The first widely adopted clinical use of language models was note-taking rather than diagnosis. That order was not an accident of ambition, and the properties that made documentation suitable are worth stating precisely.
The burden was already the main complaint
Clinicians spend a substantial part of the working day on documentation, much of it after hours.
That work is required for billing and legal purposes rather than for care, so reducing it does not reduce clinical activity.
A tool addressing an acknowledged and unloved burden meets far less resistance than one addressing a task clinicians consider theirs.
The clinician is a built-in reviewer
A generated note is read and signed by the person who conducted the encounter and knows what happened.
Errors are therefore caught by someone with complete information, before the document enters the record.
This is the safest possible arrangement for a system that can produce confident inaccuracies, and it exists in documentation without needing to be designed in.
Accuracy is checkable against a recording
The input is a conversation, and the output should reflect it, so correctness can be assessed by comparison.
Diagnostic suggestions have no such reference: whether a recommendation was right may not be known for months, if ever.
Measurable tasks get deployed and improved, because the improvement can be demonstrated to the people approving it.
The regulatory position is comparatively light
Software that produces a diagnosis or drives treatment is generally regulated as a medical device, requiring clearance.
Documentation support is usually treated as administrative, because a clinician reviews and takes responsibility for the output.
That distinction removes years from the path to deployment, and it is the single largest reason this application arrived first.
It also keeps the boundary sharp for the vendor, since a product that began suggesting what to prescribe would move into a different regulatory category overnight.
The risks are quieter than the failures being avoided
Generated notes can include plausible detail that was not discussed, and a clinician reviewing dozens of notes will not catch every instance.
Notes also become longer and more uniform, which makes the record harder to read for the next clinician and erodes the signal that a hand-written note carried about what the author considered important.
Neither of these is a reason to abandon the application, and both indicate that the review step doing the safety work is under more pressure than the arrangement assumes.