Meta tried to replace workers with AI. It didn't go as planned
What went wrong with Meta's Project OT? Employee pushback, technical limits and uncertainty began to challenge its AI ambitions.
Meta wanted to build an AI-native company. Then the humans got in the way.
According to analysis by Reuters, Mark Zuckerberg and senior Meta executives discussed a major internal restructuring in January 2026 as part of their annual leadership retreat in Hawaii. The project, code-named Project OT, short for Organisation Transformation, was designed to make Meta more “AI native”.
The idea was to build teams around AI tools and agents from the beginning instead of adding AI to existing workflows later. But turning that idea into reality proved much harder.
Meta wanted AI to reshape its teams
Under Project OT, AI agents would take over parts of employees’ daily work, with smaller groups of highly skilled staff supervising the systems. Executives explored scenarios where some teams could shrink by as much as 60%.
Meta later clarified that this did not mean cutting 60% of its overall workforce, with the scenarios also including redeployments, role closures and layoffs. The restructuring was planned in two waves, with the first in May 2026 and another potentially coming in November.
Meta carried out a 10% workforce cut on 20 May. But plans for the second wave were cancelled the night before the first layoffs. The rethink came as employees grew increasingly uneasy about AI replacing the work they were helping build.
The productivity numbers did not quite add up
The biggest problem was not simply employee resistance. The technology itself was producing mixed results. Meta’s internal data showed a 220% increase in code changes to internal platforms and infrastructure. That sounds impressive until you look at what actually reached users.
Changes that became new or improved features increased by just 36%. In other words, AI was helping Meta produce much more code, but that did not translate into an equivalent increase in finished products. There were also reliability concerns.
Reuters reported that AI agents sometimes took disruptive actions that human employees were unlikely to make. Technical and security incidents rose 40%, while staff time spent resolving them increased 70%, according to internal posts reviewed by the news agency.
For a company trying to prove that AI agents could replace significant amounts of human work, those numbers were not exactly ideal.
The human problem was growing too
Employee anxiety added another layer to the experiment. Reuters states that Meta had promoted device-tracking software for some US employees that recorded activity such as keystrokes and mouse clicks. The programme was later paused.
Employee sentiment also reportedly dropped sharply, with Meta’s half-year Pulse survey falling from 74% favourable to 55%.
Workers raised concerns about unclear reporting structures, new “pods” and the growing role of AI systems in management support. Meta said performance and promotion decisions were, and remain, made by people.
Meta is still betting big on AI
Project OT did not change Meta’s broader AI ambitions. The company still plans to spend at least $130 billion on AI chips and infrastructure this year and has moved employees into priority AI work, including training data for its models.
But the Project OT experience offers a useful reality check. AI can write more code. It can automate tasks. It can help smaller teams do more. But turning that into genuine productivity is a different challenge.
For Meta, the experiment showed that becoming “AI native” is not simply about replacing people with agents. The harder part is making sure those agents actually work, employees trust the new system, and all that extra AI-generated activity translates into something users can see.