AI does not fix unmanaged knowledge
If a team has conflicting policies, old files, unclear ownership, and private information mixed with public guidance, adding a conversational interface can make the inconsistency easier to encounter. Knowledge operations come first.
Create a source register
- TopicThe question or process the source governs.
- AuthorityWhich document, system, or person is the source of truth.
- AudienceWho is allowed and intended to use it.
- OwnerWho approves changes and answers ambiguity.
- Review eventThe date or operational change that triggers revalidation.
- RetirementHow obsolete versions are removed from the agent’s source set.
Structure for humans and retrieval
Use descriptive titles, explicit terminology, short sections, ordered steps, conditions, examples, and stable links. State effective dates or versions where they matter. Good retrieval content is usually good human documentation.
Treat questions as feedback
A knowledge gap may require a new FAQ, a clearer procedure, a terminology alias, or a policy decision. Capture the gap, resolve it with the owner, update the source, and retest the answer.
Measure maintenance health
Track unowned sources, overdue reviews, conflicting answers, unanswered high-frequency questions, broken links, and time between a real-world change and the corresponding knowledge update. These are more actionable than a vague claim that the AI is “trained.”