How to Run a Postmortem on a Bad AI Recommendation

A bad AI recommendation is not automatically a reason to abandon the workflow. It is a reason to inspect the operating system around the recommendation.

The postmortem should answer what happened, why it passed through the workflow, and what boundary or scorecard needs to change.

Start with the event

Name the recommendation, the source material, the reviewer path, the action taken, and the consequence. Avoid starting with model blame. The failure may have come from bad source data, unclear approval rules, missing telemetry, poor instructions, or an action that should have been prohibited.

Review the boundary

Ask:

• Was AI allowed to make this recommendation?
• Should this action have been draft-only?
• Was an approval required?
• Did the system of record remain authoritative?
• Was the action logged with enough context?
• Was the failure detectable before it mattered?

Update the operating loop

The fix might be a better prompt, but it might also be a narrower permission, a stronger approval gate, a new test case, a scorecard metric, a reviewer checklist, or a prohibited-action rule.

Postmortems are how AI workflows earn more trust without pretending errors will disappear. Use the Responsible AI Statement and the Boundary Matrix as the governance frame.