Example / Private knowledge
A private knowledge helper answers from the right documents.
An internal assistant is not a folder search box with a chat skin. It needs to know who is asking, which documents they may see, and when an answer should become a handoff instead of a guess.
Direct answer
A private knowledge helper should authenticate the person, retrieve only the sources they can access, cite those sources in the response, and route unresolved work to the team that owns the decision.
Reference workflow. The access, retention, and audit requirements must be defined with the organization that owns the documents.
01 / Identity first
Make access part of the request.
Document access is not an afterthought. The search request must carry the identity, tenant or team scope, and any role constraints needed to filter the result set before text is returned to the AI host.
That keeps the policy in the application boundary, where it can be reviewed and changed without rewriting conversational instructions.
Model Context Protocol · authorization02 / Answer with evidence
Citations make the answer checkable.
When a response relies on an internal handbook, contract, or runbook, show the document title and the relevant section. The user can decide whether the source fits the situation instead of treating the answer as an unexplained authority.
This also makes source maintenance visible: an outdated answer points back to the document that needs review.
03 / Keep actions separate
Knowing a policy is not permission to change a system.
The helper may explain the approved path and offer a next step, but a write action belongs behind its own named tool, validation, and confirmation. A policy answer should never silently become a payroll, access, or customer-record change.
This distinction lets teams release useful knowledge access before they decide which operational actions are safe to expose.
04 / Measure unanswered work
Use gaps as an editorial queue.
Track the questions that had no source, produced conflicting sources, or required a person. Those patterns identify missing documentation and candidates for a narrower, approved workflow.
The point is not to maximize chat volume. It is to make the next answer more reliable and the next product action more deliberate.