An unanswered question is useful feedback when it reaches the right owner with a specific gap. Do not collect an unstructured pile of failed prompts or turn every refusal into an automatic new knowledge article.
Unanswered questions have different causes
A knowledge assistant may abstain because the source is absent, access is denied, extraction failed, retrieval missed evidence, or the question itself is ambiguous. These require different remedies.
For a proposed gap queue, include a minimised question, supported scope, safe source references, failure category, impact, owner, and disposition. Avoid retaining unnecessary personal context from the original interaction.
Do not label denied access as missing content to be copied into a public collection. The remedy must preserve permissions. Similarly, a retrieval bug should not cause duplicate source writing when the approved answer already exists.
Triage a fictional reimbursement question
An employee asks about a new expense type. Review finds that no approved policy addresses it. The gap belongs to the policy owner, who can decide whether new guidance is appropriate. The assistant should not draft an answer and treat publication as automatic.
Another question fails because the relevant paragraph was split from its heading. That belongs to ingestion or retrieval engineering, not the policy team. The original source can remain unchanged if the representation is repaired.
| Gap category | Responsible next action |
|---|---|
| Approved content absent | Content-owner decision |
| Existing source not retrieved | Retrieval investigation |
| Extraction incomplete | Parser or source-format repair |
| Permission denied | Access-policy process, not duplication |
| User question ambiguous | Clarification or interface improvement |
Test whether a reviewer can select the appropriate route from a fictional item. A generic “AI could not answer” label does not provide enough diagnostic context.
Close the feedback loop with evidence
When the issue is resolved, add or update a reviewed regression fixture. Verify retrieval and final claims before marking the gap complete. A document being uploaded is not proof that the assistant can answer the original question correctly.
Gap queues can grow beyond staffing capacity. Prioritise by recurring need and consequence, and keep low-value or unsupported requests explicitly disposed. Track useful resolution rather than total collected questions.
This process is proposed. The voice integration case study describes related knowledge workflows without claiming this feedback queue. Pair it with knowledge abstention policy.
Unanswered-question FAQ
Should every unanswered prompt create a document?
No. Diagnose whether the problem is content, access, extraction, retrieval, or ambiguity.
Can the assistant publish its own guessed answer?
Not as approved company policy. Use the authorised content-review process.
What should close a gap?
A verified remedy or explicit disposition, with the original question covered by an appropriate regression test.