It is easy to list what AI can take off a clinician. It is more useful to be precise about what it cannot, because confusing the two is how products get built that doctors quietly stop using.
Examination
Nothing in a document replaces looking at someone. A great deal of clinical information arrives through observation that was never written down and never will be — how a person moves, how they look, what they are not saying.
The history that is not in the record
Patients do not volunteer the relevant thing first. Getting to it is a skill involving timing and trust, and it works because the person across the desk is a person.
Judgement under genuine uncertainty
Much of medicine is deciding what to do when the information is incomplete and will stay incomplete. That is not pattern matching with more data; it is a decision about acceptable risk for a particular person with a particular life.
Accountability
Someone has to be answerable. A signature is not a formality — it is the mechanism by which a decision has a person attached to it. This is why we draft and never issue.
What this implies for building
Aim at the work around the judgement, which is most of the day: the reading, the writing, the chasing, the explaining, the repeating. Remove enough of it and the clinician spends their time on the four things above — which is the only part that needed a clinician in the first place.
This article explains how our technology works. It is health information, not medical advice.