Nice post, offers many thought-provoking questions!
I understand what you mean by “low hanging fruit use cases” (they are the obvious area of initial application). However, I am sometimes wondering whether they are really that “low hanging”? For instance, x.ai has raised $34m in funding and has been fine-tuning the algorithms behind their meeting assistant for over 2 years now. 2 years! They still haven’t moved from limited beta to a public release, they are still reliant on two dozen “AI trainers.” And we are talking about a super narrow domain here: one simple task (scheduling meetings) in one language (English) via one channel (emails). This is genuinely hard!