As someone in this space, I think that most people vastly underestimate the actual practical considerations in converting a neat CNN or ML algorithm into a usable product. It's trivial to use the newest flavor of ML to write an academic paper and hold up a new high score for a particular problem. It is very hard to make that a usable commercial product, because often 95% accuracy is not good enough, or it's better to have a deterministic failure mode than to have a semi-random one.