Yes. An AI knowledge assistant can be designed to tell staff when it cannot find enough evidence to answer. It should explain the limit, offer a useful next step and capture the unresolved question for the document owner. It should not fill the gap with a plausible company rule.
The wording matters: “I could not find an answer in the approved sources searched” is more defensible than “The knowledge base contains no answer.” A failed search does not prove absence. The practical goal is a knowledge-gap backlog that helps your business identify what needs writing, clarification or a search repair.
Define what counts as an answer
Start with a proposed rule: an answer needs a relevant passage that supports the requested instruction and applies to the situation described. A document mentioning deliveries does not necessarily answer a question about collecting an order after closing time.
OpenAI documents file search as retrieving information from uploaded knowledge bases through semantic and keyword search. Its examples also show file citations. These are building blocks for an evidence-led assistant; they do not establish that a particular business question has been answered correctly. Source: OpenAI file search
For each substantive instruction, ask whether a staff member could open the cited document and find its basis. Check relevant conditions, such as branch, procedure version or type of request. A citation to a broadly related file should not make an unsupported answer acceptable.
This is a scoped use for custom AI agents. The business must define the evidence rule and escalation process as part of the proposed workflow.
Distinguish an unanswered question from a failed search
Use separate outcomes because each needs a different repair:
- Supported: relevant guidance answers the question.
- Partly supported: guidance answers one part but leaves another unresolved.
- Clarification needed: a missing detail could change which guidance applies.
- No supporting passage found: the completed search did not establish an answer.
- Search unavailable: the search failed or the required material was not ready.
- Conflicting evidence: relevant sources disagree and need owner review.
These are proposed workflow labels. They prevent a technical failure from being recorded as a missing policy.
Before declaring a possible gap, check whether the expected document was included, ready for searching and within the permitted scope. OpenAI's documentation instructs developers to check file readiness before use, and warns that limiting retrieved results can reduce answer quality. A narrow search can therefore warrant investigation before someone commissions new content. Source: OpenAI retrieval readiness and result limits
Ask for clarification only when it could resolve the uncertainty. “Which collection point?” is useful when procedures differ by location. Asking staff to rewrite an already clear question merely delays the handover.
Give staff a clear response and a real next step
An unresolved answer should state what was searched, what remains unsupported and what happens next. Keep it short enough for someone who is trying to complete a task.
A proposed response is:
I could not find instructions for collecting an order after closing time in the approved collection guide. The guide covers collections during opening hours. Please check with the collection supervisor before making arrangements. This question has been recorded for the guide owner.
The final sentence is appropriate only after the record has actually been saved. If saving fails, tell the staff member that the question was not recorded and give the agreed manual contact route. Do not promise a reply time that nobody has accepted.
Where part of an answer is supported, present that part with its source and label the remaining gap. Avoid blending a documented step with an invented exception.
For wider workflow choices, the comparison of AI agents and automation helps frame where interpretation sits alongside fixed routing rules.
Reusable knowledge-gap backlog
Use this record in a shared list, spreadsheet or work queue. The fields and handling rules below are proposed starting points; adjust them to your document ownership process.
| Field | What to enter |
|---|---|
| Record ID | Unique reference for this gap. |
| Staff question | Original wording, with unnecessary personal or customer details removed. |
| Relevant context | Branch, procedure, product or version needed to interpret the question. |
| Search outcome | Partly supported, clarification needed, no supporting passage found, search unavailable or conflicting evidence. |
| Evidence checked | Document titles, versions and relevant passages; state when none were retrieved. |
| Unresolved point | The exact instruction or fact that remains unsupported. |
| Document owner | Named person or responsible role; use “unassigned” until confirmed. |
| Next action | Clarify the question, repair retrieval, add guidance or resolve conflicting sources. |
| Related record | Existing gap ID where the question repeats; otherwise leave blank. |
| Occurrences | Count linked requests without treating repetition as proof of importance. |
| Work status | New, assigned, awaiting clarification, awaiting approval, ready to recheck or closed. |
| Closure evidence | Approved document location, search recheck result and reviewer confirmation. |
Handling checklist:
- Confirm whether retrieval completed before classifying the gap.
- Preserve context that could change the answer.
- Check for an existing record before creating another.
- Assign an owner and a concrete next action.
- Keep proposed guidance separate from approved guidance.
- Recheck the original question against the approved update before closing.
Structured Outputs can constrain a model's response to a supplied schema, which can help keep these record fields consistent. Treat that as a formatting aid: the evidence, classification and owner assignment still need validation. Source: OpenAI Structured Outputs
Saving the record requires an implemented connection to your chosen queue. OpenAI's function-calling guide describes the application executing a requested function and returning its result; the model's request alone is not confirmation of a saved record. Source: OpenAI function calling
Worked cases: answer, gap, ambiguity and duplication
These scenarios are hypothetical, with proposed handling rules.
Ordinary supported question. A staff member asks where to send a damaged-packaging report. The approved warehouse guide names the receiving supervisor and explains which details to include. The assistant returns those instructions with the supporting source. No gap record is needed unless the staff member identifies an omission.
Missing instruction. Someone asks whether a collection can be arranged after closing. The guide describes normal collections but says nothing about exceptions. The assistant explains the limit and refers the decision to the supervisor. The backlog records “after-hours collection authority is unspecified”, allowing the document owner to decide whether guidance should be added.
Ambiguous question. A staff member asks, “Who approves this return?” without identifying the return type. The assistant asks whether it concerns damaged goods or an unwanted order. If the staff member cannot provide that detail, the record stays awaiting clarification. The document owner should not write a universal approval rule from incomplete context.
Duplicate request. Another staff member asks, “Can customers fetch orders once the counter closes?” If it concerns the same collection procedure, link it to the existing after-hours gap and retain the wording. If it concerns a different location with different arrangements, keep a separate record. Similar language alone is insufficient for merging.
Duplicate documents. Search retrieves two copies of the same guide. Neither covers the exception. Two matching citations do not strengthen the missing instruction. Record the gap once and ask the owner which copy should be maintained.
Turn the backlog into approved, searchable guidance
Review records by the work they block and the consequence of guessing. Frequency is useful context, but a frequently asked convenience question need not outrank an unclear instruction with serious consequences.
The owner should first confirm the diagnosis. Perhaps guidance exists under an unfamiliar name, or a file was omitted. Repair discovery in that case. If the instruction is genuinely missing, the appropriate person decides the rule and approves the wording. The assistant can prepare the question and evidence for that decision.
After publication, search the updated material using the original question and a natural paraphrase. Confirm that the returned passage answers the unresolved point. Close the record only when the guidance is approved, searchable and correctly used. An uploaded file alone does not complete that loop.
The custom AI agent glossary and custom agent workflow guide provide context for defining these responsibilities.
FAQ: handling unanswered staff questions
Can we use a confidence score to decide when to answer?
Use supporting passages and applicability checks as the main decision basis. If you introduce a score, test its proposed threshold against questions your document owners have assessed. A number should not override missing evidence or conceal conflicting instructions.
Should every unanswered question become a new document?
No. Some records need clarification, a search repair or a small addition to an existing guide. Ask the owner to identify the smallest useful correction. Merge repeated gaps where context matches, while retaining the different ways staff ask about them.
Can the assistant email the document owner automatically?
That requires a configured workflow and confirmed delivery handling. Start with an agreed internal queue and notification rule. Before deployment, check current product, account, plan and regional eligibility. Tell staff a notification succeeded only when the application has evidence of success.
If your business needs an assistant that admits evidence gaps and gives document owners useful follow-up work, explore AI automation and get in touch to discuss the questions, approved sources and backlog your team needs.