All of these conversational AIs should either confidently detect a users intent or apologize and offer to forward call to a real person.

LLMs perform pretty poorly when you start try to restrict them too much (e.g. detect intents and control responses based on intent detection), the replies gets disjointed and weird feeling. An open ended prompt and chat completion gives a much better feeling most of the time but can go off the rails.

If you don't let LLMs handle the conversation entirely you get into insanely complex traditional engineering problem: (if user is complaining do x, if user asks a relevant question do y, if user asks an irrelevant question do z), which is an endless can of worms.

That's been my experience with AI "support" chatbots as well, in every single case it's been an optimistic attempt to get the bot to fix the problem I'm calling with followed almost immediately by an even longer attempt to get it to connect me to a human that can actually fix the problem. If the problem was simple enough that a fixed-program clanker can fix it then I wouldn't be having to call support in the first place.