I have benn working so so much with explainability. I think the main thing to know when trying to extend this to non vision (which I assume to be standard feature engineering based ml, let's ignore texts for a second) is that all the explainability is in the features. You have to design features that make sense. It really doesn't matter if you have the best explainability method in the world if that perfect explainability method retrasces the models prediction to features nobody understands.