Define the decision and cost of error
Specify what the person must decide, which information is relevant and what a wrong recommendation could cause. Agree the criteria for a useful result and which situations are outside scope. A narrow comparison task is more assessable than a request for an AI system to make every business decision.
Keep evidence and approval together
The workflow should distinguish observations, missing information and proposed recommendations. Use examples to assess whether the output helps the decision owner and whether uncertainty is represented honestly. Consequential actions need the agreed human review and approval boundary; the model receives no independent authority to execute a recommendation.
Bring decisions and counterexamples
Share approved inputs, past decision scenarios and examples of unacceptable recommendations. Name the reviewer and the next action after approval. We assess whether rules or bounded AI are appropriate and verify against agreed criteria, without promising correctness for unfamiliar cases or replacing professional judgement.