Pitch says it's about an asymmetry, but demo has the LLM calling a real person at a restaurant. How is that addressing the asymmetry?

But let's assume even small businesses like restaurants will adopt LLMs on the phone. So now you've got LLMs talking to LLMs that were trained to talk to people. What if we just trained the LLMs to talk to other LLMs? Surely they could complete the entire transaction in a tiny fraction of the time? They wouldn't be using any language we understand, of course.