Mathematicians both build/formulate abstract conceptual frameworks and formulate problems within those frameworks. Solving the problems has historically been one of advanced the conceptual frameworks.

Now, what has been useful for advancing physics and biology has been the abstract conceptual frameworks - using Hilbert Space for model quantum mechanics or tropical geometry for DNA modeling.

So far AIs have shown an ability to prove mathematical theories but not an ability formulate coherent-to-human mathematical models. So allowing AIs to simply replace mathematicians isn't so far a way to achieve broad advances in other fields. Indeed, AIs so far are great mathematicians only if one views mathematics as solving "meaningless puzzles ... in isolation"