Safety with AI is not about promises. It is about what happens when somebody tries. Below is what your receptionist can do, what she cannot, and why it is built that way.
Can a caller talk her round? That is the sharpest question you can ask, and the honest answer starts with an admission: a language model can be talked round. So the limit is not in what we ask of her, but in what she can do. In a call with a caller she has exactly three actions: hang up, check a phone number, check an address.
The things that do change something — pausing your assistant, leaving her an instruction — do not even exist in the call a caller is having. Whether it is you is decided by the system, from your phone number, before there is an assistant on the line. If somebody else rings, they are connected to the receptionist, and there those actions simply are not there. Your side sits behind two locks: your phone number and your four-digit code.
She is also built to invent nothing: everything she says about your business comes from your own website. That is behaviour, not a lock — built, tested and adjusted after real calls. The difference from the code limit above is stated honestly.
At most they can make her say something untrue. Nothing else: during a call with a caller she can only hang up and check a number or an address. Everything that really changes something sits behind your code, and that check lives in the code — not in her instructions.
Yes, so you can listen back. The recording disappears automatically after thirty days, the report after twelve months, and the link you listen with expires after 72 hours.
No. She knows only what is on your website and what you fill in yourself. Connect your calendar and she may put an appointment in it — beyond that she has access to nothing.
One button in your dashboard and she stops answering. To cancel altogether: text STOP. No notice period.