Meta has eliminated tens of thousands of positions since 2022. At the same time, it has committed extraordinary amounts of money to artificial intelligence infrastructure and competed aggressively for some of the world’s most valuable AI talent.

At first glance, those trends seem to tell a straightforward story: Meta is spending more on machines while spending less on people.

The reality is considerably more complicated.

Meta began shrinking and restructuring its workforce before generative AI became central to its operating model. Some later cuts were tied to performance, cost reduction or organizational efficiency rather than automation. Yet by 2026, AI had clearly become part of the company’s workforce calculations.

According to an August Reuters investigation, Meta explored reorganizing parts of the company around an “AI-native” model in which smaller groups of employees, supported by AI agents, could perform work previously handled by larger teams. Some scenarios considered reductions of as much as 60% within certain groups.

Meta stressed that this did not mean eliminating 60% of its total workforce. The scenarios varied by team and included layoffs, redeployments and the elimination of unfilled positions. They were also never fully implemented.

Still, the experiment provides an unusually concrete test of one of the biggest claims of the AI era:

If artificial intelligence makes knowledge workers dramatically more productive, will major companies eventually need far fewer employees?

Meta’s experience suggests AI can change how many people companies need and which workers they value. But it also shows how difficult it can be to turn higher AI-generated output into equally large gains in useful productivity.

Meta Was Cutting Jobs Before the AI Replacement Push

Meta’s workforce reductions did not begin as an attempt to replace employees with generative AI.

On November 9, 2022, Mark Zuckerberg announced that Meta would eliminate more than 11,000 positions, about 13% of the company’s workforce.

His explanation focused largely on business conditions. Meta had expanded rapidly after the pandemic accelerated online activity, but Zuckerberg acknowledged that he had overestimated how long that growth would continue. The company was also confronting weaker revenue growth, increased competition and changes affecting its advertising business.

AI was already one of Meta’s priority areas, but the 2022 layoffs were primarily presented as a response to overexpansion and deteriorating economic conditions.

The reductions continued in 2023.

During what Zuckerberg called Meta’s “year of efficiency,” he announced plans to eliminate roughly another 10,000 positions and close about 5,000 unfilled jobs. Meta sought fewer management layers, fewer lower-priority projects and a higher proportion of technical employees.

Zuckerberg argued that the first round of layoffs had shown advantages to operating with fewer layers and less internal coordination.

Company filings show how dramatic the reset became.

Meta reported 86,482 employees at the end of 2022, although that figure still included most employees affected by the November layoffs. By the end of 2023, headcount had fallen to 67,317—a decline of 22% from a year earlier.

But Meta did not simply continue shrinking.

By the end of 2024, headcount had risen again to 74,067, up 10% from the previous year.

That rebound matters because it complicates the idea that Meta had already entered a simple process of replacing employees with technology. The company was willing to eliminate large numbers of jobs, but it was also willing to resume hiring when its priorities changed.

In early 2025, Meta announced another reduction of roughly 5%, targeting what it described as its lowest-performing employees. Reuters reported that Meta planned to hire for many of the affected roles later.

At the same time, the company was increasing recruitment of machine-learning engineers and other technical workers.

Fewer employees in one category did not necessarily mean fewer employees everywhere.

Then AI Changed the Equation

By 2025 and 2026, however, AI had become much more directly connected to Meta’s workforce strategy.

Zuckerberg publicly predicted that AI systems would increasingly perform software-engineering work. During Meta’s first-quarter 2025 earnings discussion, he said the company was progressing toward AI coding capabilities comparable to a mid-level engineer and expected coding agents eventually to perform a substantial portion of AI research and development.

Meanwhile, Meta was competing aggressively for elite AI talent.

Reuters reported in 2025 that Meta invested $14.3 billion in Scale AI as Zuckerberg built a new superintelligence effort around Alexandr Wang. That push was accompanied by intense recruiting for highly specialized AI researchers and engineers.

Those developments helped shape Meta’s next workforce experiment.

Reuters reported that Project OT, short for Organization Transformation, examined what an “AI-native” Meta might look like.

The concept envisioned AI agents taking on more routine work while smaller groups of highly capable employees supervised, directed and built around those systems.

Internal planning described traditional product groups of roughly 10 to 20 employees potentially being replaced in some cases by pods of three to five people, with specialist resources shared between teams.

Reuters reported that executives considered workforce reductions of up to 60% in certain groups.

Again, Meta said this did not represent a proposal to cut 60% of the entire company. Some reductions would have come from eliminating vacant positions or moving workers elsewhere.

But the logic behind the exercise was clear.

If three or four AI-assisted employees could accomplish what once required ten people, the economics of staffing a technology company would change dramatically.

The problem was proving that AI actually delivered productivity gains large enough to justify that assumption.

More Output Didn’t Necessarily Mean More Productivity

This may be the most important lesson from Meta’s experiment.

According to internal information reviewed by Reuters, changes to Meta’s internal software platforms and infrastructure rose 220% year over year as AI-assisted development expanded.

Yet changes that resulted in new or improved features actually reaching users increased only 36%.

Those numbers measure different things, so the gap should not be interpreted as proof that the remaining output was worthless.

But it illustrates an important distinction.

Output is not necessarily productive output.

An engineer producing three times as much code is not automatically three times as productive if much of that code requires additional review, creates maintenance problems or never results in a better product.

The same principle applies outside software development.

AI can help workers produce more reports, presentations, customer responses, designs, analyses and documents. But companies ultimately care about whether that additional work generates revenue, reduces costs, improves products, saves meaningful time or creates some other measurable business benefit.

Meta reportedly encountered another complication.

Reuters said internal infrastructure teams warned about reliability problems as AI-assisted coding expanded. Internal posts reviewed by Reuters showed major technical and security incidents rising 40% from a year earlier, while time spent responding to them increased 70%.

Meta declined to comment to Reuters on those internal figures.

The numbers highlight a potential productivity paradox.

AI can make producing work cheaper and faster while simultaneously increasing the amount of work required to review, secure, integrate and maintain what gets produced.

If those secondary costs become large enough, raw output statistics can substantially overstate the true productivity gain.

That does not mean AI coding failed at Meta.

The more defensible conclusion is narrower: the productivity gains were not sufficiently reliable to justify every assumption executives were considering about reducing staff.

Meta nevertheless carried out a major workforce reduction on May 20, 2026. Reuters described it as approximately 10% of employees. Meta later said about 8,000 workers were affected and recorded $1.18 billion in related severance expenses during the second quarter.

But plans for another company-wide restructuring wave later in the year were dropped.

Reuters reported that employee resistance, disappointing AI productivity results and operational problems were among the pressures surrounding Project OT, although the news organization could not determine precisely what caused Zuckerberg’s final decision.

That distinction matters.

The evidence does not show that one disappointing AI metric single-handedly ended the restructuring plan. It does show that Meta encountered practical limits while testing how far AI-enabled workforce reductions could go.

Meta Still Isn’t Backing Away From AI

Those difficulties should not be mistaken for evidence that Meta has lost confidence in artificial intelligence.

Its spending suggests the opposite.

Meta said in July that it expected 2026 capital expenditures of between $130 billion and $145 billion, narrowing its previous outlook of $125 billion to $145 billion.

The company is also continuing to spend heavily on people.

In its financial outlook, Meta said employee compensation would be one of the largest contributors to expense growth in 2026 because of investments in technical talent, specifically highlighting hiring in priority areas such as AI.

That makes the idea that Meta is simply replacing humans with AI difficult to sustain.

A better description is that Meta appears to be experimenting with different combinations of humans and AI.

Some positions may disappear.

Some tasks inside surviving positions may become automated.

Some teams may become smaller.

At the same time, certain workers—especially those capable of building, improving and deploying advanced AI systems—may become considerably more valuable.

Meta’s own workforce figures capture that tension.

The company reported 75,472 employees as of June 30, 2026, only 1% lower than a year earlier. That figure still included approximately 8,000 employees affected by the May reduction who were expected to leave reported headcount largely by the end of the third quarter.

Meta is therefore reducing employment in some areas while continuing to recruit expensive technical specialists in others.

Those strategies are not necessarily contradictory.

They may be parts of the same workforce transition.

AI May Change Who Companies Hire Before It Eliminates Hiring

The debate about AI and employment is often presented as a choice between two outcomes: either artificial intelligence replaces workers or it fails to do so.

Meta suggests a third possibility.

Companies can eliminate certain positions while automating only portions of other jobs. They can expect each remaining employee to produce more. They can consolidate teams and management layers while simultaneously paying enormous sums to recruit workers capable of building AI systems.

In that environment, AI’s employment effects would appear not only in layoffs but in the composition of the workforce.

Companies also do not need to automate an entire profession before changing hiring behavior.

If AI allows ten employees to accomplish work that once required twelve, a company does not have to replace ten people with machines for AI to affect employment.

It might simply hire two fewer people when employees leave.

Repeated across teams and companies, those smaller changes could meaningfully reshape employment without producing a dramatic moment when a company announces that AI has “replaced” its workforce.

Meta’s experience suggests companies are still trying to determine where that threshold lies.

According to Reuters, Zuckerberg acknowledged internally in July that AI-agent technology had not developed as quickly as he had expected, while maintaining that he believed the benefits would improve.

That distinction between direction and timing may be critical.

Meta has not abandoned the idea that AI can allow smaller teams to accomplish more.

Instead, its experience suggests that reorganizing a massive company around that assumption before the technology reliably produces the expected gains can create problems of its own.

What Meta’s Experiment Actually Tells Us

Meta does not prove that AI cannot replace knowledge workers.

It also does not prove that large companies can remove enormous portions of their workforces and simply transfer those responsibilities to AI agents.

The evidence supports a more nuanced conclusion.

AI is already changing how Meta thinks about productivity, team structure, hiring and the skills it values.

But Meta’s experience also demonstrates that generating more work is not the same as generating more value.

The reported 220% increase in changes to internal software and infrastructure compared with a much smaller 36% increase in user-facing improvements illustrates exactly the challenge companies must solve before the most aggressive predictions about AI-driven workforce reductions become sustainable.

For workers, the important question may therefore not simply be whether AI can perform their jobs.

For companies, the question may increasingly become whether an AI-assisted employee can perform enough additional valuable work that another position no longer needs to exist.

Meta tested an aggressive version of that idea and pulled back from some of its most ambitious restructuring scenarios.

At the same time, it is spending more than ever to improve the technology.

That makes Meta neither proof of an AI employment apocalypse nor evidence that concerns about automation are misplaced.

Instead, it offers an early view of how complicated the transition may become.

The first major workforce effect of AI may not be companies replacing people wholesale.

It may be companies repeatedly asking a narrower question:

Which people do we still need, how many do we need, and what must they be able to do now?

Sources

This article was prepared with AI assistance and reviewed and approved by TrendPulse Analytics before publication.