Not only do they fill this knowledge gap, but the gap is a massive chasm. Implementation details are not just figuring out how to get something to work in a bubble without regard for the rest of a system, but finding good compromises based on very imprecise and highly product-specific requirements. AI can just get something equivalent to a poem to barely compile when real world systems are a series of heavy tomes that contain nothing at all what the PMs think they do.