just give everyone Copilot. It attempted to re-architect the company around the assumption that agent output could substitute for staff at a 60% clip. That assumption ran ahead of what the agents could reliably do — and the gap showed up in incident dashboards, not in demos.

Second, code-generation metrics are the wrong KPI. The 220%/36% divergence is a warning to every company building an AI strategy around “developer velocity” dashboards. Volume metrics flatter agents; outcome metrics expose them. Any enterprise AI rollout should be judged on features reaching users, incident rates, and maintenance burden — precisely the telemetry that sank Project OT’s second wave.

Third, workforce consent is a real constraint. A plan that employees understand as replacement will meet resistance that degrades its own feasibility — the people expected to train and supervise the replacing agents are the people being replaced. Meta’s engineers were simultaneously asked to make the agents work and to accept that the agents’ success meant their departure. That structure is unstable, and it broke.

What happens next

The cancellation of the second wave is a pause, not an epilogue. Meta confirmed Project OT as a year-long effort focused on costs, team structure, moving employees into priority AI work, and training. The $145 billion AI capex trajectory is unchanged, and the company still holds the structural option to revisit cuts as models improve. The Grind’s coverage notes the November cuts were cancelled, not the underlying logic.

But the burden of proof has shifted. For a year, the industry’s default assumption has been that AI capability curves translate more or less directly into staffing curves — that what works in a coding benchmark scales to an org chart. Project OT is the largest, best-documented counterexample so far: a company with elite engineering talent, near-unlimited compute, and full executive commitment ran the experiment at scale, gathered the internal data, and pulled back.

When the history of enterprise AI adoption is written, the 220%-code-changes-versus-36%-features number deserves a place next to the productivity statistics that defined earlier technology waves. It is the moment the replacement hypothesis met the telemetry — and the telemetry won.

Until the next model generation, every board planning “AI-native” restructuring now has a case study with numbers attached. Meta tried to delete its org chart. The org chart, aided by 40% more major incidents, deleted the plan instead.