Ceremonies
Rewriting the Definition of Done for AI-Assisted Teams
When AI writes the code, the tests, and the documentation — what does "done" actually require? A Scrum Master's guide to updating the DoD for agentic development.
Why the Definition of Done Needs an Update
The Definition of Done (DoD) is a shared commitment: every story that is called "done" has met every item on this list. In teams using AI heavily, several traditional DoD items have become ambiguous or insufficient.
Traditional DoD items that AI complicates:
| Traditional Item | The AI Complication |
|---|---|
| Code reviewed by at least one team member | Did the reviewer understand the AI-generated code, or did they approve it based on test passage? |
| Unit tests written and passing | AI can generate tests that pass without testing the meaningful behavior. Tests can be trivially satisfied. |
| Documentation updated | AI-generated documentation can be accurate at a surface level but miss the intent and context a human author would include. |
| No known security vulnerabilities | AI code generation tools introduce new classes of vulnerability patterns. Traditional SAST tools may not flag them. |
| Acceptance criteria verified | If AI wrote the acceptance criteria and AI wrote the code to satisfy them, is there a meaningful human verification step in the loop? |
A Revised DoD for AI-Assisted Teams
The following additions are not universal — your team's DoD should reflect your specific context. These are starting points for the conversation.
AI disclosure and review:
- For any story where AI tools contributed more than 30% of the code: one team member who did not write the code has read and understood the logic, not just the tests.
- AI-generated code is identifiable in the commit history or PR description.
Test quality criteria:
- Tests cover at least three edge cases not present in the happy-path scenario.
- At least one test was written by a human to verify a behavior the developer identified as non-obvious.
- No test consists entirely of AI-generated mocks that mirror the implementation.
Security review:
- AI-generated code that handles user input, authentication, or external API responses has been reviewed against the OWASP Top 10.
- Prompt injection attack surfaces identified and documented for any story that involves LLM integration.
Documentation standard:
- Documentation includes the "why" not just the "what" — the decision rationale that an AI would not know to capture.
- If AI generated the documentation, it has been edited by a human who understood the feature.
Facilitating the DoD Conversation
The Scrum Master does not own the DoD. The entire Scrum team does. But the Scrum Master facilitates the conversation that produces it and monitors adherence.
Facilitation approach for updating the DoD:
- Before the retrospective where you plan to discuss the DoD, share a pre-read: two or three recent stories where you observed the existing DoD was insufficient or misleading given AI's involvement.
- Open the conversation with a specific question rather than "should we update the DoD?" Ask: "When we called this story done and it had AI-generated tests that didn't catch the production bug — what would a better done have looked like?"
- Propose specific additions, not general principles. "AI code was reviewed by a human who wasn't the author" is actionable. "We should be more careful about AI" is not.
- Time-box the implementation. Add one or two items to the DoD. Live with them for two sprints. Inspect in the retrospective.
The DoD as a Coaching Tool
In teams new to AI-assisted development, the DoD conversation is often the first time the team explicitly names what it believes about AI's role in their process. That conversation surfaces assumptions that need to be examined:
- Does the team believe AI-generated code requires the same review standards as human-written code? (It does, and often higher.)
- Does the team believe AI-generated tests have the same value as human-written tests? (They have value, but different value — the human tests the behavior, the AI tests the structure.)
- Does the team believe that if AI produced the code and the tests and the docs and all three are consistent with each other, that constitutes done? (This is the most dangerous assumption — consistency between AI outputs is not the same as correctness.)
These are not trivial questions. A Scrum Master who surfaces them does more long-term good than one who writes a new DoD item and moves on.