Practical information and fictional examples for scoping, not claims of client work, measured results, independent assurance or certification.
Begin with the use case and who is allowed to ask
A useful knowledge assistant needs a defined audience and a bounded purpose. List the questions it should answer, the decisions it should not make, and the categories of documents it may use. “All company knowledge” is not a suitable initial source policy. Separate the model’s behaviour from the information it is authorised to retrieve. A reliable demo with one user group does not establish safe retrieval for every employee.
Review source rights and information boundaries
For each proposed collection, identify its owner, provenance, currency, access restrictions, licence or contractual concerns, and rules for corrections. Pay special attention to mirrored documents, exports, cached text and indexes, because removing access to a source folder may not remove every derived copy. Ask how a user’s permissions are checked at retrieval time, not only when material is first ingested.
Test retrieval with positive and negative cases
Choose synthetic or approved test questions across user roles. Confirm expected citations and version context, and deliberately test records a restricted user must not retrieve. Record what was in scope, which source version the test used and any unresolved connector limitations. A model’s refusal is not a substitute for enforcing retrieval permissions. Define how a subject-matter owner corrects an inaccurate or obsolete answer and when the change reaches the index.
Create an ongoing source stewardship routine
Someone must own new source approvals, version changes, access changes, incident escalation and retirement. Record when the knowledge base is refreshed and how stale material is detected. Define the review threshold for high-consequence answers, particularly when a person might interpret output as legal, clinical or financial advice. This is an AI-data governance workstream, not a certification of an entire AI system.
What buyers usually ask
Is this an AI security audit?
No. It addresses the selected data governance controls and flags other specialist assessments when necessary.
Do we need to move our files first?
Not automatically. First examine ownership, permissions and change paths for the sources in scope.
Can a chatbot be trusted if it cites documents?
Citations help review answers but do not prove source currency, authorisation or accuracy.
Start with one decision, not a platform purchase.
We can discuss a scoped engagement using the context above. Do not send confidential information through an initial inquiry.