https://research.google/pubs/pub43146/ should be required reading for anyone involved in designing or building machine learning components that are intended to go into production.
Now, that doesn’t mean that the fact that a method is classical makes it by default untangled or a good approach. I mean random forests consisting of big tree are really not that interpretable for example. And ML in general is not always the best tool for solving a problem. Gotta use our judgment factoring in the time dimension of software/product development.
Also, I think it’s a question largely of how deep learning is used. If it’s used in a very compartmentalized way, either at the product level or the code logic level, where you don’t need to be able to disentangle the component (and it’s giving huge wins over alternative methods), it can (can!) be a good choice.