AI medical scribe notes: a worked SOAP example
Short answer: An AI medical scribe takes a patient-doctor conversation and returns a completed SOAP note: subjective, objective, assessment and plan, each field traceable to something said in the room. The work is extraction rather than transcription, because a note reorganises a conversation into the structure a chart requires rather than writing it down in order.
The fastest way to understand an AI medical scribe is to see what it produces. The technology sounds abstract until you look at a real note. So here is an illustrative, end-to-end example: a snippet of a patient-doctor conversation, and the structured SOAP note an AI scribe generates from it. (The example below is illustrative and contains no real patient data.)
Getting from the first block to the second is the substance of clinical documentation AI development, and it is more extraction than transcription.
The input: a snippet of conversation
Doctor: What brings you in today? Patient: I’ve had a sore throat for about three days, and since yesterday a bit of a fever. Doctor: Any cough or trouble swallowing? Patient: Some pain swallowing, no real cough. Doctor: Your temperature is 38.1 today and I can see some redness in the throat. Let’s do a rapid strep test and I’ll start you on rest and fluids.
The output: a structured SOAP note
Subjective: Patient reports a 3-day history of sore throat with pain on swallowing, and subjective fever since the previous day. Denies cough.
Objective: Temperature 38.1 C. Oropharyngeal erythema noted on exam.
Assessment: Acute pharyngitis, rule out streptococcal infection.
Plan: Rapid strep test ordered. Advised rest and fluids. Follow up pending test results.
That structured note, plus the specific fields your EMR requires, is generated automatically from the conversation, ready for the clinician to review and sign.
How the mapping works
Turning that dialogue into that note is a pipeline, not a single call:
- Speech-to-text with a clinical model (Deepgram Nova-2 Medical) that handles terminology, accents and background noise.
- Clinical entity extraction to pull the symptoms, findings, vitals and instructions out of the transcript.
- Field mapping that places each entity into the right SOAP section and into the 10+ mandatory forms your workflow uses.
- Confidence scoring, so any uncertain field is flagged for the clinician rather than silently filled.
Why this matters for medical charting
Charting is where clinicians lose hours every day. The value of an AI scribe is not “transcription,” it is structure: mapping a natural conversation to the exact fields a chart requires, with the clinician as the final reviewer. That is the difference between a raw transcript and a note you can actually put in a record.
The human stays in charge
Every field is reviewable, and low-confidence extractions are surfaced for confirmation. The scribe removes the typing, not the clinical judgment. That is the only way this belongs in real care.
What the note leaves out, and why
A generated note is shorter than the conversation, and what it drops is a design decision rather than a loss of fidelity.
Small talk, repetition and the clinician thinking aloud do not belong in a chart. Neither does a symptom the patient mentioned and then corrected. The system has to decide what was clinically asserted rather than merely said, and that decision is where quality separates.
The failure to watch for is the opposite: a detail that was said, was clinically relevant, and did not reach the note. Omissions are the most common error in a generated note and they are invisible on review, because a note that reads well gives no sign of what is missing. This is why evaluation has to compare against the transcript rather than against a reviewer’s impression.
Where specificity is lost
The note above is accurate and less precise than it could be, which is the normal case and worth understanding before you buy.
Clinicians speak in clinical shorthand rather than in coding language. The transcript may fully support a more specific diagnosis than the drafted note states, because the drafting model reproduced how the clinician talked rather than how a coder would write it.
That gap is the entire reason documentation improvement programmes exist, and it is also the opportunity: a system holding the transcript, the draft and the structured data can flag where greater specificity is supported, at the moment the clinician is still in the encounter. That turns a query sent two days later into a prompt answered in two seconds. The wider practice is covered in clinical documentation improvement.
What a reviewer should check
Four things, in this order, and it takes under a minute once the habit forms.
Attribution. Did any symptom get assigned to the wrong speaker. Most likely when a family member was present.
Medications and dosages. The highest-consequence field and the one where speech recognition is weakest.
Omissions. Anything discussed that does not appear. Harder than it sounds, which is why systems that show the transcript alongside the draft are worth more than those that do not.
Plan and follow-up. Frequently stated late in a consultation, at speed, and the part most often truncated.
The signature on the note means exactly what it always did. No vendor disclaimer changes that, and any rollout that does not say so plainly in training has created a risk rather than removed one.
Tools referenced
- Deepgram Nova-2
- LangGraph
- FastAPI
EpochC does AI medical scribe development — AI scribing, clinical documentation, medical transcription and charting for healthcare. See the scribe case study, our healthcare AI practice, or start a project. Audio work runs on our speech-to-text development.