The control question is real. I've found the sweet spot is using AI for implementation but writing the architecture decisions yourself. When AI owns both, you end up with code that works but that you don't fully understand — which becomes a problem the moment something breaks in production.