1 & 2 are totally dependent on the company being willing to let their agents do things that they haven’t traditionally let humans do. For example, issue refunds, or do things that cost money but generate good will. I am skeptical that companies will be OK with their agents doing those things on their own volition.

3. Cool so the user didn’t indicate if they were satisfied. What then?

4. You can’t use a SOTA model right now for reasoning, there’s too much latency for a conversation. So you’re either using an older, but significantly less capable model, or you’re paying out the nose for fast mode. If the former then you can’t trust the agent to do the right thing (see points 1&2). If the latter, there’s no cost savings over a human. So which is it?