The problem
Support teams often combine keyword rules, manual review, and a general model. The result is difficult to tune and easy to over-escalate.
The Jev questions
Ask Jev for the primary topic, a severity score, and whether a human should review the case immediately. Keep the output fields aligned with the actual queue operations.
The production loop
Measure time to first response, escalation precision, and reopened tickets. A good Jev classifier is one that improves the queue, not one that merely looks confident.