I take note of when I provide input that disagrees with the LLM's suggestions, or adds additional insight. If I was just accepting the model's suggestions all the way then I don't need to be there, if I feel that I am contributing to a better result then I should be there.

This is starting to happen less, however.