Will an AI co-host promise your guest a refund?

No. A refund is never the AI's decision — not at any confidence, not on full autopilot, not when the guest asks nicely. You have probably seen the stories: a company's chatbot promises a discount or a refund nobody approved, and the company ends up honoring the words its software said. That failure mode is exactly what a host imagines when they think about an AI answering guests, so it deserves a concrete answer, not a reassurance. Here is the machinery that makes a money promise impossible to automate.

A refund is a decision, not a fact

Most guest questions are questions of fact. The wifi password, the check-in time, where to park — these have one right answer, written down by you, and an AI that answers only from your facts can handle them safely. A refund request is a different kind of question. It has no answer in your house rules; it has a judgment call attached to your money, your cancellation policy, and the particulars of one stay. The system is built around that distinction: facts can be answered, decisions can only be surfaced. There is no fact sheet entry that would let a draft reply say yes to money.

The list that always escalates

Every incoming message is screened for a fixed set of sensitive topics: refunds, money-back requests, chargebacks, disputes, cancellations — and alongside the money list, safety, injuries, medical issues, anything legal, and discrimination. A message that touches any of these routes to you, full stop. This is a stricter rule than the one ordinary replies live under. A normal answer must clear an 85 percent confidence bar before autopilot may send it; a sensitive topic never sends at any confidence. One hundred percent certain is still your call, because the problem with a refund reply was never that the AI might word it badly — it is that the decision was not the AI's to make.

Two checks, and the stricter one wins

"The model decides what counts as sensitive" would be a weak guarantee — the model that wants to be helpful would also be grading its own restraint. So two independent checks run on every message. The model reports whether the topic is sensitive, and a separate, deterministic keyword screen reads the guest's own words for refunds, disputes, cancellations, and their relatives. The screen can only raise the risk level, never lower it. If the model says the message is fine and the screen says it is about money, the screen wins, the message escalates, and the disagreement is recorded so it can be counted later. An over-helpful model cannot talk its way past a regular expression.

What the guest hears, and what you see

The guest is not left hanging. They get the honest version of the situation: the host is being looped in, and they will hear back shortly. You get the whole thread immediately, with a drafted acknowledgment attached — and the decision itself stays where it belongs. If you want the reply to carry your exact words, it can: the approval loop accepts a reply you write yourself, and only you can approve anything at all. Guests generally read "let me get the host on this" as a careful operation, not a stall — and since the question lands in front of you with context attached, the wait is usually minutes.

The honest cost

This design means an AI co-host does not take money questions off your plate, and you should walk away from any tool that claims otherwise. What it takes off your plate is the volume around them: the wifi, the parking, the early check-in, the pool heater — the twenty routine messages that bury the one that actually needs you. The escalation is the product working, not failing. And the boundary is adjustable in only one direction: you can reserve more categories for yourself — some hosts keep the personal welcome, some keep anything touching reviews — and those route to you no matter how confident the draft is. Nothing you can configure makes a refund auto-send. New accounts start with every reply held for review anyway, so early on you will watch this boundary work before any reply goes out on its own.

You can test the boundary yourself before trusting it: paste your listing into the demo and ask for your money back. Watching the reply decline to decide — on your listing, in your browser, before any signup — says more than this page can.