In 30 seconds
- AI-generated work must meet the same Definition of Done.
- Add explicit checks for review, understanding and data safety.
- Make it a team agreement, not a tool setting.
Example additions
- A human has reviewed and understood every AI-generated change
- Tests were written or checked against the spec, not just generated from the code
- No sensitive data was shared with tools outside policy
- Generated code follows team conventions
- For AI features: evals pass the agreed threshold
Why it matters
The Definition of Done protects the quality of every Increment. AI makes it easy to produce work that looks done but isn't understood or properly checked. Explicit criteria close that gap.
Agreeing it
Discuss it in a retrospective, start with two or three additions, and review them after a few sprints. Pair it with human-in-the-loop review.
Myth vs reality
Myth · AI work needs a separate Definition of Done.
Reality · One shared bar for all work.
Myth · Tools enforce the Definition of Done.
Reality · Tools help; the team owns it.
Myth · Adding AI checks slows teams.
Reality · It prevents rework and incidents later.
FAQ
How many AI-specific items should we add?
Two or three to start.
Should we label AI-generated code?
Many teams note it in pull requests for review context.
Does this apply to documentation?
Yes, AI-drafted docs need review too.