You do not find an uplift number. You bound it, then narrow it with evidence.
Everyone else starts from a vendor claim and discounts it by feel. That produces a figure with no defence the first time somebody serious attacks it, and somebody serious always does.
Start from what the work is made of
Decompose the task. Measure what share of it each component consumes. Identify which components the system actually touches. The ceiling is the share it touches, and it holds regardless of how good the model is. No system can save time on work it does not do.
Take clinical documentation. Fourteen minutes per encounter, of which roughly six is transcribing what was said and three and a half is structuring it into a note. Two and a half is clinical reasoning added afterwards. The rest is coding and review.
A system that transcribes and structures cannot touch the reasoning, and review takes longer rather than shorter once there is generated text to check. The arithmetic maximum lands near two thirds.
Which makes the claim checkable
A vendor claiming an eighty per cent reduction is not claiming a better model. They are claiming to have removed clinical judgement from the encounter.
That sentence is defensible, it is checkable by anyone with the decomposition in front of them, and it is worth more than any survey number we could have collected instead.