There is a version of this problem that sounds simple. Train a model on medical conversations. Have it produce a structured note. Ship it to physicians and watch the time savings add up.
We tried that version. It produced notes that were technically complete and clinically imprecise. Physicians spent more time editing than they saved dictating. The tool had learned to sound like medicine without learning to practice it.
Where general models fall short
The difference between a general model and a specialty trained one shows up in the cases that matter most. A general model will document a cardiac condition. A cardiology model will distinguish between a condition that is managed by a device and one that has resolved — a distinction that changes the billing code the risk score and in some cases the treatment plan.
In psychiatry the gap is wider. A general model transcribes symptoms. A psychiatry model understands formulation — the difference between what a patient reports and what a clinician is actually assessing. Those are not the same document and they should not be generated by the same model.
Why we trained per specialty from the start
We made the decision early to build separate models for each specialty rather than fine tune a single general model. It is slower and more expensive. It also produces notes that require near zero editing after the first week of use.
Physicians are not looking for a tool that reduces documentation time by half. They are looking for a tool that removes documentation from their mental load entirely. That only happens when the output is accurate enough to trust without reading every word twice.
We are currently live across internal medicine cardiology and psychiatry. Each model was trained on specialty specific encounter data reviewed by practicing clinicians in that field. The cardiology model does not know psychiatry. That is intentional.
Generalist tools have their place. Clinical documentation is not one of them.
