Give your chatbot a harness before you give it your knowledge base
An AI assistant that answers from company knowledge is useful right up until it answers from the wrong permissions, or nobody can say what it told someone six weeks ago.
Open-ended is the risky default
An assistant wired straight to a general-purpose model with broad data access will eventually answer a question it should not, using a document the person asking could not normally see. That is not a model failure. It is a harness that was never built.
Answer only from what is governed
- Query only a governed knowledge index, never the open internet
- Filter responses by the asker's own access level, not the index's
- Tag every answer with its source document and a confidence signal
- Escalate sensitive questions to a person instead of answering them
An audit trail, not just an answer
Every question, every document touched, and every answer returned stays logged. That turns “the chatbot said something wrong” from a guess into something a compliance team can actually trace back and explain.
What this takes
Guardrails to scope and filter, Policy to decide who sees what, Audit to keep the record. None of it depends on which model answers the question.
This is Guardrails and Policy working together at the front door, with Audit keeping the record behind them.