The demos promise a prompt in and a finished asset out. In practice that is the one thing AI is worst at, because a finished asset is defined by constraints — topology, scale, UVs, how it sits next to the twelve other assets already in the scene — and a model generating in isolation knows none of them.
The places it does earn its keep are duller and far more useful.
Exploring before committing
The expensive part of concept work is not drawing, it is choosing. Generating twenty directions in the time one used to take does not make the choice for you, but it makes it an informed one. You reject nineteen faster and you reject them for better reasons.
The output is almost never the deliverable. It is the argument for what the deliverable should be.
The technical middle of the pipeline
This is where it quietly pays. Writing a script to rename three hundred objects by a convention. Reading an error you have never seen. Roughing out a tool that would have taken an afternoon and now takes twenty minutes, so it actually gets built instead of being worked around forever.
None of this shows up in a showreel. All of it is time back.
The part that does not delegate
Taste. Whether the shot reads. Whether the silhouette is doing its job at the distance the camera sits. Whether the material is convincing or merely detailed.
A model will happily produce something competent and characterless, and it will do it quickly enough that the speed is tempting. The judgement about whether the result is actually good is the job, and it stays with you. Anyone who hands that over ships work that looks like everyone else’s.
How I work now
AI is in the loop from the first idea to the last frame, and it never makes a decision unsupervised. I direct it the way I would direct a render: knowing what the result should be before I start, and rejecting everything that is not it.