Follow up to this question where effective prompting is discussed. Here are some ideas: Revising type-setting of old classics. Maybe even translating.
Extracting out of general ideas and approaches of proofs: You could give it a proof to, say Claude, and then ask it to give a high level overview of the proof. Now, this could be some scaffolding to build a deeper understanding by carefuler reading. It can also be used to compare and contrast proofs of similarish results.
Generating code: With AI we can shift the abstraction level to simply pure thought. By simply describing what we need, it can generate out the required programs for the specific usecase. This could be used for example to do numerical simulations.
Generating counter examples and new proofs: Well there are enough examples of this in recent times.
I am thinking now what more use cases can be there. I can imagine that in a few years that AI can reach a point where it could generate out lecture series for a given book if you feed in sufficient things, or even write out whole books on topics that are somehow semi accurate.