Meta's Project OT: 220% more code, 36% more shipped

Meta ran the experiment every AI vendor is selling you, at a scale you will never reach, and wrote down the numbers. Project OT — internal shorthand for Organization Transformation — was a plan to replace as much as 60% of some teams with AI agents. Reuters obtained the internal documents and reported on August 26 that it collapsed. The reason it collapsed is the most useful AI productivity data anyone has published this year.

What actually happened

Per Reuters' review of internal posts and three people familiar with the project, Meta cut roughly 10% of staff in May and planned a second wave for November. That second wave was halted.

The numbers underneath tell you why. In an internal post cited by Reuters, CTO Andrew Bosworth reported that code changes to Meta's internal platforms and infrastructure rose 220% year over year. New and upgraded features actually reaching users rose 36%.

That is the whole story in two figures. Output went up 6x faster than outcomes.

The second pair is worse. Major technical and security incidents climbed 40%, and the time employees spent responding to them rose 70% (summarized by TechSpot). Meta's half-year Pulse survey put employee sentiment at 55% favorable, down from 74%. Infrastructure teams had raised reliability concerns as early as March.

We wrote in July about Meta's agents running behind schedule. This is the audit trail for why.

Why AI output metrics will mislead your business

Every AI tool you can buy reports the thing it can count. Lines generated. Pull requests opened. Tickets drafted. Calls summarized. None of those are the thing you sell.

Meta had unlimited budget, elite engineers, and its own models, and still shipped a 220/36 split. A five-person shop running Claude Code against a legacy Rails app will not beat that ratio by trying harder. You beat it by measuring somewhere else.

Three things we do on every automation engagement:

Pick the outcome metric before you turn the tool on. Orders shipped. Invoices cleared. Tickets closed without a second touch. Write the number down in week zero. If your only baseline is "we feel faster," you have already lost the argument you will need to have in six months.

Count the rework. Meta's incident time went up 70%. That cost is real and it lands on the same people who were supposed to be freed up. Track escalations, reverts, and bugs traced to AI-written changes — and subtract that time from your savings. Most AI ROI decks skip this line entirely because it is the only one that can go negative.

Automate where the output is verifiable. A reconciliation script either matches or it does not. A generated marketing email needs a human read every time. The first compounds. The second is a treadmill.

The pitch is that agents replace headcount. The evidence from the best-funded version of that experiment is that unsupervised volume creates incidents, and incidents eat the savings. Automate the narrow, checkable work. Keep a human on the loop where being wrong is expensive.

Sources: Reuters — Mark Zuckerberg had a bold plan to replace Meta staff with AI, TechSpot.