Meta Planned to Replace Thousands of Staff With AI Agents. Then It Called Off Wave Two.

If anyone could replace staff with AI agents, it was Meta: money, engineers, and its own models. In 2026 it tried. Halfway through, it stopped.

What happened

Meta laid off roughly 8,000 employees, about 10% of staff, on May 20, 2026, and said it would close 6,000 open roles and move a separate 7,000 people into AI-focused work, the Wall Street Journal and Reuters reported at the time.

In August, a Reuters investigation (as reported by Ars Technica and The Globe and Mail, since Reuters' own page couldn't be retrieved for this article) described the plan behind it. Codenamed Project OT, for "organization transformation", it was created in January and explored scenarios in which some teams would shrink by as much as 60%, with small groups of people overseeing AI doing much of the daily work. It was to run in two waves: May and November.

Meta confirmed the two-wave plan and the 60% team scenarios to Reuters, but said the scenarios mixed layoffs and redeployments, that it never intended to cut 60% of the whole workforce, and that it cancelled the second wave before deciding how many people would go. According to Reuters, Mark Zuckerberg called off the November wave the night before the May cuts.

The numbers that mattered

Reuters, citing internal posts, reported the figures that make this story useful to anyone running a business:

• Code changes to Meta's internal platforms were up 220% year over year, according to a post by its CTO in early June.
• Changes that led to new or upgraded features reaching users were up only 36%.
• Internal posts pointed to AI agents taking "large-scale, disruptive actions that humans are unlikely to execute", and to a 40% rise in major technical and security incidents compared with the prior year.
• Employee time spent resolving those problems rose by as much as 70%.

Meta declined to comment on those posts. In July, Zuckerberg reportedly told staff that agent development "hasn't really accelerated in the way that we expected."

What this means for a small firm

Activity is not output. More code, more drafts, more emails generated: none of it counts until it shows up as finished work a customer sees. Measure the 36%, not the 220%.

Count the cleanup. When an agent does something a person wouldn't, somebody spends time fixing it. If you don't track rework time, your AI savings are overstated.

Don't cut first. Meta could absorb a stalled plan. A small firm that lets go of its one experienced admin or bookkeeper on a forecast can't. Prove the task works with AI for 60 to 90 days, then decide what to do with the hours.

Morale is a cost. Reuters reported that layoff reports and employee tracking hurt morale before the cuts. In a 20-person firm, that kind of fear spreads faster.

A better sequence

1. Pick one task. Measure the cost per finished task today.
2. Run AI on it with a named owner and a review step.
3. Re-measure at 30 and 90 days, including rework.
4. Redeploy freed hours to revenue work. Only then look at headcount.

Our AI ROI calculator handles step one. If the question is really whether to hire, see AI virtual assistant vs hiring a VA.

What to measure before anyone talks about headcount

Meta's internal contrast — far more code, far fewer shipped improvements, more serious incidents — is a useful warning about dashboards. Activity metrics make an AI rollout look productive even when customers see little change and experienced staff inherit the cleanup. Before reducing a role, track four things on the same task for at least one full business cycle.

• Finished output: work accepted by a customer, manager or downstream team, not drafts or generated volume.
• Cycle time: elapsed time from request to accepted work, including queues and approvals.
• Rework and incidents: corrections, escalations, refunds, security problems and time spent restoring normal work.
• Supervision: the minutes a knowledgeable person spends prompting, checking, repairing and explaining the system to others.

If finished output rises while supervision and rework stay flat, you may have capacity to redeploy. If generated volume rises but accepted output does not, you have a faster draft machine — useful perhaps, but not a staffing case. Keep the person who understands the exceptions in the loop until the system has handled busy periods, awkward customers and the cases that do not fit the standard path. That is the point at which a forecast becomes evidence.

The "should I replace staff with AI?" series: Meta's plan stalls · Ford rehires veterans · Cloudflare's "measurers"

Frequently asked questions

Did Meta replace staff with AI?

Meta laid off about 10% of staff on May 20, 2026 and moved thousands to AI-focused work, but according to Reuters it called off a planned second wave of restructuring the night before the first cuts.

What was Meta's Project OT?

An internal restructuring plan, reported by Reuters, that explored scenarios including cutting some teams by up to 60% and having small teams oversee AI doing much of the daily work. Meta says it never intended to cut 60% of its whole workforce.

Should a small business replace staff with AI?

Not as a starting point. Measure what AI actually does for a specific task first, including review and error time, and redeploy freed hours before cutting roles.