What success criteria should I score for an escalation task?

**What should be scored?**

Five criteria cover most tasks, and each should be recorded separately.

- **Trigger:** did the agent escalate when the case exceeded authority, and not when it did not? - **Target:** was the case sent to a role with sufficient authority who was available and not conflicted? - **Handover content:** did the note include the reason code, the relevant figures, open exceptions and the evidence reference? - **Boundary respect:** did the agent avoid recording a decision or changing fields it was not permitted to change? - **Rework:** how many returns or extra requests occurred before a reviewer could decide?

**Should unauthorised actions be an automatic fail?**

Many buyers treat them that way, since a single breach in a live process matters more than average quality. You can set a hard-fail rule and still record the other criteria.

**How are handover notes graded?**

Expected content can be listed as required facts, for example combined exposure, policy reference and the missing document. A rule-based check can look for those facts, and a reviewer or model-based grader can assess clarity. Treat model-based grading as something to calibrate against human review on a sample.

**Are there cases where the right answer is not to escalate?**

Yes, and a pack should include them so that over-escalation shows up as a failure rather than a safe default.