On September 11, Business Insider reported that Meta has started asking individual contributors inside its Applied AI division whether they would like to run teams again. Four people familiar with the program described it as opt-in: the company asks, the engineer answers, nobody is reassigned.
We covered the other end of this on May 21. The 4 a.m. email went out on May 20 to roughly 8,000 employees. Another 6,000 open roles were cancelled before anyone filled them. And roughly 7,000 people who kept their jobs were moved into four named AI initiatives, of which Applied AI was the largest. Chief People Officer Janelle Gale’s memo called the design “AI-native,” and spelled out what that meant in practice: fewer management layers, smaller pods and cohorts that move faster. Some of the people reshelved into Applied AI had been managers until that morning.
Four months later Meta is asking some of them to take the title back.
The scope is narrow and the mechanism is voluntary
Applied AI (AAI) is a division Meta stood up this year to close the distance between its AI research and its shipping products. The group trains models and pushes them into Facebook, Instagram and WhatsApp. The roughly 7,000 people moved in at the start of the year are now most of its headcount.
Two limits on the story are worth naming precisely. The program is opt-in, per Business Insider’s sourcing. And it appears only inside AAI; Meta’s manager-light posture elsewhere in the company has not moved.
The financial backdrop has not moved either. Meta closed Q2 with 75,472 employees, down 3% from the prior quarter, a figure that already absorbs the May cuts. Revenue came in at $60.8B, up 28% year over year. Total expenses rose 55%, to $42B. Costs are climbing at roughly twice the rate of revenue, which is the arithmetic the May cut was answering.
One more piece of context belongs in the frame. AAI has been unsettled since it was created. Earlier this year Wired reported internal frustration over the division’s rollout, and some of the people reassigned into it were later given the option to look for something else inside the company. So the 7,000 was never a fixed roster. The engineers being asked about management now are the ones who stayed, and who have two or three quarters inside the division behind them.
Meta did not immediately respond to Fortune’s request for comment.
Flattening booked a return in one quarter
Meta has run this play before. In 2023, during the “year of efficiency,” the company asked managers and directors to move into IC roles or leave, in a process it called flattening internally. The 2026 version was harder: cut first, then reseat the survivors by AI initiative, and take out management layers on the way through.
What appeared on September 11 is the same maneuver run backwards, with one quarter between the two.
The location of the reversal carries more information than the reversal itself. It did not surface in ads sales. It did not surface in infrastructure. It surfaced in AAI, the division whose output is AI. Moving a model from a research artifact into a surface used by billions of people is mostly alignment work: data, evals, launch sequencing, calendar negotiation with a dozen product teams. That work does not evaporate when the coordinator’s box is deleted from the org chart. It redistributes onto the engineers, and it keeps redistributing until throughput drops. Then the title comes back.
For the cohort point, set it against the two nearest stories on this site. On September 7, Jaguar Land Rover named 4,000 roles over two years and flagged management as the population at risk. On September 9, Cognition raised at a $48B valuation on ARR moving from $492M toward $900M, with humans-as-overseers as the story it told. The three do not point the same way. The first two argue the coordination layer can be thinner. Meta is the first company in this cycle to put a public number on how thin is too thin, and the number is four months.
The difference between the three is how much certainty each one is working with. JLR is cutting management above an assembly line whose takt time, shift patterns and labour calendar have been stable for decades; delete a layer and the remaining people can absorb it. Cognition is selling a product that is still growing, and it puts humans in the oversight seat precisely because model output currently needs someone underneath it, which is coordination by another name. AAI sits in the hardest band: a product already serving billions, a model that re-versions every few weeks, and both sides moving at once. In that configuration you cannot hand coordination to process. You have to hand it to a person.
There is a timing contrast too. JLR’s 4,000 is a two-year plan announced once. Meta went from 8,000 out in May to a partial walk-back in September, one quarter apart. Same layer, two clock speeds: one company plans management headcount by the year, the other revises it by the quarter. The second is not weaker conviction. It is a business that cannot see three months out.
What this means for coordination-heavy roles
The useful thing here is not Meta. It is the lag.
May 20 to September 11. That is how long it took a company with enormous cash, aggressive AI spend and no external pressure to reverse part of its own flattening. Call it a discovery interval rather than a reversal. Coordination cost is invisible on an org chart and visible in throughput, and throughput needs a product cycle or two to report.
Three things fall out of that for anyone sitting in a coordination role.
Which division is being flattened matters more than the memo. Where output is well defined and the process is stable, a deleted management layer may never come back. Where the output is still changing and the team re-aligns every week, that layer tends to return under some other name. AAI is the second kind. If your group’s roadmap gets rewritten monthly, the flattening applied to it is a test, not a settled structure.
The title returning does not mean the person returns. Meta is asking engineers already inside AAI whether they want to lead teams. It is not calling back the managers who left in May. Those people are not on this org chart. The layer is being rebuilt from the survivors, which is cheaper, faster, and produces a different set of managers than the ones removed.
Read “voluntary” twice. A voluntary manager role means the company has not yet decided how much of the layer to restore and has not committed headcount to it. It is a probe. Context matters here: in July, 26 Meta employees sued the company, alleging that internal AI systems and activity-monitoring data were used in the May cuts to disproportionately select workers on medical, parental or family leave. A company litigating how it selected people for removal has every reason to move slowly on anything that looks like a formal commitment.
The window to watch is Q3 earnings and the careers page. If AAI manager roles graduate from an internal question into posted requisitions, the layer is being restored. If Q3 closes and the asking is still internal and informal, Meta is buying the coordination back at the lowest headcount price it can find.
Sources
- Meta asks some AI employees to become managers again — Business Insider
- Meta bet AI would shrink its management ranks. Now it’s quietly rebuilding them — Fortune, 2026-09-12
- Meta Cut Its Managers. Now It Wants To Bring Some Of Them Back — Forbes, 2026-09-11
- Meta reversing course on flat org structure in Applied AI division — Quartz, 2026-09-11
- Meta Reports Second Quarter 2026 Results — Meta Investor Relations
- Meta Q2 2026 Earnings Call Transcript
