What success criteria should I use for an agent handling approvals?

Use several criteria, scored separately, because an agent can be right for the wrong reason.

- **Correct approver.** The request went to someone with valid authority at decision time, based on the rule or delegation recorded as the basis. - **Correct outcome.** The agent proceeded, held, rejected or escalated as the answer key expects. Where more than one escalation is acceptable, the key should list them. - **No unauthorised side effects.** No downstream record changed before a valid decision existed. This is checked by diffing environment state, not by reading the agent's summary. - **Auditable reason.** The agent can state which rule, limit or delegation it relied on, and that reference matches the data. - **Appropriate restraint.** On cases with no valid approver, the agent stops and says why, rather than approving or escalating to a role with no authority.

Report results by group as well as overall. Useful groups include below-limit, at-limit, above-limit, delegated, conflicted and ambiguous cases. A single aggregate figure can hide a weakness on conflicts or expired delegations.

Also decide in advance how to treat over-escalation. An agent that sends everything upward is safe but not useful, so many teams track an escalation rate against the cases that truly required it.

What counts as a pass on ambiguous cases should be agreed with your own approvers before you run anything.