What to check in an AI-generated answer

A fluent answer can still be wrong. The reader needs to see the material it relies on and know when to ask for another review.

Start with the source

If an application answers from company documents, identify which documents it may use and who is allowed to see them. Show the passage or record that supports the response. An answer without a relevant source should be treated differently from one grounded in an approved document.

Test the difficult cases

Before release, assemble representative questions, including missing information, conflicting records and requests outside the intended scope. Read the answer alongside each retrieved passage. An answer can sound certain while omitting a condition or using a superseded policy. Record those examples so the team can repeat the checks after the source material changes.

Decide where the person steps in

A draft for an internal team may need a different review from a message sent to a customer. Establish who checks the output, how an uncertain answer is flagged and how a user can correct the underlying material. The review should follow the consequences of an error.