One example where it may be truly revolutionary is astrophysics research. I saw a lecture by an astronomer talking about it. Astrophysics works well with the limitations of AI. A lot of their work is building simplified models to study large complex astrophysical systems. Imagine someone modeling the formation of planets in a circumstellar disk.
The key is that these are very contained problems. You're just trying to simulate a little toy solar system. The code can all be completely self-contained and doesn't need to deeply integrate with tons of other existing software and infrastructure.