Rolling out AI-assisted development across thirty-five engineers

Context

AI coding assistants arrived faster than most engineering organisations could absorb them. The failure mode I wanted to avoid was the common one: a few enthusiastic developers using a tool well, everyone else either ignoring it or trusting it uncritically, and no shared standard for what “reviewed” means when a machine wrote the first draft.

Leadership & decisions

Sequenced, not announced. I ran GitHub Copilot as a pilot with a subset of engineers before any org-wide rollout, and later repeated the process for Claude Code. Adoption problems surface in a pilot cheaply; they surface across thirty-five people expensively.

Standards before speed. The rollout shipped with review policy attached: what may be generated, what must be read line by line, and what an engineer is accountable for regardless of who wrote it. Velocity gains that arrive without review discipline turn into defect rates later.

Training as the real cost. The licence is the cheap part. I treated enablement — how to prompt, when not to, how to review generated code — as the actual investment, and ran it squad by squad.

Honest measurement. I have not instrumented this rigorously. The figure below is our own estimate across the teams that adopted it, on work suited to AI assistance, and I present it as an estimate rather than dress it up as a measurement.

Outcome

All four squads work with AI assistance as a normal part of delivery, under a shared review standard. On work well suited to AI assistance, our own estimate was a reduction in development effort of roughly half.