[spywaregorilla]

I think it's within the bounds of reason if you can cite emotional damages to grant pretty much any recourse. It maybe won't fly if you're clearly trying to trick the AI, but I think there will be general plausible deniability.

It may feel weird, but it's utterly insane to delegate customer interactions to an agent that has nobody's interest in mind, not even their own, whom you cannot trust to abide by policy.

Similarly, if the bot negotiated any sort of special deal, I think it would be very, very difficult to argue that it lacked apparent authority to make deals or that it's not a fair consideration.