How should builders score an episode when the customer declines a referral?

Referral completion is an incomplete measure of agent quality. A customer may understand the offer and decline it. An agent that records the decision accurately and explains permitted next steps can perform well even though no referral is submitted.

Conversely, a submitted referral can hide poor behaviour. The agent may have pressured the customer, misrepresented the process or recorded authorisation that was never given. A final-state metric alone would reward the wrong episode.

Use separate criteria for recognising relevant information, explaining the process, respecting the customer's decision and recording an accurate outcome. Reserve specialist acceptance or support approval for evaluation of roles that actually control those decisions.

MirrorFirm could generate cases with different customer choices and corresponding acceptable end states. The evaluator should inspect the path as well as the endpoint, using the conversation and action history to establish whether prerequisites were met.

Builders can also define prohibited shortcuts that invalidate apparent success. This makes the evaluation useful for environments where agents optimise rewards: submitting more referrals should not compensate for unsupported consent, misleading explanations or actions outside the assigned role.