Every AI coding demo runs on a fresh project with clean conventions. The code most of us are paid to maintain is 10+ years old, carries generations of patterns, and has a wiki that stopped being accurate in 2021, while most of the actual business logic lives in developers' heads.

That gap is where AI rollouts quietly stall. The licenses get bought, the tooling helps on new files, and it stays close to useless on the code that fills most of the week. Not because the tools are weak, but because nobody has told them anything about the system.

I spend a couple of days on site with teams fixing this, and the work is stranger than it sounds. Almost none of it is configuration or specific tooling. Most of it is getting a machine to read the code and describe the system back to you. Reviewing the output with the people who have maintained it for years while they argue with what the tooling wrote. The arguing is the valuable part. By the end of the day, the repository knows things that used to live in three people's heads. This way, no tooling needs to guess how the codebase works and behaves; it's all documented.

I'll show what comes out of that day, what a model reliably gets wrong about a system it has only just read, and why the teams that do this in the wrong order walk away convinced AI has nothing to offer old code.

Vendor-agnostic. Copilot, Claude Code, and Codex all need the same thing.