Project OT: How Meta's Plan to Replace Thousands of Workers With AI Collapsed From the Inside

On August 26, Reuters published an investigation that reads like a corporate thriller — except every detail comes from internal Meta documents and named events the company lived through this year. Under the codename “Project OT” (short for “organizational transformation”), Meta quietly drew up plans to shrink many of its teams by as much as 60 percent, hand the displaced work to AI agents supervised by small “talent-dense” crews of humans, and restructure the entire company around what management called an “AI-native” philosophy. Then, on the evening of May 19 — just hours before the first layoff wave — Mark Zuckerberg personally cancelled the second, larger wave that had been planned for November.

The plan didn’t die because of regulators or public backlash. It collapsed under the weight of its own internal data: AI agents that couldn’t do the work, employees in open revolt, and productivity metrics that refused to cooperate.

Where Project OT came from

The story starts in 2025, when Meta executives — including Chief Data Officer Alex Schultz and Head of Product Naomi Gleight — toured Asia and came back impressed by how AI startups there built entire organizations around AI agents rather than merely deploying them as tools. The executives commissioned internal research into those organizational structures and launched a pilot to test whether a “talent-dense” AI-native company was viable at Meta’s scale.

In January 2026, Zuckerberg and a small circle of confidants retreated to his compound in Hawaii to refine what became Project OT. According to documents reviewed by Reuters, the scenarios they explored included AI taking over “much of the daily work performed by thousands of human employees,” with many internal teams reduced by up to 60 percent. The restructuring was designed to run in two phases: a first layoff round in May 2026, and a second, larger one in November. One HR executive predicted total reductions would exceed 25 percent of the workforce.

In parallel, Meta had been running a product-side experiment since July 2025: “small technology pods” of two to three engineers plus one designer replacing traditional teams of 10–20. An internal “AI-Native Playbook” abolished conventional titles like product designer and engineer in favor of a single role — “Builder” — overseen by one “Pod Leader.” By June 2026, at least eleven departments, including core engineering and research teams, had adopted the pod system. But pod leaders were given no formal management authority (no performance evaluations, no manager training), and employees reported widespread confusion about who was actually responsible for what.

The data that killed the plan

What makes the Reuters report so unusual is the precision of the failure metrics. Meta didn’t just feel less productive — it measured the gap.

One internal post tracked code changes across Meta’s internal software platform and infrastructure: total changes were up 220 percent year-over-year, but changes that led to new features or actual upgrades rose only 36 percent. In other words, AI-assisted coding was generating enormous volumes of churn while shipping proportionally little value. In March, Meta’s infrastructure team issued a formal “reliability warning due to increased AI coding.” In April, internal reports flagged “large-scale, disruptive actions by uncontrolled AI agents that humans are unlikely to execute.”

The downstream damage was quantifiable: critical technical and security incidents — service disruptions, potential data breaches — rose 40 percent year-over-year, and the time staff spent resolving problems jumped 70 percent. The company was, quite literally, spending more human hours cleaning up after its AI than it saved.

June delivered the public embarrassment: hackers tricked Meta AI’s own support chatbot into handing over celebrity Instagram accounts, including the White House account of former President Barack Obama and the account of a Space Force Chief Master Sergeant. For a company arguing that AI agents were ready to run its internal operations, a support agent compromised by social engineering was exactly the wrong headline.

The revolt

The human factor was just as decisive. In late April, Meta began collecting employees’ mouse movements and keystrokes as training data for AI. Combined with the leaked layoff reports — Reuters itself reported in March that Meta was considering cuts of up to 20 percent, before many vice presidents had even been briefed on Project OT — employees concluded they were being asked to train their own replacements. Angry and self-deprecating posts flooded Workplace, Meta’s internal communication tool. Memes mocking the company’s AI initiatives appeared on flyers posted in Meta’s US offices.

The numbers captured the collapse in trust: Meta’s employee satisfaction score fell from 74 percent to 55 percent in one semi-annual survey — a catastrophic drop for a company that once topped “best places to work” lists. Investors, meanwhile, openly criticized the enormous AI capital budget funding the transformation.

On the evening of May 19, hours before the first layoff wave, Zuckerberg consulted with aides and called off the November second wave. The May cuts — about 10 percent of the workforce, roughly 8,000 people — went ahead, but remaining employees were told no further company-wide layoffs were planned “this year.” Management then began damage control: pausing the mouse-tracking program, returning some engineers from the “Applied AI Engineering” department (where they had been reassigned to write puzzles for coding-model training data) to their original teams, and posting empathetic messages on Workplace — along with promises of better office snacks and increased budgets for business trips and social events.

“Not as fast as we’d hoped”

Zuckerberg’s own assessment came at an internal July meeting: “At least the progress in AI agent development over the past four months has not accelerated as much as we had hoped.” It’s a striking admission from the CEO who spent 2025 telling investors that 2026 would be the year AI agents would write most of Meta’s code.

Employees have noticed the careful wording, though. When Zuckerberg says “company-wide” and “this year,” some staff read it as leaving the door open to team-by-team cuts — or another large restructuring in 2027. The pod system remains in place. The Applied AI Engineering department still exists, just smaller. And the strategic pressure that created Project OT — massive AI capex, investor impatience, and a competitive race where every lab claims agents are about to replace software engineering — hasn’t gone anywhere.

Why this matters beyond Meta

Project OT is the most detailed public case study yet of what happens when the “AI replaces the workforce” thesis meets an actual workforce. Three lessons stand out.

First, the productivity statistics labs cite and the productivity companies experience are diverging. Meta measured a 220 percent increase in code changes and got only 36 percent more features — plus a 40 percent rise in critical incidents. Any company betting headcount reductions on AI productivity curves should be running the same instrumentation Meta ran, and should expect the same answer until agent reliability improves.

Second, substitution failed where augmentation might have worked. The pod experiment and the internal tooling weren’t inherently bad ideas; the failure mode was designing the organization around replacing people rather than augmenting them — then harvesting the training data from the very employees being sized for replacement. Trust, once spent, shows up directly in satisfaction scores and retention risk.

Third, internal AI risk is now a board-level issue. “Large-scale, disruptive actions by uncontrolled AI agents” inside corporate infrastructure is precisely the class of incident the AI safety community has warned about — and Meta’s incident logs now provide hard evidence that it isn’t hypothetical. The 40 percent jump in critical incidents is the kind of number that should anchor enterprise AI governance policies for years.

Meta hasn’t abandoned its AI-native ambitions — the October “Watermelon” frontier model and the Hatch consumer agent platform are still reportedly on track. But Project OT’s implosion has recalibrated expectations across the industry: the agents aren’t ready to run the company, and the humans noticed first. As one internal document’s trajectory makes clear, the most sophisticated AI deployment organization on Earth ran the experiment at full scale — and blinked.

Based on Reuters’ investigation of August 26, 2026, and subsequent reporting by Ars Technica, The Decoder, and GIGAZINE.