Help staff recognise when an AI answer is uncertain

What this means for your practice

Specify how unsupported questions are handled and provide a useful human route. A plausible sentence is not evidence that the answer is grounded in an approved source.

A practical conversation with your supplier

Ask an intentionally unsupported policy question and an ambiguous question. What does the user see, and where can they get help? Inspect the supporting source when an answer is given. Agree how reported mistakes are reviewed so the same unsupported response does not remain an accepted part of the working process.

Plan the first improvement

Write down what a good no-answer response should help the staff member do next. It might identify the missing information or direct the question to the owner of the relevant procedure. Simply saying sorry may stop an unsupported claim but still leave the work blocked. Give users a way to report an answer that appears unsupported and keep the review responsibility clear. Include questions that look close to an approved topic but ask for a detail the source never gives.

These are useful tests because a plausible answer can sound more convincing than an obvious mistake. Also test conflicting sources and ambiguous wording. Your acceptance decision should consider how the assistant handles those situations, not only straightforward retrieval. Explain to staff how to inspect supporting material and when to use the human route. Review repeated unanswered topics as possible gaps in the knowledge base, while keeping the decision to create new guidance with the accountable owner.

How to check the result

Measure unsupported answers and successful handovers to the responsible team.