ML does not include integer linear programming (nothing learnable there) and similar mathematical optimizations. ML can be a part of it but is not integral to it. Just like finding the shortest paths between two points does not include ML. ML can be used to learn the weights through time (traffic sensitive routing) but the search part is separated from it.

IMO, ML is nowhere close to useful in these problems when most of the clients just use pen and paper. It should be fairly easy to beat that with some simple search heuristics.