Blog · AI for small and midsize businesses
When AI does not know: design useful answers without guessing
An assistant should be useful when evidence is missing. Define when to ask a specific question, explain the limit of available information or refer the case with context. Test these paths deliberately instead of relying on a generic instruction to avoid guessing.

In this article
An employee asks whether an unusual expense is covered. The policy describes ordinary expenses but says nothing about this case. A confident answer would invent company authority. A useful assistant explains the evidence limit and helps the employee reach the right next step.
Distinguish three kinds of uncertainty
| Situation | Useful behavior | What to avoid |
|---|---|---|
| Question is ambiguous | Ask for the missing detail needed to search correctly. | Choosing an interpretation without saying so. |
| Evidence is absent | State what the approved sources do not establish and refer. | Filling the gap with general industry practice. |
| Sources conflict | Identify the conflict and route it to the source owner. | Selecting the most plausible wording as authority. |
Make clarification specific
“Please provide more information” shifts the problem back to the user. Ask for the field that changes the search: location, date, product variant or type of request. Explain why it matters when that is not obvious. Avoid collecting personal information that the task does not require.
In a hypothetical policy assistant, “Does this apply to me?” may need the employee's location or contract category. The assistant should ask only the permitted relevant detail. It should not request a complete personnel file to answer a general procedure question.
Refer with context
A useful referral names the approved team or route, summarizes the unresolved question and identifies any relevant source. If no service expectation is documented, do not invent a response deadline. Distinguish “I did not find supporting information” from “the company does not allow it”; those statements mean different things.
For an undocumented expense, a suitable draft might explain that the available policy does not establish approval for the described situation and direct the employee to the policy owner. It can list the facts already provided so the person does not need to restart the explanation.
Test fallback quality
Create cases with missing evidence, ambiguous requests and conflicting sources. Label the expected behavior before testing. Measure unsupported answers, useful clarifications, correct referrals and unnecessary refusals. An assistant that refuses everything may avoid some errors while failing its business purpose.
Microsoft's evaluation guidance includes queries outside source coverage. The NIST profile provides a broader basis for evaluating limitations and operating risk.
Assign follow-up for unanswered questions
The source owner reviews repeated gaps and decides whether to clarify documentation, add an approved route or preserve an intentional boundary. The technical team checks retrieval failures separately; a document may exist even when the search missed it. Track categories without retaining unnecessary sensitive conversation details.
Review referrals with the receiving person. Can they resolve the case from the transferred information, and does the user understand what remains undecided? Track referrals that return for missing details. Improving that handoff can be more useful than increasing the number of questions the assistant attempts to answer. Keep the unresolved decision visible until someone authorized makes it.
Bring examples of unsupported answers and difficult referrals to a knowledge assistant evaluation to design fallback behavior that helps people continue their work.