Policy Answers in n8n

What it does

1. A staff member types a question into a simple form.
2. Crisis words get the 988 line and alert the on-call clinician. A question about a medication or a dose is refused with ask a clinician.
3. For a real policy question, it finds the closest sections of the current policies.
4. The AI only points at sentences. The answer is built from those exact words, with the policy, the section and the version shown.
5. If the policies do not cover the question, it says so, names who owns that policy and adds the question to a list of gaps.

What keeps it safe

• The AI writes nothing a person reads, so it cannot invent a rule. My first version let it write the answer, and 10 of 21 trick answers got past my checks, so I threw it out.
• Only the current version of each policy can be quoted. An old one cannot.
• If the AI fails, the answer says it cannot answer reliably right now and shows the closest sections. It is never logged as a policy gap.

How I tested it

• 83 questions against an answer key: 44 of 44 one section questions, 10 of 10 two section questions and 13 of 13 uncovered questions were right. Uncovered ones got not in the policies, not a guess. 5 of these questions were reworded after a first miss, and in their first wording all 5 came out wrong.
• 30 new questions I had not tuned for, run with the real AI: 30 of 30 got the right kind of answer, all 6 the policies do not cover came back not in the policies, and both medical questions were refused. 18 of the 22 answerable ones cited every section expected; the 4 misses were two section questions.
• No sentence outside a current policy, no paraphrase and no old version ever appeared. 387 rule checks passed.

Built with

• n8n
• Gemini
• Google Sheets