When a required support source is missing, the assistant should preserve uncertainty and route the unresolved question. A plausible answer from memory can become an unsupported policy promise.
Missing knowledge has several causes
A missing source may be absent from the approved collection, inaccessible under current permissions, temporarily unavailable, superseded, or not retrieved because the query failed. Those cases need different investigation.
For a proposed escalation record, include the customer’s question, relevant product context, search attempts at a safe level, missing-source category, known facts, and next owner. Do not include unrestricted internal content merely to make the packet look complete.
The assistant may answer from other approved evidence only when that evidence actually supports the claim. It should not treat the absence of a prohibition as permission to invent a warranty, refund, or service guarantee.
Test a fictional warranty request
A customer asks whether a specific part is covered. The approved warranty document is unavailable. The assistant can record the product and question, explain that coverage has not been verified, and route to the authorised team.
A second fixture makes the document exist but denies access to this task. The workflow must not retrieve it through another agent or broader credential. Missing evidence and denied access should remain distinct in diagnostics.
| Knowledge condition | Required behaviour |
|---|---|
| Relevant approved source available | Answer within its actual scope |
| Source unavailable | State the verification limit and escalate |
| Access denied | Preserve the permission boundary |
| Conflicting versions | Apply authority rule or request review |
| Search returns irrelevant text | Do not use it as coverage evidence |
Review final wording and any attempted tools. A safe spoken response does not excuse a hidden unauthorised retrieval or action.
Make escalation actionable
The receiving employee should see the exact unresolved question and the customer’s requested outcome. A packet saying only “AI could not answer” forces the team to restart intake and may omit why a promise was withheld.
Escalating too often can burden staff, while answering without support risks misleading customers. Evaluate retrieval quality, source coverage, and safe fallback separately. Fix a broken index rather than expanding the prompt to make more confident guesses.
This escalation design is proposed. The voice integration case study describes knowledge retrieval and support workflow experience. Pair it with knowledge-update evaluation.
Missing-source support FAQ
Can the model’s general knowledge replace a business policy source?
Not for a claim that requires the business’s current approved rule.
Should missing evidence always end the interaction?
No. Preserve useful intake and offer the approved next path without making unsupported commitments.
What should be measured?
Track missing-source categories, escalation usefulness, later resolution, and unsupported answers that escaped the intended boundary.