Ground every answer in sources and require citations, check claims against live systems rather than docs alone, and don't let the agent send answers it can't support. Start with human review and loosen it as accuracy proves out.
Controls that work
- Require a citation for each factual claim
- Prefer code, config, and logs over older documentation
- Make 'I couldn't verify this' an acceptable outcome that routes to a human
- Review a sample of sent answers every week
Signs it's happening
- Answers mention settings or API fields that don't exist
- Reopen rate rises on AI-handled tickets
- Customers quote the AI's answer back when escalating
Every Decimal answer carries citations to the evidence it used. See hallucination and grounding.