I could probably write a book to answer this question :D Blender has gone "everything nodes" route, and in particular it created the Geometry Nodes system. I'm very good in geonodes and pretty much specialize in it, and yet I think the system is severly flawed: the nodes in compositing and shading work very well, but in geonodes they are too low-level, and you don't get the easy learning curve of usual node systems (learning geonodes is very hard), while you get all the annoyances of using a very shoddy programming language and having to manage node positioning and fighting with noodles...

...And here's where AI comes into play, If AI could be contained into steps: - Input node: describe where the starting data comes from and AI automatically loads a file from hard drive or Internet or generates a primitive - Select node: describe a pattern by which to select elements of the geometry - Modify Geometry node: perhaps should be split into multiple nodes as there's so many ways to modify the geometry. - Sample/connect data: create an attribute and describe a relation of it to something else to create an underlying algorithm populating this attribute. - Save node: do you want to output the data through the usual pipeline, or maybe export to a file, or save to a simulation cache?

This way AI could do low-level stuff that I think it excels at, because this low-level stuff is so repeatable AI can be well trained on it. While the high-level decision making would be in control of an artist.