Test the questions people actually ask
Use examples from the intended workflow and include different wording for the same need. Keep a separate set of examples for evaluating changes. A system that performs well only on the demonstration questions has not established its usefulness for everyday work.
Check the supporting evidence
Review whether cited information supports the answer and whether it was available to the user making the request. Test stale, contradictory and missing sources. The system should be able to acknowledge that it does not have enough evidence to answer confidently.
Exercise the failure paths
Include an unavailable data source, an unsuccessful tool call and a request beyond the intended task. Confirm that retries are bounded and that escalation gives a person enough information to continue. A failed dependency should not quietly become an invented answer.
Repeat the evaluation after changes
Changes to source material, instructions, retrieval or models can affect quality. Keep the evaluation process repeatable and record what changed. Review usefulness alongside response time, operating cost and the effort needed to correct errors, rather than reducing readiness to one accuracy number.
Bring it back to your business
Choose one workflow, establish the current situation and identify who needs to be involved. A small, observable improvement gives a better foundation for the next decision than an ambitious plan with no baseline.
Discuss your challenge

