The honest answer is: the AI never picks the frame. Selection, tone, the decision of what the photograph is about — those stay with the photographer. Anything else is just process.
where it helps
Culling. The first pass through 600 frames is mechanical: out-of-focus, eyes-closed, exposure errors. A vision model handles that in twenty seconds and gets it right almost every time. The hour I get back goes into looking longer at the remaining 60.
Captioning. For internal archives only — a draft alt-text the photographer rewrites, never a final publication.
Search. "Find the frame with the red coat near a window." That used to be twenty minutes of scrolling.
where it does not belong
Composition advice. Tone curves. Anything described as "enhancement". A model that has been trained on the average internet photograph will always optimise toward the average internet photograph. That is the opposite of editorial work.
a note on local
Everything above runs on the workstation, not in someone else's cloud. The photographs are the client's. The model can be ignorant of that and still be useful.