A business knowledge assistant helps staff find and use approved internal information through questions and source-linked answers. Its value depends on document quality, permission boundaries and a clear response when evidence is missing. Zoha can scope an assistant around a defined knowledge collection, with evaluation examples, content ownership and controlled access.
Updated
On this page
Choose a collection with a clear owner
Start with a specific information problem: staff struggle to find a current procedure, product details are scattered across files, or support teams repeatedly ask the same internal questions. Name the collection and its owner before connecting a large shared drive.
Remove obsolete versions and identify which document takes precedence when two sources disagree. A knowledge assistant cannot reliably settle an unresolved business policy. The initial collection should contain material the intended users are allowed to see, with a process for keeping it current.
Apply permissions outside the answer prompt
A useful answer is still a problem if it reveals information the user should not receive. Define user groups and the sources each group may access. The retrieval and application layer should enforce those boundaries; asking a model to keep a secret is not an adequate permission mechanism.
Consider a hypothetical staff assistant with general procedures and restricted personnel files. Ordinary staff should not be able to retrieve the restricted files through a summary, a source link or a cleverly phrased question. Test those boundaries with suitable sample material before allowing real restricted information into the system.
Specify what a supported answer looks like
| Question type | Expected behavior | Review method |
|---|---|---|
| Answer exists in a current source | Answer with a relevant source reference | Reviewer checks meaning against the cited passage. |
| Sources conflict | Identify the conflict or refer to the owner | Reviewer confirms the unresolved policy is not silently chosen. |
| No supporting material | Say the collection does not support an answer | Test with questions deliberately absent from the collection. |
| User lacks permission | Do not reveal restricted content | Test source retrieval and links across role boundaries. |
| Document is replaced | Use the approved current version | Check update and removal behavior after a controlled change. |
Source references help a person verify an answer, but a citation alone is not proof that the claim is correct. The review should check whether the cited material actually supports the response.
Test a useful set of business questions
- Include frequent questions and realistic variations in wording.
- Add questions with similar terms but different business meanings.
- Test missing answers, conflicting documents and outdated examples.
- Check the languages and document formats included in the scope.
- Measure agreed answer quality and permission outcomes using human review.
- Record failures and update sources or behavior before widening use.
NIST's AI Risk Management Framework is a useful reference for organizing AI risk and evaluation responsibilities. The practical acceptance criteria still need to come from your business. A demonstration with a few easy questions should not stand in for review of the collection's difficult cases.
Plan for changing documents and questions
Define who adds, replaces and removes sources, how updates become available and how failed answers are reported. Keep access administration and recovery information with the agreed business owners. Review retention of questions and responses because staff may put sensitive information into a query.
We can scope a pilot, its source set, access model and review process before expanding the assistant. Supported formats, integrations and ongoing maintenance should be explicit. The handover should explain how your team checks an answer, reports a problem and keeps the underlying information useful.