Which scheduling exceptions most often need human correction after an assistant interaction?
Replay synthetic time-zone, cancellation, rescheduling and duplicate-request cases against an expected calendar outcome.
What we could measure together
Measure: Incorrect or conflicted bookings per reviewed scheduling attempt (exceptions/attempt).
Freeze eligible work, observation window and acceptance criteria with the owner; retain incomplete work and exceptions.
Acceptance boundary: Use test calendars only; retain failed attempts in the denominator and require owner-approved success rules.