# AI-assisted change review

## Task and authority
- Intended user outcome and human-written acceptance criteria:
- Model / provider / version or configuration:
- Permitted data and approved retention arrangements:
- Tools and permissions available to the assistant:
- Actions requiring human approval:
- Change owner and reviewer:

## Independent evidence
- [ ] A reviewer understands the implementation and can maintain it.
- [ ] Tests compare behaviour with requirements, not only generated output.
- [ ] Negative cases, boundaries and failure paths have independent expectations.
- [ ] Dependencies, licences and configuration changes have been reviewed.
- [ ] Secrets, personal data and customer material are handled under agreed rules.
- [ ] External or retrieved content cannot silently grant the assistant new authority.
- [ ] Tool output and generated instructions are treated as untrusted input.
- [ ] High-impact writes, deployments and communications have enforceable approval boundaries.

Record evidence, reviewer and remaining questions next to each check.

## For an AI feature or workflow
- Representative evaluation examples and expected outcomes:
- Uncertain answers, refusal and escalation behaviour:
- Provider outage and stale-document handling:
- Cost and latency observations under realistic load:
- Manual fallback, monitoring owner and stop conditions:
- Change triggers requiring a fresh evaluation:

## Decision and operation
- Accepted limitations and approving owner:
- Required fixes before wider use:
- Pilot scope and observation period:
- Handover, support and rollback reference:

Example — fictional: a document assistant may draft an answer from approved internal material but may not send that answer to a customer. Review the citation against the source and keep the send action with an authorised person. A correct answer in one demonstration does not establish reliability across the document collection.

Treat model configuration, source documents and tool permissions as part of the deployed system. Re-evaluate when any of those change; keeping the same prompt does not mean the behaviour stayed the same.