I stole this from LinkedIn.
I support a call center and we're about to implement an AI agent. We're paying for a model that essentially can talk and has "learned how to learn", but is otherwise dumb. It's trained on a very small amount of information, anything we'd give to a real agent, plus the public info on our website.
The result of this should be a bot that says, "I don't know, should I transfer you to a real person?" a lot, but should hopefully never hallucinate or teach someone how to build a bomb or something.
Dunno how others do it though