Mark Zuckerberg spent the first half of 2026 telling anyone who would listen that AI agents were about to change everything inside Meta. This week, he told his own staff that the plan is behind schedule.
At an internal town hall on Thursday, Zuckerberg said that AI agent development over the past four months had not "accelerated in the way we expected." He also admitted that the company's large restructuring, which included laying off about 8,000 people and reassigning 7,000 more into AI-focused teams, was not as "clean" as it should have been. The expected benefits of that new structure have not arrived yet.
To understand why this matters, consider the scale of the bet. Meta is projected to spend between $125 billion and $145 billion on AI infrastructure this year alone. That is more than double what the company spent in 2025. The layoffs were not just a cost-cutting exercise: they were framed as a way to free up money for this AI push while simultaneously building the AI teams that were supposed to replace some of that lost capacity. Both sides of that equation are now in question.
Inside Meta, the situation is worse than a simple delay. Reports from Wired describe morale inside the Applied AI team, which has around 6,500 staff, as "hitting rock bottom." Employees assigned to the team say they spend their weeks on repetitive tasks like generating puzzles to test AI models, work they describe as having no clear purpose. The kind of soul-searching you would expect from a group of people who watched their colleagues get laid off and were then handed busywork.
Zuckerberg's admission also quietly unpicks a story the broader tech industry has been telling for the past year: that companies cutting staff are proof that AI is working, that software is doing jobs humans used to do. Meta's own CEO is now saying that is not what happened at his company. The layoffs were driven by budget pressure and the need to fund AI spending, not by AI actually delivering productivity gains. Those are very different things.
The wider picture is not much more encouraging. Research from MIT covering 300 AI deployments found that 95% of generative AI pilots fail to deliver any measurable financial impact. Analysts estimate that 40% of AI agent projects will be cancelled before the end of 2027. A Carnegie Mellon study that staffed a simulated company entirely with AI agents found the best-performing model completed just 24% of its assigned tasks. Even real-world deployments that looked promising have run into trouble: an autonomous booking agent at a major airline rebooked over 1,200 passengers onto wrong flights during a single weather disruption.
The failure pattern is consistent across companies. AI agents work well when given a narrow, specific job with clear rules and a human watching the results. They struggle when given broad responsibilities, messy data, or the expectation that they will replace entire roles without organisational redesign. Zuckerberg planned for the second scenario and is now living with the consequences.
For business operators currently being pitched AI automation as a way to reduce headcount or transform operations quickly, this is a useful data point. The world's best-funded AI company, with thousands of engineers and $145 billion to spend, is three to six months away from seeing results it expected to have already. Narrow, supervised, well-defined uses of AI tools are delivering real value right now. Wholesale replacement of job functions is not, and any vendor telling you otherwise is selling a timeline that even Meta cannot meet.