> deploying and testing on real world is much easier and faster than letting loose an ML model on it.

No this is what you're not getting. It is "doing the things I would do to deploy and test in the real world, but faster and in the background while I do other things", it is not "letting loose an ML model on it". This is the new capability. If you have any process like "do {action}, wait until {something}, check {something}, determine if it matches expectation", it is now possible to run that loop way more times in way more variants without either spending the time on it synchronously oneself or writing a script to do it. (If you do that specific action loop often enough, it's probably worth writing the script anyway, but that's also much quicker to do now.)

The AI doesn't need training on the domain, it just needs to be told "these are the things I would do, please do them for me and report back".

I'm sympathetic to not everything being nail-like, but I really think you're leaving a lot of chips on the table if you can't imagine any of this kind of action-check-evaluate loop you have that you could offload.