Why AI-accelerated teams keep breaking production — and what the ones that don’t are doing differently