The plan had a name. Project OT, short for Organization Transformation, was set at Mark Zuckerberg’s annual leadership retreat at his Hawaii compound in January. On August 26, Reuters correspondent Katie Paul published the planning documents, built from scores of internal posts and recordings and more than 20 people with knowledge of Meta’s internals.
The design: AI takes over much of the daily work of thousands of employees, and a smaller “talent-dense” layer of humans supervises the virtual workers. In scenario planning, executives modelled cutting many teams by as much as 60%. One HR executive projected a culling as large as or larger than the roughly 25% Meta cut three years ago.
Two waves. May, then November.
On the night of May 19, hours before the first wave went out, Zuckerberg called off the November planning. Meta laid off 10% the next morning anyway.
Meta confirmed Project OT to Reuters, describing a year-long effort covering cost cutting, team redesign, and shifting staff to priority work — including producing training data for its own AI models. It confirmed the two waves and the up-to-60% scenarios for some teams, said it never intended to cut 60% of the company, and said leaders killed the second wave before deciding how many people overall would lose their jobs.
We covered the 4 a.m. email on May 21: 8,000 out, 6,000 open roles cancelled, 7,000 survivors reshelved into units like Applied AI Engineering. On August 1 we priced it, at $147,500 a head against $1.18B of severance. Both articles described execution. What Reuters published is the half of the plan that never executed, and the reason.
The org chart went through, even though the layoffs didn’t
In July last year, VP of product management Ime Archibong announced the first pilot: five “small tech pods,” each two to three engineers plus a designer, all equipped with AI tools. The pods dropped the six-month planning cycle for four-week sprints.
By October, a member of his team posted an “AI-Native Playbook” telling other teams how to follow. It was specific. Traditional product designer and engineer titles disappear; pod members share one title, “builder.” Layers of middle management get removed. Pods report up to a single high-level unit head. Day-to-day priorities come from “agent-assisted analysis.”
By June, at least 11 units had implemented pods, including engineering and research teams. Alongside came a “village approach” to managing people: performance ratings and promotions decided by Org Leads, supported by HR and unspecified “AI systems”; each Org Lead covering 30 to 50 people; Pod Leads running the day-to-day with no formal management authority. One staffer assigned to lead a pod wrote on an internal board that he was not going through manager training and could not access ratings or manager tools.
Meta also stood up an HR tool to flag “Irreplaceable Talent,” with an internal document naming a hypothetical “10X Performer.” Two people familiar with the matter said Meta intended to spend layoff savings on outsized pay packages for AI engineering stars.
None of that was cancelled. Only the second layoff wave was.
Meta measured the reason itself
Three internal numbers stopped the plan.
The first is throughput. CTO Andrew Bosworth posted in early June that employee AI use had driven code changes to Meta’s internal platforms and infrastructure up 220% year over year. Changes that reached users as new or upgraded features were up 36%. Same transformation, two growth rates, a gap of roughly six times.
The second is reliability. Infrastructure teams flagged “reliability warning signs” as early as March. An April post said unchecked AI agents were taking “large-scale, disruptive actions that humans are unlikely to execute.” Major technical and security incidents rose 40% year over year, and staff firefighting time rose 70%. In early June, attackers exploited Meta’s new AI-powered customer support bot to reach high-profile Instagram accounts, including the dormant Obama White House page. That was the only piece the public saw.
The third is people. Meta’s half-year Pulse survey put employee sentiment at 55% favourable, down from 74%. In April, Reuters reported that Meta had mandated tracking software on US employees’ machines to capture keystrokes and mouse clicks for agent training. Staff concluded they were training their own replacements and buried the internal network in angry posts, some replying to executives with pictures of elephants. Labour organising picked up.
Engineers reassigned into Applied AI Engineering were writing software-engineering puzzles to improve the coding skills of Meta’s models. Internal posts called the work rote and boring. After the layoffs, Meta let some of them transfer back and paused the mouse-tracking program.
At a town hall in early July, Zuckerberg conceded he had misjudged the timing. Agent technology, he said, had not “accelerated” as fast as he expected. He still expects more benefit in three to six months.
Deferred, not withdrawn
After killing the second wave, Zuckerberg told staff he did not expect other company-wide layoffs this year, and said he wanted to give people more stability.
Read the two qualifiers he has stuck to: “company-wide” and “this year.” Team-specific cuts are not company-wide. Performance-based dismissals are not company-wide. Moving the action into next year does not violate this year. And Meta’s own statement is that leaders cancelled the November wave before determining how many people would lose their jobs, which is the same sentence as: nobody was spared, a number was left uncalculated.
The cohort comparison is instructive. On August 21, Apple cut 200 roles and promised new ones in the same sentence via leak-then-confirm, leaving nothing in a public filing. Meta used an internal memo with two qualifiers. Different routes, same outcome for anyone trying to audit it later: no verifiable number enters the record.
The financial pressure has not eased either. Meta plans to invest at least $130B in AI chips and infrastructure this year, which LSEG estimates will consume its 2026 operating cash flow. The motive to compress labour cost is intact. Only the fastest route to it has been shown, internally and in writing, not to work yet.
Who this actually lands on
Middle managers. The layer Project OT removed was not engineers. It was management. Pods report to Org Leads, Pod Leads carry the responsibility without the authority, and one Org Lead covers 30 to 50 people. That structure is the core of the AI-Native Playbook, it was not cancelled, and it is running in at least 11 units. If your role’s value rests mainly on “I run a team,” in this kind of reorganisation your layer goes before your work does.
Titled product roles. Product designer, engineer, and adjacent titles collapse into “builder” inside a pod. Title collapse is not paperwork. It resets promotion ladders, evaluation criteria, and the portable job title you carry out the door. Ask about it in interviews: is the level system organised by function, or by pod? The second answer means the résumé you take to your next employer in three years carries no title the industry recognises.
Anyone redeployed onto training data. Meta confirmed on the record that part of Project OT was shifting staff into producing training data for its models. That is a live internal path: you do not get cut, you get moved from a product role onto a role that writes exercises for a model. China industrialised the same step into dedicated facilities; Meta ran it inside its own org chart. The test is simple. Does your output ship to users, or to a model? If it ships to a model, the life of your role is set by when the model learns.
Everyone currently being handed AI tools. The 220%-versus-36% pair is one of the very few agentic-coding productivity measurements taken inside a large company and seen outside one. It does not say AI is useless. It says there is a large gap between output volume and shipped value, and that gap is usually unmeasured at the moment the layoff decision gets made. If your employer is citing AI efficiency as the reason for cuts, the question worth asking is which end of the pipeline got faster, and whether anyone measured downstream.
Meta took seven months to answer that question about itself. The answer was to put the second knife back.
