Meta's project, codenamed Project OT, was meant to turn the company into an 'AI-native' organization: small teams of people would manage AI agents, while some teams would be cut, in some cases by up to 60%. As a Reuters investigation showed, the plan was effectively scaled back: Zuckerberg canceled the second wave of layoffs, limiting cuts to roughly 10%. Internal telemetry, meanwhile, recorded a rise in serious incidents and response times, and a customer-support AI bot became a channel for hacking major Instagram accounts.

image

What happened

On August 26, 2026, Reuters published an investigation based on Meta internal documents, recordings, and interviews with more than 20 employees. It describes Project OT ('Organization Transformation'), which originated at Zuckerberg's January retreat in Hawaii: the end goal was an 'AI-native' company, with some teams planned to be cut by up to 60%. Before the May 20 layoffs, Zuckerberg canceled the second company-wide wave of cuts and limited them to roughly 10% of staff. Internal telemetry showed a 40% year-over-year increase in serious technical and security incidents and a 70% increase in response times, alongside more modest growth in actual product improvements. In June, a reputational problem was added: hackers used Meta's customer-support AI bot to gain access to major Instagram accounts, including the Obama-era White House archive account. Meta officially confirmed Project OT as a year-long initiative to cut costs, restructure teams, and move some employees to creating training data.

Context

The claim that 'agents will replace teams' had existed mainly as a declaration: promises of CEO agents and 'AI-native' organizations were not accompanied by verifiable numbers. Project OT became the first known case where such a plan at the scale of a large company ran into operational telemetry. The economics did not add up: the rise in incidents and response costs ate into the expected savings on headcount, and the relatively modest final cuts look like a direct consequence of that gap. Notably, the way the short story was officially closed: Meta shifted the focus to cost-cutting and retraining people for tasks such as generating training data. This is a typical big tech pattern, where roles shift from writing code and content to data and eval cycles for models.

Why this matters for the industry

This is the first detailed field breakdown with numbers of how 'replacing people with AI agents' works in practice at a company the size of Meta, and the aggregated result so far is negative: incidents +40% y/y and response times +70% give competitors and engineers concrete data against the argument that 'agents are ready to replace teams.' The effect is already in negotiations: offers to 'replace your team with agents' now meet counterarguments with numbers, and pilot deals are shifting toward an 'agent plus human' format and contractual reliability metrics. The hack of the customer bot outlines an unoccupied product layer — not 'another agent,' but secure-operation infrastructure: per-agent telemetry, permission boundaries, action audits, human-in-the-loop for privileged and irreversible operations. Teams launching customer-facing agent bots are already forced to review the scope of permissions and add human confirmation for sensitive operations, and expected directions in eval methodology are shifting from one-off benchmarks of demo tasks to endurance: incident frequency, recovery time, response cost.

Why this matters for users

The key risk for users was shown by the June incident: attackers gained access to other people's accounts through an automated support channel, meaning the place a person turns to when a problem has already happened. Practical consequences: be more cautious about any support-bot requests to verify identity or change data, enable two-factor authentication, and treat unexpectedly 'returned' access to an account as a possible incident, not luck. For news readers, the material works as a filter: when a company announces an agent-driven restructuring and layoffs, look at operational metrics — number of incidents, support response speed, product quality — rather than statements about the future. Finally, it is worth watching the reorientation of roles: the confirmed move of some employees to creating training data shows which engineering and content professions in big tech are changing profile right now, and this directly affects career plans for specialists.

What is still unknown / limitations

The +40% and +70% metrics are observational: the definition of a 'serious incident,' the comparison baseline, and normalization are not disclosed, so a strong formula that 'the outcome refutes capability claims' would be an exaggeration. The rise could have been partly caused by the restructuring itself and the departure of some engineers during the ~10% cuts, rather than the quality of the agents as such; a direct causal link between Project OT and the metrics is not proven by the material. Details of the June hack — the mechanism of the bot's compromise and the exact scale of the hackers' access — remain outside the published data.

Sources

Author

Look at AI, editorial team