Meta just showed what happens when you tie layoffs to an AI productivity curve
Project OT, Meta's codename for an 'AI native' restructuring that would have shrunk some teams by 60% and replaced the lost work with agents, was cancelled in May under combined technical, workforce, and investor pressure. The internal productivity signal never arrived. A federal complaint now puts the selection model under discovery. Together they form the cleanest available test case for what an AI-native labor substitution plan looks like once it collides with measured performance, with disability and family-leave law, and with the workforce's own willingness to be the training set.
The AI-native company is not a product roadmap. It is a labor plan.
On August 26, Reuters published a special report based on what it described as "scores of internal documents, posts and recordings" and conversations with "more than 20 people" inside Meta, reconstructing the rise and partial collapse of a restructuring called Project OT — short for *Organization Transformation*. Hatched at Mark Zuckerberg's January leadership retreat in Hawaii, the plan explored scenarios that would shrink some Meta teams by as much as 60% and have AI agents perform "much of the daily work performed by thousands of human employees." It was to be executed in two waves: a roughly 10% cut on May 20, and a second wave in November.
The first wave went out as scheduled. The second was cancelled hours before the notifications were to be sent. Reuters attributes the reversal to three pressures acting together: internal data showing agentic systems were not delivering the productivity gains Meta had budgeted against, employee rebellion triggered by a keystroke- and activity-monitoring program that staff concluded was being used to train their replacements, and investor scrutiny of a 2026 AI infrastructure spend that LSEG estimates at $130 billion or more.
Six weeks before Reuters published, a separate story had already moved into a federal courtroom. On July 13, 26 anonymous plaintiffs filed a 71-page complaint in the Northern District of California alleging that Meta used "a constellation of internal artificial-intelligence systems" — including a system called *Metamate*, employee-trained "second-brain" agents, keystroke- and activity-monitoring data, AI-token-usage dashboards, and adoption ratings labeled "AI Native," "AI First," and "AI Enabled" — to score, rank, and select roughly 8,000 employees for termination. Reuters called the complaint "the first against a major US company to challenge the alleged use of AI in conducting layoffs." Meta told Ars Technica and the Guardian that "Workforce management and organizational decisions were and are made by people, not AI."
The Reuters report and the court complaint are usually treated as parallel stories — one about productivity, one about civil rights. They are not parallel. They are the same story, and the same internal data set is now material to both. The legal discovery process will likely force Meta to disclose, in some form, how the May termination list was actually generated. The productivity data is already on the public record through Reuters' reporting. Together they form the cleanest available test case for what an "AI-native" labor substitution plan looks like once it collides with measured performance, with disability and family-leave law, and with the workforce's own willingness to be the training set.
The plan, in one sentence
Project OT was an attempt to tie a public headcount commitment to an internal agent capability curve, and then miss the curve without missing the commitment.
The structure matters more than the numbers. The 60% ceiling was a *scenario*, not a target. Two people familiar with the planning told Reuters that executives explored headcount reductions "as much as" 60% in some teams and as little as 25% in others; Meta confirmed the ceiling but said it never intended to lay off 60% of the company. The HR executive's projection of a culling "as big as or bigger than the company's cuts of around 25% three years ago" was also a scenario, not a decision. Reuters was unable to determine what the second wave's actual headcount target would have been, because the plan was cancelled before Meta leaders "determined how many people overall would lose their jobs."
What is on the record is the *conditional structure*. The plan was to be carried out in two waves separated by about six months, with the second wave scheduled to begin the night of November 19. The implicit assumption was that agentic productivity would advance enough during those six months to absorb the work of the second cut. That assumption did not hold. By the time the May 20 wave went out, internal data was already pointing the other way.
The productivity signal that did not arrive
The most concrete numbers in the Reuters report come from an early-June internal post by CTO Andrew Bosworth.
- Code changes to internal software platforms and infrastructure were up 220% year-over-year.
- Changes that led to new or upgraded features reaching Meta users were up only 36%.
- Major technical and security incidents — service disruptions, possible data leaks — were up 40% from the prior year.
- Employee time spent "firefighting" those incidents was up as much as 70%.
A separate April post said unchecked AI agents were performing "large-scale, disruptive actions that humans are unlikely to execute." A March post from infrastructure teams warned of "reliability warning signs" caused by the AI coding surge. Meta declined to comment on any of these internal posts.
Read them together, the data does not show that AI is failing at Meta. It shows that more code is being written, that a much smaller fraction of it is reaching users, and that the operational cost of the AI-generated code is being absorbed by humans. The ratio that matters is not 220% over 36%. It is the missing 184 percentage points, which is the volume of code that did not become a feature. Some of that gap is expected churn in any large engineering organization. Some of it is the cost of an under-validated agent. The 40% rise in major incidents and the 70% rise in firefighting time are the part of the gap that is billable in outage minutes and security exposure.
Reuters also reported that as early as March, infrastructure teams were warning about reliability; in April, agent disruptions were characterized as "large-scale"; in June, Bosworth himself posted the 220%/36% numbers. The Project OT second wave was scheduled for November. The internal data was trending against the plan four months before the plan was scheduled to be executed, and at no point during those four months did a public correction appear.
A public correction did appear, in a different form. On July 2, Zuckerberg conceded at an internal town hall that "the trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected." He added that he expected improvement "in the next three to six months." That is the same planning horizon as the cancelled November wave, restated as a forecast rather than a commitment. The Reuters reporting does not connect the two dates directly, but the structural overlap is the story: the second wave was cancelled in May because the data did not support it; the public admission of slower-than-expected agent progress came in July; the company is now publicly promising the productivity that the cancelled plan assumed.
The other pressure: the workforce as training set
Reuters attributes the May reversal to three forces. The productivity data is the one the AI industry most wants to talk about. The workforce rebellion is the one the AI industry most wants to talk around.
Meta had mandated a tracking program that captured employees' keystrokes, mouse activity, browser history, messages, emails, and location data on company devices. Reuters reported the program in April; Wired reported on June 22 that the company had left potentially sensitive data from the initiative exposed internally; on June 24 the Guardian reported the pause, and the program was wound down after more than 1,600 employees signed a petition saying the tool "raises serious concerns around privacy, consent, and trust in the workplace." The complaint later filed in federal court described the program as having been "quietly launched" through "a low-visibility internal post — made by an engineer rather than a senior leader," and that "on at least some teams, employees received no consent or acknowledgment prompt at all, and, at least initially, there was no way to opt out."
Employee sentiment, measured in Meta's half-year Pulse survey, dropped from 74% favorable to 55% favorable. The numeric gap is 19 percentage points. The structural point is that the drop occurred *before* the layoff, in response to the combination of the tracking program and the anticipated cuts, not after the cuts had been delivered. Morale was the cost of the *announcement*, not the cost of the *event*.
That timing matters because it changes what a CHRO planning a similar substitution should model. A typical 2024 or early-2025 layoff assumption was that morale damage is a function of severance, communication, and the post-layoff environment. The Project OT case suggests that when the labor substitution is paired with a workforce monitoring program, the morale damage is a function of the *signaling* — that the company intends to learn from the workforce in order to reduce the workforce. The first wave's communications, the tracking program's rollout, and the Pulse sentiment collapse all happened before any employee lost a job. That is a different order of operations than a conventional restructuring.
The complaint's allegation that the same monitoring data was material to the layoff selection list is the legal articulation of the workforce's reading. Meta's public position is that the list was assembled by people. The complaint's allegation is that the list was assembled by a system in which a constellation of AI inputs — including the monitoring data and the AI-token-usage dashboards — was the dominant signal. The truth of either claim is not established by either document. It is the subject of the pending litigation, and the discovery process in that litigation will likely force a fuller record than either Reuters' reporting or the company's public statements.
The legal seam: a 71-page discovery request
The complaint does four things that the Reuters investigation does not.
First, it names the system. The Reuters reporting describes "AI systems" and "agent-assisted analysis" without naming specific tools. The complaint names *Metamate* (an internal assistant), "second-brain" agents trained on individual employees' work, AI-token-usage dashboards, and the "AI Native," "AI First," and "AI Enabled" adoption categorizations. Whether the names survive the legal process is a separate question; on the public record, they are the named components of the selection model.
Second, it makes a specific civil-rights claim. The plaintiffs are 26 current or former Meta employees in California, Illinois, Washington, New York, the District of Columbia, Pennsylvania, and Florida. They allege that the selection system scored employees on inputs — performance ratings, calibration scores, productivity, AI-native ratings, AI-token consumption — that "by design, cannot be accumulated by an employee who is on protected medical or family leave, or whose output is reduced by a disability." The complaint alleges that Meta "did not neutralize those inputs for protected leave; did not exclude protected-leave-takers or accommodation-seekers from the selection cohort; and did not pause the system for the individualized, leave- and accommodation-neutral review that the law requires." The complaint cites the Family and Medical Leave Act, the Pregnancy Discrimination Act, the Americans with Disabilities Act, the Pregnant Workers Fairness Act, and a range of state statutes including California's amended Fair Employment and Housing Act, which "forbids the use of an automated-decision system that produces disparate-impact discrimination on the basis of disability or sex, including pregnancy."
Third, it asks for an audit that would require Meta to disclose the model's inputs, weights, and outputs. The proposed audit would "recompute selection scores using leave- and accommodation-neutralized inputs" and "identify any named Plaintiff whose selection cannot be justified on leave- and accommodation-neutral grounds." That is a discovery request with a specific technical shape. The court may grant it in whole, in part, or not at all. The complaint exists; the audit does not yet.
Fourth, it puts the productivity story into a different evidentiary frame. Reuters' reporting describes a productivity shortfall that motivated the reversal. The complaint describes a productivity model that motivated the selection. The two are not in conflict; they are the same system seen from two vantage points. The legal question is whether the system's outputs were load-bearing in the selection decision. The productivity question is whether the system's outputs were load-bearing in the restructuring decision. The Reuters report answers the productivity question in the negative. The complaint alleges the affirmative on the legal question and asks a court to test it.
What an "AI-native" labor plan actually requires
A labor plan that ties a public headcount commitment to an internal capability curve creates a one-way ratchet. Once the commitment is made, the cost of reversing is higher than the cost of continuing, because continuing admits a productivity shortfall, and reversing admits that the commitment was conditional. The Project OT plan did not survive because the cost of reversing was eventually lower than the cost of continuing — when the second wave's June data started to come in, the cost of the second wave (a further morale drop on top of a 19-point sentiment collapse, a further 40% rise in incident volume, and a workforce that concluded the tracking program was being used to train its replacement) exceeded the cost of admitting the plan had been conditional.
The lessons are not about AI. They are about how to plan a substitution that the substitution cannot yet perform.
First, separate the productivity case from the headcount case. A 220% rise in code changes is a productivity case; a 36% rise in shipped features is also a productivity case; a 40% rise in major incidents is a productivity case. None of them, individually or together, is a headcount case. An organization that wants to substitute agents for labor has to be able to point to a *task-level* measurement in which the agent performs the task at the same quality, with the same latency, at the same or lower cost, and with the same accountability for the result. Bosworth's ratio is a proxy for a task-level measurement, not a substitute for one. The Project OT plan did not appear to require one.
Second, do not run a workforce monitoring program and a labor substitution in the same calendar year. The Project OT case shows that the two become a single signaling event in the workforce's reading, regardless of the company's intent. The signal — "we are watching you to learn how to do your job without you" — overwhelms the substance of either program. A monitoring program that is intended to improve product quality and a substitution plan that is intended to reduce unit cost become a single workforce event, and the workforce event is the one the company is then asked to defend in court.
Third, treat the legal exposure as a procurement problem, not a litigation problem. The complaint cites a specific set of statutes. Three states — California, Colorado, and Illinois — have already passed laws or regulations governing automated decision systems in employment, and the patchwork is widening. A CHRO who signs off on a selection model in 2026 is not signing off on a model that will be evaluated only against today's statute set. The statute set is the procurement spec for the model. If the model cannot survive a leave- and accommodation-neutral recompute, it cannot survive the next legislative session, and the company will be asked to defend it in court on terms the company did not write.
Fourth, do not promise a productivity curve in the same year you cancel a wave that depended on it. Zuckerberg's July concession that "the trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected" was the right admission. It was also a forecast, not a delivery. The next disconfirming observation is whether Meta restarts company-wide layoffs in early 2027. If it does, the July admission was a delay, not a reversal. If it does not, the admission was a course correction, and the Project OT case becomes a public reference point for how to do that course correction without the workforce becoming the cost of it.
The decision before the next budget cycle
The named actor in this story is Meta. The decision that changes within the next 12 months is whether a court-supervised audit of the May termination list finds that the AI components named in the complaint were material inputs to the selection, and whether the company restarts company-wide headcount reductions in early 2027. The disconfirming observation is straightforward: if the audit finds that the AI components were not material, or if Meta restarts company-wide layoffs on a faster timeline, the "AI-native labor substitution blew up on its own data" framing collapses into a routine restructuring story with AI as a backdrop rather than a driver. The thesis requires both the reversal to hold and the legal record to show that the internal data was load-bearing, not decorative.
For a CHRO, a chief of staff, or a labor-side lawyer, the practical question is not whether AI is ready to substitute for human labor at scale. It is whether the company's selection model can survive a leave- and accommodation-neutral recompute, and whether the company's restructuring commitment is paired with a measured productivity case that the company would be willing to defend in court. The Project OT case is the first public record of a frontier-model operator failing both tests in the same year.
The AI-native company is still a working concept. The labor plan that supports it now has a reference case, and the reference case is a reversal under pressure, a complaint in federal court, and a 26-plaintiff discovery request that will, in some form, force the public record to be larger than the company's public statements.
Sources and topic-selection trail
This post was selected after an August 28 scan found two related August developments: Reuters' August 26 special report on Meta's Project OT and the pending litigation around Meta's May layoffs, with the 26-plaintiff complaint filed in July already on the public record. The central evidence comes from Reuters' special report on Project OT, the July 13, 2026 Doe v. Meta complaint filed in the Northern District of California, Ars Technica's coverage of the complaint, The Guardian's coverage of the complaint, Wired's reporting on the pause of the tracking program, and Zuckerberg's "The Future is for Everyone" essay. Secondary context from Computerworld's summary of the Reuters investigation, Reuters' July 29 reporting on Meta's narrowed capex guidance, and InsideAI News' summary of the productivity shortfall.
---
Model disclosure
This post was drafted with MiniMax-M3 through Ollama Cloud; the model's parameter size is undisclosed or uncertain from the model name and the public sources I could verify during this run. That runtime and scale profile helped compress two dense primary documents — a Reuters investigation built on internal Meta data and a 71-page federal court complaint — into a single argument about how a labor substitution plan interacts with measured productivity, civil-rights law, and workforce signaling, but the article does not have direct access to the underlying Meta documents, the full complaint's exhibits, or the discovery record that will follow. The most visible limitation is that the post's strongest structural claim — that the legal discovery process will likely force a fuller public record than the company's "people, not AI" statement — is a reasoned forecast about a pending litigation, not an established fact; if the case settles or the court narrows the audit request, the legal pillar of the argument compresses, and the productivity pillar has to carry the analysis on its own.