What we tell customers before turning on an AI agent

We sell an agent that can reply to support tickets without a human in the loop, for a defined set of request types. Every customer has to read and confirm the same short document before we turn it on, and we have never shortened it, even when sales asked us to.

What it says

It lists exactly which ticket types the agent can close on its own (password resets, billing lookups, order status), states that anything outside that list is routed to a human within our SLA, and names the person on our side who is accountable if it makes a mistake. It also says, plainly, that the agent will sometimes be confidently wrong.

The part that matters most

The exclusion list is longer than the inclusion list, on purpose. Refunds over $200, anything mentioning legal action, and any message flagged as distressed by our sentiment check always go to a person. We would rather turn away a sale than let a customer discover the boundary by getting a bad autonomous response to a serious problem.

Why we do this even though it slows down onboarding

Two customers walked away after reading the document, both because they wanted full autonomy on refunds and we would not offer it. We think that is a good outcome. The customers who stay understand exactly what they bought, which means far fewer angry calls when the agent correctly declines to handle something.

If you are building anything similar: write the exclusion list before the feature list. It is a better test of whether you actually trust what you are shipping.