The model by MIT's Daron Acemoglu estimates that ~5% of U.S. tasks can be completely and profitably automated by AI within ten years.
It was expressly not a head-count forecast, and didn't attempt to quantify the headcount reduction that AI augmentation could enable.