Our office chat bot answered confidently from a policy that had already changed

Our team has an internal chat bot that searches company documents and answers routine questions. It is useful for finding travel forms, expense categories, and the correct page of a handbook. Last Tuesday it told a coworker that manager approval was required before booking any hotel.

The answer included a link. The linked page had been replaced two months earlier.

The current policy allows booking within a stated limit without advance approval. The old page still existed in an archive folder, and the bot turned it into a clean, present-tense instruction. My coworker lost a day waiting for approval that nobody needed to give.

This was not a dramatic hallucination. Every sentence could be traced to a real document. The failure was freshness and authority: the bot found words but did not know which version governed.

We are considering a modest rule set. Policy answers must show the source title, owner, and effective date. Archived documents should be excluded or visibly labeled. If two sources conflict, the bot should present the conflict instead of choosing the smoother answer. For expenses, security, employment, or legal obligations, the user should confirm against the current official page before acting.

The hard part is that warnings can become wallpaper. If every response says “verify this,” people will stop seeing it. I would rather design the result so the evidence is easy to inspect and stale sources are harder to retrieve.

What would make an internal AI answer trustworthy enough for routine use without pretending it is the policy owner? Is the critical control document hygiene, source dates, a narrow approved index, or a human escalation path when the answer affects money or obligations?