The Numbers That Killed Meta's AI Workforce Plan

At a leadership retreat at Mark Zuckerberg’s Hawaii estate in January, Meta executives sketched out an “AI native” version of the company: agents handling the bulk of day-to-day work, overseen by small, “talent-dense” human “pods,” with general-purpose “builder” roles replacing narrower job titles and whole layers of middle management gone. Internally it was code-named Project OT, for Organization Transformation, according to internal planning documents reported by Reuters. It was supposed to run in two waves — the first in May, a second in November — with some teams facing headcount cuts of up to 60%, on top of eliminated open roles and the removal of lower performers. The first wave alone matched or exceeded the scale of Meta’s 2023 “Year of Efficiency” layoffs. The second wave never happened: Zuckerberg canceled it the evening of May 19, hours before the first round of cuts landed the next day.

What sank it, by Meta’s own internal numbers, was a widening gap between how much AI-generated code was flowing into the company’s systems and how much of it actually reached users. Code changes to Meta’s AI platforms and infrastructure rose 220% year over year, but new or improved features that shipped to users grew just 36% over the same period — and the time employees spent fixing the technical and security problems that came with the AI-written changes grew 70%. In July, Zuckerberg told staff that “the trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” and that the company’s bet on AI agents “hadn’t come to fruition yet.”

The other half of the collapse was internal. When employees realized that software tracking their mouse clicks and keystrokes — introduced to measure “builder” productivity — was also generating the training data for the agents meant to replace them, Meta’s internal network filled with public pushback, and labor-organizing efforts inside the company picked up. Internal sentiment scores fell 19 points over the course of the rollout.

Zuckerberg’s move wasn’t a full reversal — it was a pause on the second wave, not a cancellation of the AI-native reorganization itself. Meta says the restructuring continues in a smaller, slower form: leaner AI-assisted teams rolling out across the company without the second mass layoff that was supposed to accompany them. The metrics that forced the retreat — the output-to-shipped-feature gap, the incident-rate increase, the sentiment drop — are the ones that will determine whether that quieter version fares any better than the one that got shelved in May.