There is a satisfying version of this story in which Meta spent years declaring war on middle management, discovered that artificial intelligence could not run the company, and quietly asked the managers to come back.
It is also too neat.
Fortune reported that Meta’s Applied AI division had asked some individual contributors whether they would volunteer to move into management roles. The report, drawing on Business Insider’s interviews with four people familiar with the matter, described a selective request inside one organization—not a company-wide reversal. Meta did not confirm the details to Fortune.
That caveat should stay attached to every conclusion. But the reported move is still interesting because of where it occurred. Applied AI exists to move models from research into products. It sits between teams with different clocks, incentives and definitions of “ready.” If any part of Meta would rediscover the need for explicit coordination, this is it.
The lesson is not that managers always win. It is that AI lowers the cost of producing components more quickly than it lowers the cost of making those components work together.
Flattening solved a real problem
Meta did not invent its aversion to layers out of nowhere. In a large technology company, management can become a forwarding service. One person turns an executive request into a slide, another turns the slide into a planning document, and a third schedules the meeting in which engineers explain what they already knew.
Mark Zuckerberg’s 2023 “year of efficiency” attacked that accumulation. Managers were moved into individual-contributor roles or asked to leave. In May 2026, Meta cut roughly 8,000 positions and dropped plans to fill another 6,000, according to Fortune, with managers disproportionately affected.
The appeal is obvious. Fewer layers mean shorter communication paths and more people building. Add code assistants, automated testing and generative tools, and the argument becomes even stronger: a small group of capable engineers can produce far more than it could three years ago.
For bounded work, that argument holds. An engineer improving a self-contained service may need fewer hands and less supervision. A team with stable interfaces can make decisions locally. If a manager exists only to collect updates, an automated dashboard is probably better.
The trouble begins when the work is not bounded.
Applied AI is mostly boundary work
Meta’s AI program now touches recommendation systems, advertising, assistants, business agents, model APIs and new consumer products. On its second-quarter earnings call, Zuckerberg said AI was helping teams ship faster. The company also described infrastructure spending, research releases and product ambitions large enough to make “AI” less a department than a dependency running through the business.
Applied AI’s job is not simply to write more code. It has to convert probabilistic model behavior into product behavior that can survive billions of users, fixed latency budgets, policy requirements and limited compute.
Imagine three teams moving faster at once. Researchers produce more model variants. Infrastructure engineers change routing and capacity plans to keep costs under control. Product teams use AI tools to prototype more features. Each group looks more productive on its own. Collectively, they create more handoffs, more incompatible assumptions and more decisions about who gets scarce resources.
The work moves from inside the boxes to the lines between them.
Someone has to decide whether a model improvement is worth additional latency. Someone has to tell a product team that its launch cannot have the capacity it expects. Someone has to own the failure that appears only when a new model, serving system and application release meet in production.
Those decisions are not status reporting. They require authority across teams.
This is why I resist the easy claim that AI eliminates management. It can eliminate portions of managerial work: summarizing, scheduling, drafting plans, tracking tasks and moving information. It does not automatically eliminate contested priorities. A model can assemble the trade-offs. It cannot decide which vice-president absorbs the cost unless the company gives it power that no serious board is ready to delegate.
Manager and bureaucracy are not synonyms
The argument for bringing back selected managers depends entirely on what job they are being asked to do.
A bureaucratic manager adds a checkpoint because the organization once had a checkpoint. A useful manager removes uncertainty about ownership. The first asks for another review. The second decides which review can be skipped and accepts responsibility for the result.
In applied AI, the valuable version of management looks closer to systems integration than personnel administration. The job is to keep research, infrastructure and product interfaces compatible; arbitrate when they are not; and protect engineers from spending half their week negotiating across organizational boundaries.
Without that role, coordination does not disappear. It becomes informal. Senior engineers pick it up because they understand the dependencies. They attend more meetings, chase owners and resolve priorities without the title or authority to make the decision stick. A flat org chart can conceal a large shadow management layer.
That is often the worst of both systems. The company loses accountable managers but keeps the coordination work, now distributed among people whose performance is measured by technical output.
If Meta is asking former managers to resume the role, the important detail is not the title. It is whether they receive decision rights.
Scale changes what “efficient” means
Meta reported $60.8 billion in second-quarter revenue, up 28% from a year earlier, and $42 billion in expenses, up 55%. It spent $31.1 billion in the quarter on capital expenditures including finance-lease principal, largely for servers, data centers and networking. Those are Meta’s figures, and they describe a company under pressure to convert unusually large AI spending into products and revenue.
Against that scale, another management layer can look offensively inefficient. It can also be cheaper than a month of duplicated model work or an underused cluster reserved for a launch that slipped.
Efficiency is not the smallest number of people between an engineer and the CEO. It is the amount of useful, integrated output the organization produces for the resources it consumes. Sometimes fewer layers improve that ratio. Sometimes they leave expensive decisions unowned.
Meta’s reorganization is therefore more useful as an experiment than as a morality play. The company flattened aggressively, accelerated internal production and then reportedly found a place where coordination capacity needed to be restored. That does not invalidate the earlier cuts. It suggests the cuts were a rough instrument and the repair must be more precise.
What other companies should take from this
Executives looking at Meta should not conclude that they need to preserve every middle-management role. They should map where work crosses boundaries.
Teams with stable interfaces and local decisions can remain flat. Teams responsible for connecting a changing model, a constrained infrastructure layer and several product organizations need named owners. The faster AI makes each team, the more frequently those interfaces will be tested.
There is a second warning. Restoring managers can become an excuse to recreate the old meeting structure. If the new role is defined by headcount, reporting and performance reviews, Meta may bring back bureaucracy without fixing integration. If it is defined by authority over cross-team trade-offs, the role may allow more engineers to spend time engineering.
That is the concrete test for Applied AI. Do the returning managers have responsibility for a model-to-product boundary, a capacity decision or a launch outcome? Or do they simply have more direct reports?
AI can draft the weekly update either way. Only one version solves the coordination problem.