Why the Support Triage Bot Never Free-Writes an Answer

The tempting shortcut

Wire a webhook straight into an AI node, hand it the customer's message, let it answer. It works immediately, it feels impressive in a demo, and it's the fastest possible thing to build.

Why that's a liability, not a feature

An AI model that's allowed to answer anything will eventually invent a policy, a price, or a promise that was never actually true. For a bot speaking on behalf of a real business, that's not a minor inconvenience — it's the business's name attached to something it never said.

The actual design

The FAQ this workflow answers from lives as a plain array of question/answer pairs directly in the workflow. The prompt tells Gemini explicitly to answer only from that list, nothing else, and to return strict JSON with an `answerable` boolean rather than free-form prose.

If the question isn't covered — a billing dispute, a complaint, anything ambiguous — the workflow doesn't attempt an answer at all. It escalates to a human with an AI-written summary of what the customer actually needs instead.

The result

Tested against a real pricing question, it answered correctly, grounded word-for-word in the actual FAQ text. Tested against a real billing complaint, it correctly refused to guess and escalated with an accurate summary instead of a made-up answer.

The measure of a good support automation isn't how much it can answer. It's how reliably it knows what it can't.