AI coding assistants have stopped being a curiosity and become a default part of the toolchain. The question is no longer whether to use them, but where they create value and where they quietly create damage.
Where they genuinely help
- Bridging unfamiliar territory: new languages, new SDKs, old codebases with sparse documentation.
- Mechanical translation and refactoring with test coverage as the safety net.
- Boilerplate and scaffolding that would otherwise consume an afternoon.
- Explaining unfamiliar code during onboarding and auditing.
Where we keep them out
We do not let assistants author acceptance criteria, security-sensitive logic, or anything close to the money path without a human sweep. The reason is not skepticism about output quality; it is about ownership. Code without a human's mental model attached to it becomes unmaintainable the moment the assistant is gone.
The expensive failure mode is not a bad line of code. It is confident, plausible code that nobody on the team can explain.
The rule of thumb
Use assistants for anything you could verify yourself in an hour. Use them for nothing you would struggle to review. Review becomes the discipline, and it needs to become more rigorous, not less, as generation gets cheaper.