Community and Maturity Model, Not Tools, Drove Meta’s AI Gains

Meta’s Horizon Experiences team grew AI tool adoption from 30–40% to over 80% weekly active users in 7 months by building a grassroots community and a self‑assessment maturity model. The six-dimension, five-level model let teams identify gaps in workflow integration, prompt skills, trust, and quality. Bottom‑up ownership, not mandates, drove adoption.

By tying AI to engineering excellence goals—test coverage, code quality—they saw measurable wins. An unsupervised agent workflow raised code coverage to over 90% with just 3 hours of human review, versus an estimated 1,920 manually. Legacy code migrations cut time by 75% using supervised pair‑programming.

But the speed introduced larger, boilerplate‑heavy diffs that strain code review. Trust and quality remain persistent concerns; hallucinations and slop haven't disappeared. Teams warn that accelerating code generation while slowing review can be net invariant. Measuring true productivity gains—not just tool usage—is still unsolved.

The approach works by creating safe sandboxes and layering AI onto existing goals, not by chasing full autonomy. However, accountability for AI‑authored code, long‑term codebase familiarity, and anti‑slops governance are open challenges that engineering leaders must address now.