An AI assistant needs more than a large collection of documents. It needs relevant, current business information and clear instructions about when it should ask for help. Treat knowledge preparation and answer testing as an ongoing operating process.
Organise information around real questions
Start with the questions customers actually ask. Group product information, service scope, process details and practical support answers so they can be maintained. Remove contradictions and obsolete documents. Record who owns each source and when it should be reviewed. A clear, current answer is more useful than several versions of the same policy with no indication of which one applies.
Define the boundaries explicitly
Write down the topics the assistant may answer, the information it should collect and the situations it should escalate. Decide how it should respond when the information is missing or uncertain. Avoid promising decisions the assistant cannot make. For sensitive or specialist questions, keep the route to qualified people clear. Agree what information the handover should include.
Build a small answer-quality test set
Collect realistic questions, including incomplete wording, ambiguous needs and requests outside the intended scope. Define what a useful answer should contain and what it should avoid. Review the assistant’s response, the source information and the next action. Update the test set when your services change or when support conversations reveal a new gap. Keep humans responsible for reviewing the outcome.
Your practical checklist
- Approved sources selected.
- Content owners assigned.
- Outdated or conflicting answers removed.
- Escalation boundaries documented.
- A practical test set reviewed.
