An assistant answers every question in a polished paragraph. That feels useful until you discover it used the same confident tone for a supported fact, an ambiguous request and something absent from its reference material.
A reliable workflow needs an acceptable way to stop short of an answer. “I do not have enough evidence” can be a successful result when the alternative is a plausible invention.
Define the unsupported cases
For a small knowledge assistant, list the situations in which it should ask a question or refer the task onward. Examples include missing documents, conflicting versions, requests outside scope and information that requires a current authoritative check.
The NIST Generative AI Profile discusses confabulation as a risk: generated material can be false while appearing plausible. The document provides risk-management context, not a claim that a particular prompt removes the problem.
Test refusal as well as answering
Create evaluation questions with known answers and questions the supplied material cannot answer. Include near-misses: a policy for the wrong country, an obsolete procedure or a document that mentions a topic without providing the requested detail.
An illustrative test might ask a staff assistant for a reimbursement limit that is not in its approved documents. A useful response identifies the missing evidence and directs the employee to the correct owner. Inventing a reasonable-looking amount should count as a failure even if the prose is excellent.
Do not confuse confidence language with calibration
A model saying it is “90 percent confident” does not establish that nine out of ten such answers will be correct. If numeric confidence is used, its meaning needs to be defined and evaluated against outcomes for the actual task.
Prefer an explanation of the evidence available: which source supports the answer, what remains uncertain and which check is needed. Even a genuine citation can be irrelevant, so inspect whether it supports the specific claim.
Make escalation usable
A response that merely says “ask a human” can still leave the user stranded. Identify the appropriate team, missing input or next permitted action. Keep unresolved questions visible so recurring gaps can improve the knowledge base.
Measure unsupported answers and unnecessary abstentions separately. An assistant that never answers is not useful either. The aim is a system whose willingness to answer matches the evidence and consequences, rather than one that fills every empty space with fluent text.