The personalized feedback could be really good for learning math. You could develop an "async" model for learning, where the student works on a proof (on their own) and tutor checks in once in a while to give tips and add corrections.
Note for basic arithmetic and algebra, the computer could check the answers automatically, but for proofs and more advanced/subtle arguments, the tutor will serve as the "math compiler" checking the student's work. Here's the discussion where the notion of a "math compiler" came up https://news.ycombinator.com/item?id=11267877