Answering
The answer comes from your knowledge base, not from a general model
You fill the knowledge base with your own texts: return terms, delivery times, warranty, sizing. The AI agent answers from those. If the answer isn't there, it doesn't invent one, it says it isn't sure and offers to bring in a person.
You notice that difference mostly at the edges. A model allowed to say anything gives a plausible-sounding answer about your return window that happens not to be your return window. That is the kind of mistake you only see when a customer holds you to it.