Meta Weighed Cutting Some Teams by 60% for AI, Then Backed Off. Why?
Meta tested whether AI doing half the work means half the staff. Employee pushback and weaker-than-expected productivity gains led it to drop further cuts.
The 20-second read
If AI can handle half of an employee's work, can a company cut its staff in half?
Meta tested that question with a real reorganization. It weighed cutting headcount by as much as 60% on some teams and having a small number of employees manage AI agents. Amid employee pushback and productivity gains that fell short of expectations, it withdrew plans for further cuts.
The key issue is not how much work AI handles. It is how much more value the company creates as a result.
1. Meta's plan was more than a simple layoff
According to internal documents Reuters published on Aug. 26, 2026, Meta has run a reorganization effort since January called Project OT (Organization Transformation).
The core idea was that AI agents would handle a substantial share of routine work, while a small number of employees supervised multiple AI systems. One concept would have reshaped product development teams of 10 to 20 people into small teams of three to five.
Meta cut about 10% of its workforce on May 20. The night before, executives canceled a second, larger restructuring that had been planned for November. Thousands of employees were also reassigned to new work.
Employees worried they would be training AI while also putting their own jobs at risk. Positive responses in an internal employee satisfaction survey fell from 74% to 55%. Reuters could not confirm the exact reason for the change in plans.
2. AI produced more code. Why didn't results follow?
- +220%: code changes on internal platforms
- +36%: changes tied to user-facing features
These are year-over-year increases in two different metrics. They do not mean only 36% of code was shipped.
Inside Meta, AI use sharply increased code output. Code changes on internal platforms rose 220% from a year earlier, while changes tied to new and improved features delivered to users rose 36%. The two figures measure different things, so they do not mean only 36% of code was deployed.
Stability was a bigger problem. According to internal posts Reuters reviewed, major technical and security incidents rose 40% year over year, and employee time spent responding to incidents rose by as much as 70%.
Even if AI speeds up development, it does not necessarily cut the cost of checking results and fixing errors. Speed of work is not the same as total productivity.
3. Automating half the work does not make half the staff unnecessary
Suppose AI saves half the working hours of a 10-person team. On paper, five people could handle the same workload. But employees are not fully interchangeable units of work.
Interpreting customer needs, handling exceptions, verifying faulty AI output and taking responsibility for final decisions all remain after automation. As AI generates more output, there may be more output to review.
So the first thing to examine is tasks, not headcount: which steps are automated, what new oversight work that creates, and where the freed-up time will be reinvested.
4. The hidden cost: where do new hires gain experience?
Routine work can look low in productivity, but it is also how new hires learn the product and the customer. If AI takes over organizing materials, basic analysis and first-draft code, companies need to design new learning paths.
Alternatives could include reviewing real customer cases, planting errors in AI output for trainees to find, and analyzing the reasoning behind experienced staff's judgments. Without this kind of training, labor costs may fall in the short run, but the foundation for the next generation of skilled workers could weaken.
- No. The figure of up to 60% was a scenario for shrinking some teams. Meta said it had no plan to cut 60% of its total workforce.
- No. It canceled the second round of restructuring planned for November, but running small teams, adopting AI in work, and reassigning employees all continued.
- You can't conclude that. The key point is that code changes surged, but improvements delivered to users did not grow as much, and incident response costs also rose.
- Along with learning to use AI tools, it is important to build skills in verifying results, handling exceptions, understanding customer context, and making final decisions.
What to watch
Insight Times Editorial Desk
