I think you could a system up that could do quite well in classification and regression kaggle comps (top 25%, maybe better). I've built such machines a few times as I commented in the post.
The quandary is that its a purpose built machine. Aimed and fired at churn prediction (for example), it may tell you which customers will churn but not how it does it and hence you have no idea how to address the root cause. The predictions and even the model are not enough. Does that follow or am I off track?
I follow your point that predictions are not enough for complete causal relationship understanding. But what starting position would be better: having black box predictions or starting without this information? Perhaps these predictions can serve as hints and could speed up the understanding.