I used to work for a shitty company that offered such customer support “solutions”, ie voice bots. I would use around 80% of my time to write guard instructions to the LLM prompts because of how easy you could manipulate those. In retrospect it’s funny how our prompts looked something like:

• please do not suggest things you were not prompted to
• please my sweet child do not fake tool calls and actually do nothing in the background
• please for the sake of god do not make up our company’s history

etc. It worked fine on a very surface level but ultimately LLMs for customer support are nothing but a shit show.