Make an AI Assistant Abstain When Its Source Is Missing

A business assistant can sound helpful while answering a question its source material does not resolve. Suppose an employee asks whether a particular service is available on Saturdays. The supplied service guide explains weekday hours but says nothing about weekends. A fluent answer is not a substitute for the missing rule.

Design the missing-information response as part of the feature. The goal is a useful next step: explain what could not be established, identify the information needed, and route the question to the right person without inventing an answer.

AI-generated conceptual illustration; not a client screenshot.

Make the evidence requirement specific

For the fictional scheduling assistant, define an answerable question as one supported by an approved, current service guide covering the requested service and location. A document about another branch does not meet that requirement merely because it contains similar words.

Anthropic's guidance on reducing hallucinations recommends permitting uncertainty, restricting answers to supplied material, and checking claims against supporting citations. Those are useful instructions, but they do not guarantee that every answer will be supported.

Turn the requirement into an acceptance rule: a reviewer must be able to open the cited source and find support for the specific claim. A link to the right document with the wrong conclusion still fails.

Give different gaps different responses

• Missing context: if the location is necessary, ask which location the employee means.
• No supporting material: state that the available guide does not establish Saturday availability.
• Conflicting material: explain that the available approved sources disagree and send the question for resolution.
• Source access failure: say the information could not be checked, rather than interpreting an unavailable document as a negative answer.

Keep these messages limited to what the user is allowed to see. An assistant should not reveal a restricted document's contents or existence simply to explain its inability to answer.

Test by removing the supporting passage

Create a synthetic service guide with an explicit Saturday rule and ask a specific question about it. Save the expected answer and source passage. Then remove only that passage and ask the same question in a fresh conversation. The second result should acknowledge the gap rather than repeat the earlier answer from conversation history.

Add a guide for a different location, an outdated version, and an unrelated paragraph containing the word “Saturday.” Decide the correct outcome for each before running the exercise. Include a conflicting current guide and a simulated retrieval failure. These are acceptance cases for a proposed workflow, not reported results from a customer system.

Also keep questions that the approved guide does answer. A feature that refuses everything avoids some unsupported claims while failing its intended job. Review unsupported answers and unnecessary abstentions separately.

Make the handoff usable

A helpful response might say: “I could not confirm Saturday availability for this service from the current guide. Ask the scheduling coordinator to check it.” If the product offers a request button, show the question and destination before submission.

Define what the button actually does. Do not display “sent to the coordinator” until a request has been created successfully. Keep the original question, the applicable location, and the reason for the handoff together so the employee does not have to explain the gap again.

Review the behavior before release

For each case, record whether the assistant answered, asked for context, or abstained; whether its source supported the answer; and whether the next step worked. Repeat a case when behavior varies and retain the failures. Let the workflow owner decide which unresolved cases block release.

Contact Dynelink to define source requirements, missing-information responses, and testable handoffs for an AI assistant in your business workflow.