Use this checklist before moving an AI pilot into daily use. Record an owner and evidence for each answer.
Task and data
- Is the task specific enough to test?
- Can the intended user verify the result?
- Can the system retrieve the required information with the user's permissions?
- Are missing, stale and conflicting records handled?
- Are data use, retention and provider arrangements documented?
Evaluation and actions
- Does a versioned test set represent ordinary work and failure cases?
- Are acceptance criteria agreed with the team using the result?
- Are reading, drafting and writing separate permissions?
- Do consequential changes require the agreed approval?
- Can operators stop writes without losing the work already done?
Operations and handover
- Is someone responsible for failed tasks and incidents?
- Can a retry duplicate an action?
- Can a reviewer trace the evidence behind a result?
- Have usage costs, response times and review effort been measured?
- Is there a tested rollback path?
Make a release decision
For each unanswered question, record the impact, next action and responsible person. Resolve gaps that could cause an unauthorised action or unrecoverable change before release.
Review legal classification and data protection with the appropriate specialists. This checklist is an engineering tool, not a compliance certification.
Record the controls and their owners before approving a pilot.