Intercom

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Do not build this if the AI agent is resolving real tickets. That deflection is the product now.

An agent answering from your help centre before a human sees the conversation, and doing it well enough that customers do not immediately ask for a person, is the thing being sold. Building it means retrieval over your own documentation, a model, careful refusal behaviour, and measurement of whether it actually helped — and getting the refusal behaviour wrong is worse than having no agent at all.

WHEN TO KEEP PAYING
- Enough volume that deflection is measurably cheaper than headcount
- You would rather somebody else own the failure mode of a bot confidently telling a customer something untrue

WHEN TO BUILD
- The messenger and the inbox are ordinary work, covered under Chatwoot and Crisp elsewhere in this catalogue

IF YOU BUILD THE AGENT, THESE ARE NOT NEGOTIABLE
- **Answer only from retrieved documentation, never from the model's own knowledge.** If nothing relevant is retrieved, say so and hand over. A confident answer assembled from nothing is the failure that costs a customer
- Show the sources. A cited answer is checkable; an uncited one is a rumour
- An obvious, always-present route to a human, in one click, with no discouragement. A bot that makes escaping difficult is the thing people write about publicly
- Hybrid retrieval — full text and vectors together — because vectors miss exact product names and identifiers
- Record every question, what was retrieved, what was answered, and whether the person then asked for a human. That last field is the only honest measure of whether it works
- Never let it take an action. Answering is one risk; refunding, cancelling or changing an account on a model's judgement is another entirely

AND THE PART THAT IS NOT SOFTWARE
The agent is only as good as the help centre behind it. Questions that retrieve nothing are your writing backlog, and half the deflection comes from writing five articles rather than from any model.

THE ONE-LINE VERSION
Buy the agent, or build it with citations, a refusal, and a one-click human. And fix the help centre first — that is where most of the deflection actually is.