AI Assistant for Corporate Rules

Only from approved knowledge bases and corporate repositories. Search returns excerpts from current documents along with versions, and the response shows the document, section, and rule version the conclusion is based on. The model explains the rule in plain language, but the source can always be opened.

No. Access rights are inherited from the corporate environment: the system first determines which documents and sections are available to a specific employee, and only then searches for an answer. In the pilot, the absence of closed documents in the results for the test role is checked with a separate acceptance criterion.

The request is sent to a person. Low confidence, version conflicts, or a missing rule create a task for the knowledge owner; the assistant does not invent the missing rule on its own. It also shows which questions are not covered by documents at all.

A control sample of questions is fixed before launch, then we measure on the same data set: source accuracy (whether the correct document and its current version were selected), answer completeness (whether all mandatory rule conditions were named), access security, and the share of answers that the knowledge owner corrected or sent back for revision.

We start with one type of recurring question: one knowledge owner, one document set, one working channel, and verifiable acceptance criteria. This makes the impact visible in a limited scope, and the knowledge base can then be expanded using a proven approach.