S, O, A and P: what belongs in each section, and which ones a recording can fill

A SOAP note has four sections that differ less in subject matter than in where their content comes from. Subjective is the patient's account of the problem. Objective is what was measured or observed. Assessment is the clinician's interpretation of the first two. Plan is what was decided. For anyone building or buying documentation software, that ordering is the whole story: each section sits further from anything a microphone can hear than the one before it.

The structure is Larry Weed's, from the problem-oriented medical record, and roughly half a century old — the StatPearls reference entry on SOAP notes credits him with it. It has outlived every documentation fashion since because it keeps evidence and inference in separate boxes. A note that blurs the two is harder to read six months later and much harder to defend.

Four sections, four different sources of truth

Subjective is the section a recording serves best, because a consultation is mostly a patient describing a problem. The history of present illness is the bulk of it, and which of its elements a clinician actually says out loud rather than carrying silently is set out in the HPI's elements and what a consultation contains. Get that section wrong and everything downstream inherits the error, because the other three are all built on it.

Objective is the section most often got wrong, because half of it is not audible at all. A blood pressure of 148 over 92 exists in a device reading or a chart entry, never in the conversation. Examination findings exist in the conversation only when the clinician narrates them — chest clear, no ankle swelling, abdomen soft — which some clinicians do by habit and others never do at all. Two clinics running identical software will get very different Objective sections for that reason alone.

The four misfilings that show up in every chart

• Interpretation written into Subjective. "Patient is anxious about the lump" is a conclusion. "Patient says she has not slept since finding the lump" is the report. The first belongs in Assessment, and filing it under Subjective quietly converts a clinician's observation into something the patient supposedly said.
• Patient-reported numbers written into Objective. Home blood-pressure readings, a weight from the patient's own scales, a glucose figure read out from memory. All are legitimate content and all belong in Subjective with attribution, because Objective is the section that asserts you measured it.
• A recap written into Assessment. Two tidy paragraphs summarising what happened in the room, with no diagnosis and no differential. That is a summary of the visit, not an assessment of the patient, and it is the single commonest failure in a generated note.
• Discussion written into Plan. "We discussed treatment options" is not a plan. A plan is a commitment with specifics attached — the drug and dose, the test and why, the review interval, what should prompt an earlier return. Why that section keeps coming back thin has its own diagnosis.

Two sections a recording can source, two it cannot

Subjective, and the narrated part of Objective, are transcription and structuring problems. They are hard in the ordinary engineering sense — accents, overlapping speech, a relative answering for the patient — but the information genuinely exists in the audio, so a system that gets them wrong can be measured and improved against the recording itself.

Assessment and Plan are not that kind of problem. Clinicians reason in silence: the differential is formed while the patient is still talking and is rarely narrated. A model asked to produce an Assessment from a transcript containing no reasoning will still produce one. It will be fluent, it will be clinically plausible, and it will be the model's differential rather than the clinician's. That is the failure that matters, because it is invisible in review — a plausible paragraph reads like the clinician's own thinking, so it gets signed.

The mapping step that decides which span of a transcript lands in which section, and what happens to the spans that fit nowhere, is a design problem in its own right and is worked through in from transcript to structured note. The argument here is narrower: the four sections are not four equal buckets, and treating them as one generation task is what produces confident Assessments.

What the four letters do not tell you

• Which sentences are from today. The structure has no field for provenance, so text carried forward from a previous encounter is indistinguishable from text written now — the mechanism behind copy-forward and a chart that stops being readable.
• Who is asserting it. SOAP predates machine drafting by decades and has nowhere to record that a draft was generated, reviewed and adopted, which is why the attestation on a machine-drafted note has to exist as its own record rather than a line of text at the foot.
• How accurate is accurate enough. Section-level correctness is not one number: a wrong dose in the Plan and an ungainly sentence in Subjective are not comparable errors, which is the whole argument of the metric a documentation pilot is judged on.

Two of the four sections are records of what happened. The other two are records of what a clinician thought — and only one system in the room can produce that second kind.

A scribe built around this division writes into a shape it can defend: transcribe into Subjective, source Objective value by value, and put Assessment and Plan in front of the clinician as proposals to confirm or replace rather than text to skim. That review-before-commit posture is the same one any agent writing into a system of record needs, for the reasons AI agents in production sets out, and it shapes how we approach AI agents and automation in clinical settings. The rest of ambient documentation and the clinical note — diarisation, templates, signature queues — assumes this division holds, and the wider picture sits across our healthcare and MedTech work.

• clinical documentation
• SOAP note
• ambient scribe
• note structure