You should have a look at a PCB netlist file... it's just "there's a chip called U1, it has package TQFP-48, and pin 1 connects to the net called "Net-U1-1", pin 2 connects to the net "+5V", etc etc. Obviously in a machine-readable text-format, but that's the gist of it. An AI can't glean anything from that (aside from making assumptions based on the name of the nets, which is just a horrible idea), even a human would have to either make the name-based assumption, or reverse-engineer the original schematic to understand what net is actually doing (complete with reading component datasheets). You think an AI can handle this when 99.9% of humans aren't able to?
I'm not sure if you're saying that you're a layman with respect to software engineering, or PCB layout as well. But if you're a software layman, this doesn't need to be a mystery to you. I'd encourage you to imagine teaching another layman how to do PCB layout, since a computer is the ultimate extremely fast, pedantic, stupid layman. Try to imagine a list of written-out instructions that you could hand to a temp worker to do PCB layout. Because a computer program is just that, a written-out pile of instructions.
Regarding machine learning and AI, you perhaps appear (like many) to think that AI is some sort of magical panacea that just solves any problem. But the problems you mentioned (OCR, speech recognition) are recognition problems: take a large file (image of words, or speech audio) and output text (which has precisely one correct answer). In the case of image recognition, the successive layers in the neural net recognise progressively more complex/higher level concepts -- lines/bends, then junctions and loops, then digits. Similar for audio. PCB autorouting as a huge space of different-looking but essential equally acceptable outputs, so the solution is not unique; the output is a huge amount of data because it's a fully routed board, not some condensed text output; and it's really hard for me to imagine what the different layers of the neural net would be doing. What does the first layer of the neural net do?
One last thought: 4 year old humans can do OCR and speech recognition. They can't lay out PCBs better than autorouters. Just imagine trying to teach a 4 year old to route a PCB, and you should have a good sense of what trying to program a computer is like.