In 30 seconds

Run this regularly, monthly is a reasonable default, for each live AI workflow. It catches drift, stale knowledge and creeping problems before they become incidents, and forces the honest question of whether the workflow is still worth running.

Results and quality

  • Recent results have been reviewed against real outcomes.
  • The golden test set was rerun and results compared to last time.
  • Any drop in quality has been investigated for drift.
  • Overrides and user feedback have been reviewed for patterns.

Failures and exceptions

  • Failed cases since the last review have been examined.
  • Recurring failures have been traced to a cause and addressed.
  • The exception queue is being worked and is not silently growing.
  • Real failures have been added to the test set.

Knowledge and dependencies

  • Knowledge sources are current and within their review cadence.
  • Superseded sources have been removed from the searchable set.
  • Changes to external models, services or integrations have been checked.
  • Upstream changes to depended-on sources have been noticed.

Safety and access

  • Guardrails and limits are still in place and enforced.
  • The workflow's permissions are still appropriate and no wider than needed.
  • The pause or stop mechanism still works.
  • Any incidents since the last review are documented and prevented.

Ownership and value

  • The workflow still has an active, named owner.
  • The workflow is still delivering enough value to justify operating it.
  • If it is no longer worth running, retirement has been considered.

What this checklist does not prove

Completing it does not guarantee the workflow is perfect. It confirms someone is watching the things that quietly go wrong, which is what keeps a live workflow trustworthy over time. See how to assign ownership after launch.