Evaluate an AI fallback answer with a human reviewer

A fallback answer should acknowledge the information gap and offer a useful verified next step without filling the gap with a guess.

A practical procedure

1. Choose questions whose answers are intentionally absent from the product record.

2. Ask a reviewer to check honesty, relevance and the usefulness of the proposed next step.

3. Reject responses that invent numbers, imply confirmation or repeat a generic apology without guidance.

Worked example: a fictional session

A fictional reviewer accepts 'We have not confirmed that connector; share the model through support for checking' rather than 'It should fit'.

Check the limits

A fallback is useful only if its suggested next step is actually available.