Maybe brainstorming before a human actually does all the work. Or finding trends, if the data set is large enough that the ai-trend finding is worth all the re-checking and confirmation that you have to do manually.

I hear it's good at building software architecture with very strict rules in place so the time saved is worth the necessary manual check afterward.

Yea, aggregating data and brainstorming right now, but it doesn't seen worth the trouble yet, maybe in another decade.