Will an AI receptionist or chatbot make things up to my customers?

A good setup is blocked from quoting prices or making promises unless those are locked in word for word, and it hands anything outside that to a human rather than guessing. You should also be able to read back what it actually said to customers, especially in the first few weeks of it running.

If a provider cannot show you where its answers come from and what it is forbidden to say, that is the point to walk away, not after something has already gone wrong with a customer.

The businesses that avoid this problem entirely tend to start narrow: a handful of common questions, answered exactly, with everything else routed to a person. Widening it happens later, once you have watched it perform on the easy questions first, not on day one with your whole business handed over at once.

Ask specifically what happens when a customer asks something outside that narrow list. A good answer names the fallback in one sentence: it says it does not know and offers to get a person, rather than guessing to sound helpful. That single behaviour is usually the clearest sign of a setup built properly.

It is also reasonable to ask for a written list of the exact things it is not allowed to say, such as guaranteeing a price or a timeframe it cannot control. A provider who has already thought about this will have the list ready.

It is also fair to ask how often that list gets checked once the tool is live, not just at the start. A setup nobody looks at again tends to drift, quietly, until a customer notices before you do.

A Three Fixes report only recommends tools scoped to what you have told us, and it never invents a detail about your business it was not given.