But software engineering problems are more fuzzy and less amendable to mathematical analysis, so exactly how can those AI policies developed for math be applied to software engineering problems?

Not sure which way the difference puts the pressure. Does the fuzziness require more prudent policies, or allow us to get away with less?

Don't use them for the parts that are fuzzy.

I mean it should be obvious that making executive decisions about what the code should do exactly should only be left to a RNG powered model if the choices made are unimportant.