This is a very naive take. AI output still needs to be guided and iterated over, architecture decisions (even at the code level) still need experience to judge correctly.