No single answer - a generalist's AI exposure depends on which part of the work you focus on. The bar shows the range, lowest to highest craft; see them below.
Roto
High
AI matting and rotoscoping have become genuinely good, pulling mattes and tracking edges that used to be frame-by-frame manual labour. What stays human: fine hair, motion blur and semi-transparent edges, which still need a person to check and clean.
FX / simulation
Mixed
FX splits sharply by the task, which is why this one is a band rather than a single verdict. Where the look is a well-learned natural pattern - ocean surface, smoke, fire, clouds - AI can produce convincing results fast, and a full 3D simulation would be slow and expensive. But where a shot needs precise, art-directable behaviour - exact timing, interacting with specific geometry, hitting a director's note - AI still drifts and makes mistakes, so studios fall back to controllable 3D sims. So how much AI can do here honestly depends on the shot.
Compositing
Low
AI takes isolated tasks - cleanup, roto, some matte work. What stays human: the precise node-based assembly of a final shot, where every operation must be exact, controllable and consistent across the whole sequence.
Matchmove / tracking
Low
Tracking is already heavily solver-automated, and AI is improving the hard cases. What stays human: verifying and correcting the solve so geometry locks perfectly to the plate - an error here breaks the shot.
Lighting
Low
AI relighting is promising but still research-stage for production. What stays human: the lighting decisions themselves - mood, motivation, continuity and matching across a sequence - which are deliberate creative choices.
Tools: Houdini, Maya, Photoshop, Unreal Engine, After Effects (context - we score the work, not the program)