Oh, I know how to prompt AI. Getting it to spit out workable code doesn't mean you don't have to review the code, or make sure it's integrated correctly.
You also have to make sure it's not generating blatantly braindead code, which makes the review and debugging cycle take longer.
I remain unconvinced that it's suitable for domains where there is a right and wrong answer, like engineering or law.
I've found more value in the systems that do a good job understanding the problem description and then returning references to documentation and prior art on techniques, as opposed to the actual code.
I don't need a virtual junior dev I need to hand hold, I actually have those and mine get better. I want a virtual "person who worked on something like this once and knows the links to the good articles".