Test a Model on the Exceptions You Actually See

A model can handle ordinary support tickets well and still fail on the exceptions that consume most of your team's time.

Build a small evaluation set from representative, appropriately sanitized cases. Include a missing order number, conflicting dates, an unclear request and a case that must be escalated. Write the expected decision before looking at the model's answer.

Keep a held-out set for later changes so you do not optimize only for the examples in the prompt. Track failures by category rather than averaging away the difficult cases. This gives an OpenAI model comparison a business purpose: you can see whether a candidate handles the decisions your workflow needs, including when it should ask for more information.

OpenAI documentation.