Turning a spoken patient encounter into a structured clinical note happens in a few distinct stages, even though it feels like a single step to the clinician.

The first stage is transcription: converting audio into text. This step benefits from speaker diarization — distinguishing who said what — since a visit typically involves at least a clinician and a patient, sometimes more. Getting speaker roles right (which speaker is the clinician, which is the patient) matters for producing a note that attributes symptoms and statements correctly.

The second stage is structuring. A raw transcript isn't a clinical note — it's a conversation. A language model trained on clinical documentation patterns extracts the relevant clinical content (chief complaint, history, exam findings, assessment, plan) and organizes it into the note format the clinician has selected, whether that's a built-in template or a custom one.

The third stage is the part that matters most: review. The generated note is a draft. The clinician reads it, corrects anything that's wrong or missing, and only then does it become part of the patient's record. This is a deliberate design choice — the model assists with structure and drafting, but clinical accuracy and sign-off remain the physician's responsibility.

Underneath all of this, token usage and processing cost scale with the length of the conversation and the complexity of the note being generated — which is why documentation platforms track usage per note rather than treating every visit as identical.

From Consultation to Note: Behind the AI