Project OT was the codename Meta gave to the most ambitious workforce experiment any large technology company has attempted this decade, and according to a Reuters special report published on 26 August 2026, it did not survive contact with reality. Project OT — short for Organization Transformation — was designed to hand much of the daily work of thousands of employees to software, leaving small, “talent-dense” groups of humans to supervise the machines.

The plan was drawn up at Mark Zuckerberg’s Hawaii compound in January 2026 and, in its most aggressive scenario-planning form, explored shrinking some teams by as much as 60 per cent. One wave of cuts went ahead in May. A second, larger wave scheduled for November did not. On the evening of 19 May, hours before the first round of notices went out, Zuckerberg reportedly cancelled it.

What makes Project OT worth reading about is not that a big company changed its mind. It is why. Reuters describes three forces converging: agent software that produced enormous activity without matching output, investors who had grown impatient with a capital budget approaching $130 billion, and a workforce that revolted when it concluded the monitoring software on its laptops was training its own replacement. We have written before about Bill Gates’ warning that this would be a turbulent era; this is what turbulence looks like from inside a payroll of roughly 80,000 people.

What Project OT Actually Was

project ot meta ai restructuring layoffs b office swivel chair

Meta has never published a description of the programme. Everything below comes from Reuters reporting based on internal planning documents and interviews, plus the company’s own statement responding to it.

The retreat where the plan began

In January 2026, Zuckerberg and his senior leadership gathered for their annual offsite at his estate in Hawaii. The output was an “AI native” vision for the company: a Meta in which software agents did the routine work of engineering, design, marketing operations and programme management, and humans did the parts that agents could not.

The scenario that alarmed people

Scenario planning is not the same as a decision, and Meta has been careful to say so. But the scenarios were dramatic. Executives modelled cutting the size of many teams by up to 60 per cent through a mix of layoffs and redeployment. Reuters reports that some engineering units were on track for headcount drops of around 30 per cent by the end of May alone.

Pods, villages and builders

The organisational half of Project OT was as radical as the headcount half. Traditional teams were to be replaced by “pods” of two or three engineers and designers. Unit heads, retitled “Org Leads”, would each oversee 30 to 50 people under a “village approach”, while Pod Leads steered day-to-day work without formal management authority. Job titles collapsed into a single generic label: “builder”.

The two-wave structure

Project OT was sequenced deliberately. A first wave in May would take out roughly 10 per cent of staff and prove the model. A second wave in November would follow once the agents had absorbed the work. The second wave was the part that never happened.

Project OT at a glance

ItemDetail
CodenameProject OT (Organization Transformation)
OriginLeadership retreat, Hawaii, January 2026
Premise“AI native” company; agents do routine work
Maximum team cut modelled60 per cent
Wave one20 May 2026 — went ahead
Wave twoNovember 2026 — cancelled 19 May
New team unit“Pod” of two to three people
New generic title“Builder”
2026 AI infrastructure budgetAt least $130 billion
SourceReuters special report, 26 August 2026

Who was in the room

Reuters names a familiar senior cast around the programme, including chief technology officer Andrew Bosworth, chief data officer Alex Schultz, head of product Naomi Gleit, vice-president of product management Ime Archibong and chief financial officer Susan Li. That matters because it places Project OT in the operating core of the company rather than in a strategy function. This was not a think-piece; it was a plan with owners.

How the story surfaced

Almost nothing about Project OT was disclosed at the time. It emerged in fragments across five months of reporting, and only the August special report assembled the fragments into a programme with a name. The sequence is worth laying out, because it shows how far the plan had already travelled before anyone outside Meta could see its shape.

DateWhat happened
January 2026Leadership retreat in Hawaii; the “AI native” plan takes shape
13 March 2026Reuters reports Meta is weighing cuts of 20 per cent or more
April 2026Keystroke and click tracking reported on US employee devices
23 April 2026Staff are told roughly 10 per cent of jobs will go
19 May 2026The November wave is cancelled, hours before wave one lands
20 May 2026Around 8,000 people are laid off
June 2026At least 11 units have adopted the pod structure
July 2026Zuckerberg tells a town hall that agents progressed slower than hoped
26 August 2026The special report describing Project OT is published

What Meta says about it

The company did not deny that Project OT existed. Its statement acknowledged the exercise and framed the outcome as a redeployment success: “This ultimately resulted in moving thousands of employees to do priority work on several newly-established teams… Ultimately, we didn’t move forward with every scenario from the exercise – and it was never assumed we would.”

Why Meta Abandoned the Second Wave of Project OT

project ot meta ai restructuring layoffs c three stacked office trays

The cancellation was not announced. It was inferred by Reuters from internal documents and from the fact that November came and went without the cuts Project OT had scheduled. Three explanations run through the reporting, and they reinforce each other.

The agents did not deliver

This is the load-bearing reason. Meta’s internal AI push generated a spectacular amount of activity. Code changes to the company’s AI platforms and infrastructure rose 220 per cent year over year. But new or improved features that actually reached users rose only 36 per cent. Activity went up sixfold relative to outcomes.

Worse, the activity created work

The same period saw technical and cybersecurity incidents spike 40 per cent, and the time employees spent resolving them rose 70 per cent. Agents were not just failing to replace human effort; in the aggregate they were consuming it. Any organisation that has run an intelligent automation programme will recognise the shape of this: throughput measured at the wrong layer looks like progress right up until the incident queue tells you otherwise.

The activity-versus-outcome gap, charted

Reuters gives four internal year-over-year figures. Read together they are the clearest argument against the second wave of Project OT that anyone at Meta could have made.

Meta internal metrics, year-over-year change (2026)
Code changes to internal platforms +220%
Time spent resolving incidents +70%
Technical and security incidents +40%
Improvements reaching users +36%
Bars scaled to the largest figure (+220%). Source: Reuters, 26 August 2026.

Zuckerberg conceded the point in July

At a town hall in July, Zuckerberg told staff that agent technology had not accelerated as quickly as anticipated. That was a public admission, months after the private decision, that the premise underneath Project OT had slipped.

Investors had stopped applauding

The capital side mattered too. Meta committed at least $130 billion to AI chips and infrastructure for 2026, and that number attracted the kind of scrutiny that follows any very large bet made on a promise. It is the same question being asked across the sector, and one we examined when Cornell researchers looked at whether there is an AI bubble. Announcing a second mass layoff to fund a programme that was not yet producing measurable output would have invited an uncomfortable comparison.

The Employee Revolt That Undermined Project OT

project ot meta ai restructuring layoffs d upright megaphone cone

The third force was the one the planning documents were least equipped to model. Reuters reported in April that Meta had installed software on US employees’ devices to capture keystrokes and mouse clicks. The stated purpose was to teach agents how humans use a computer.

Staff drew the obvious conclusion

Employees decided the tracking was training their replacements, and said so. Meta’s internal network filled with angry posts, jokes, and images of elephants — the elephant in the room being the layoffs nobody would name. The monitoring programme was later paused.

The sentiment number moved

Meta measures staff sentiment in a half-year Pulse survey. Favourable responses fell from 74 per cent to 55 per cent, a 19-point drop. For a company that treats internal metrics as management instruments, a swing that size is not a mood; it is a signal.

Meta Pulse survey — favourable employee sentiment
Before the restructuring 74%
After the tracking disclosure 55%
A 19-point fall in one half-year cycle. Source: Reuters, 26 August 2026.

The pod model confused its own managers

Structure changes landed before the tooling that was supposed to support them. One newly appointed pod manager posted internally: “I’m not going through manager training, and I’m not getting access to ratings & manager tools.” Reuters also reported that performance decisions were to involve “AI systems”, a description Meta later clarified by saying people, not software, make those calls.

Why this matters more than it looks

Surveillance framed as data collection is a change management failure before it is an ethics problem. The company needed its staff to demonstrate their work so the agents could learn it, and simultaneously needed those staff to believe their jobs were safe. Project OT asked for both at once, and got neither.

What Project OT Did Change at Meta

project ot meta ai restructuring layoffs e solid briefcase with handle

Cancelling the second wave did not undo the first. A great deal of Project OT was executed, and the company Meta runs today is materially different from the one that met in Hawaii.

The cuts that did happen

On 20 May 2026, Meta laid off roughly 8,000 people — about 10 per cent of the workforce — and closed a further 6,000 open roles. Reporting at the time put the combined effect, including internal moves, at something closer to a fifth of the company being reshuffled in a single quarter.

Thousands were moved, not removed

Meta’s own framing emphasises redeployment, and the numbers support it. Roughly 7,000 workers were redirected into newly created AI-focused teams rather than dismissed. In at least one case Reuters describes, engineers in the Applied AI Engineering unit were reassigned to produce training data — building the corpus the agents needed rather than being displaced by them.

The structure stuck

By June, at least 11 units had implemented the pod structure. The village model, the Org Lead span of 30 to 50 people, and the “builder” title survived the cancellation of wave two. Project OT failed as a headcount plan while succeeding, at least partially, as an org-design plan.

Planned versus delivered

ElementPlannedDelivered
Wave one layoffs~10% of staffYes — ~8,000 roles, 20 May
Wave two layoffsNovember 2026No — cancelled 19 May
Team reductions up to 60%Modelled in scenariosPartial — ~30% in some units
Open roles closedIncluded in the planYes — ~6,000
Redeployment to AI teamsCore to the designYes — ~7,000 people
Pod and village structureCompany-wideYes — 11+ units by June
Keystroke and click trackingAgent training inputDeployed, then paused
Further 2026 layoffsExpected under the planRuled out publicly by Zuckerberg

The promise of stability

After the May notices went out, Zuckerberg told employees he did not expect further company-wide layoffs that year, and that he wanted to give the team more stability. Whatever else Project OT achieved, it made that sentence necessary.

Workforce actions at Meta, May 2026 (roles)
Employees laid off 8,000
Moved into new teams 7,000
Open roles closed 6,000
Bars scaled to the largest figure (8,000 roles). Figures are approximate.

Project OT and the Wider AI Replacement Story

project ot meta ai restructuring layoffs f three descending solid blocks

Meta is not the first company to promise that software would absorb headcount, and it will not be the last. What separates Project OT is the scale of the bet and the quality of the paper trail that survived it.

The pattern is now familiar

Firms across sectors have announced hiring freezes and role reductions justified by autonomous AI agents, then quietly rehired or redeployed when the agents proved better at drafting than at owning outcomes. The failure mode is consistent: agents handle bounded tasks well and unbounded responsibility badly, and most jobs are mostly unbounded responsibility.

Where the case for Project OT was weakest

The plan assumed capability would arrive on a schedule. Capability arrives when it arrives. Committing a headcount reduction in November to a capability that had to exist by October is a sequencing error, and it is the specific error Project OT made.

How the reasoning compares

AssumptionWhat the plan expectedWhat the data showed
Agent outputSubstitutes for human work+220% activity, +36% shipped
Operational loadFalls as agents take overIncidents +40%, firefighting +70%
Staff reactionManageable with messagingSentiment 74% to 55%
Monitoring dataNeutral training inputRead as replacement evidence
Small podsFaster decisionsManagers without tools or training
Capital marketsReward the efficiency storyQuestioned the $130bn budget

The reporting caveat

None of this is a filing. Meta has confirmed the exercise existed and disputed the framing of parts of it. The figures come from documents Reuters reviewed and people it interviewed, and the company’s position is that scenario planning was never a commitment. Treat Project OT as very well sourced, not as company disclosure.

What Project OT Teaches About AI Workforce Planning

Most organisations will never run anything at this scale. The lessons still transfer, because the mistakes were structural rather than specific to Meta.

Measure shipped outcomes, not activity

A 220 per cent rise in code changes against a 36 per cent rise in delivered improvements is the entire story in two numbers. If your automation dashboard counts actions, tickets touched or drafts produced, it will show success during a period when nothing is actually reaching a customer.

Watch the incident queue

Rising incidents and rising remediation time are the leading indicators that automation is being deployed faster than it is being governed. They moved before anything else at Meta did, and they moved in the wrong direction.

Sequence capability before commitment

Cut headcount after the capability is demonstrated in production, not on the schedule you hope it will arrive. Project OT inverted that order and had to reverse itself in public.

Never make people train their own replacement in the dark

If you collect behavioural data to build automation, say what it is for, say what happens to the roles it touches, and accept that ambiguity will be resolved pessimistically. Meta’s sentiment score is what unexplained monitoring costs.

A short pre-flight checklist

Question to answer firstWhy it matters
What outcome metric moves?Activity metrics hide failure
Who owns an agent’s mistakes?Unowned errors become incidents
Is the capability in production?Roadmaps are not capacity
Have you told staff the purpose?Silence is read as the worst case
Do new managers have tools?Structure without tooling stalls
What is the reversal plan?Meta needed one and used it

Project OT: What Happens Next

The second wave is cancelled, not deleted. Meta has kept the structure that Project OT introduced, kept the capital plan, and kept the ambition. What it has given up, for now, is the timetable.

Three things worth watching

The next Pulse survey will show whether 55 per cent was a floor or a trend. The next capital-expenditure guidance will show whether $130 billion was a peak. And the ratio that killed the second wave of Project OT — activity to shipped output — is the one number that decides whether the plan ever comes back.

The honest read

Zuckerberg was early rather than wrong about direction, and wrong rather than early about pace. That distinction is doing a lot of work, and it will keep doing it across every company that writes an AI headcount plan in the next two years.

Project OT: Frequently Asked Questions

Did Meta officially announce Project OT?

No. The programme was described in a Reuters special report on 26 August 2026 based on internal documents. Meta responded to the reporting and confirmed the exercise took place, but has not published its own account.

Were the May 2026 layoffs part of Project OT?

Yes. The roughly 8,000 roles cut on 20 May, and the 6,000 open roles closed alongside them, were wave one. Wave two, planned for November, is the part that was abandoned.

How many people would the full plan have affected?

Scenario planning explored reductions of up to 60 per cent in some teams. Reuters describes the projected total as comparable to or larger than the 25 per cent reduction Meta made in 2022 and 2023.

Why did the agent strategy fail?

It produced activity rather than delivery. Code changes to internal AI platforms rose 220 per cent year over year while improvements reaching users rose 36 per cent, and incidents and remediation time rose sharply at the same time.

Is the tracking software still running?

Reuters reports the keystroke and mouse-click monitoring on US employee devices was paused after the internal backlash. Meta has said people, not software, make performance decisions.

Could the second wave return?

Nothing in the reporting rules it out. Zuckerberg said he did not expect further company-wide layoffs in 2026, which is a statement about a year, not a strategy.

References