It seems the mockery these companies are getting is mostly focused on the dangers of giving a statistical model prone to "hallucinations" access to anything important. Which definitely seems like a bad idea. And won't stop being a bad idea unless something fundamentally changes about the underlying technology.

But it's worth taking a step back to see how bafflingly pointless a lot of the purported uses are in the first place.

Spending money online isn't hard. If I spend a lot of time making an online purchase, it's probably because I'm being extremely meticulous about what I choose to purchase and from where. In any case, it's not something that requires much help, and isn't a task I want to delegate to automation unless it's *exceptionally* reliable.

Or some things are unnecessarily hard because you're dealing with kind of broken systems (ironically enough, LLM-based phone support is probably the single most broken thing a customer will run into). It would be nice to not have to deal with that, but I don't have much faith in the "agent" to succeed under adverse conditions.