I'd argue that LLMs are not yet capable of the last step, and because most sufficiently large AI-generated codebase are an unmaintainable mess, it's also very hard for a human developer to take over and go the last mile.