Lawrence Weed introduced the Problem-Oriented Medical Record in the late 1960s, and with it the SOAP structure: Subjective, Objective, Assessment, Plan. The format was never intended to describe how medicine is actually practiced moment to moment — clinical thinking doesn't move through those four steps sequentially in the exam room. SOAP was designed to organize the written record of what happened, not to map cognition in real time. That distinction matters when we consider how ambient AI scribes interact with it.
Physicians who have been writing SOAP notes for years have developed an implicit understanding of what goes where: the patient's report of symptoms and their context belong in the Subjective section; physical examination findings and vital signs belong in the Objective; the clinical interpretation and differential belong in the Assessment; the management decisions and follow-up plan belong in the Plan. Ambient AI systems don't reason through clinical content the way a physician does, but the good ones produce drafts that respect this architecture — and understanding where the AI maps content well versus poorly is essential for effective use.
The Subjective Section: Where Ambient AI Performs Best
The HPI — History of Present Illness — is typically the most time-consuming narrative section of an outpatient note. It requires weaving the patient's words, symptoms, timeline, associated factors, and relevant context into a coherent paragraph that orients any reader to why the patient is there and what the clinical picture looks like. For most physicians, this is the section that takes the longest to write from scratch after a busy clinic.
Ambient systems handle HPI generation well relative to the other note sections, because the source material — the patient's direct account of their symptoms — is present in the audio in organized form. A patient who describes chest pain that started three days ago, worse with exertion, relieved by rest, with no radiation, no shortness of breath, and a similar episode two years ago that resolved on its own, has essentially narrated the HPI. The ambient system's job is to recognize the clinical relevance of each element and assemble them into standard HPI structure using appropriate medical language.
The Review of Systems and the relevant social and family history updates also draw primarily from patient-reported content and are reasonably well-handled by current systems, with the caveat that brief patient disclosures — a mention of recent travel, a family history update, a note that the patient has stopped smoking — can be missed if the system treats them as conversational rather than clinical content. These are the segments worth paying particular attention to during review.
The Objective Section: Where Integration Matters
Vital signs, physical examination findings, and relevant recent laboratory or imaging results populate the Objective section. This is the section where ambient AI's limitations are most defined by the surrounding technical infrastructure.
Vital signs in most practices are documented by clinical staff into the EHR before the physician enters the room. If the ambient system integrates with the EHR and can pull those values into the note draft, the Objective section is populated with accurate structured data. If the integration doesn't exist, vital signs appear in the note only if the physician verbally states them during the encounter — which is not standard practice for most clinicians. The result is an Objective section that may be incomplete by default.
Physical examination findings are captured from what the physician narrates aloud. Some physicians have developed a habit of thinking aloud during examination ("lungs clear to auscultation bilaterally, no accessory muscle use") that translates directly into structured exam documentation. Physicians who examine silently and document after the fact find that ambient capture of exam findings is less complete. This is a workflow adjustment that takes conscious effort — narrating examination findings in real time while also examining the patient requires some practice to feel natural.
The Assessment: Clinical Judgment That AI Assists, Not Replaces
The Assessment section is where ambient AI faces its most significant challenge and where physician oversight is most critical. The Assessment requires synthesis: taking the patient's reported symptoms, the examination findings, the relevant history, and the clinical context, and arriving at a working diagnosis or differential. That reasoning process happens in the physician's mind during the encounter. The ambient system can capture what the physician says aloud about their clinical thinking, but it cannot generate that thinking.
Systems vary in how they approach Assessment generation. Some reconstruct the Assessment primarily from explicit clinician utterances — if the physician says "I think this is most consistent with viral upper respiratory infection, possibly with a secondary bacterial component," the system builds the Assessment from that narration. Others use generative approaches that attempt to synthesize an Assessment from the clinical entities identified in the transcript. The generative approach can produce well-written clinical prose, but it introduces the risk of hallucinated content — a diagnosis or clinical interpretation that wasn't actually the physician's conclusion, inserted because it was contextually plausible given the symptoms described.
This is the section of the ambient-generated draft that warrants the most careful review. The Assessment should reflect the physician's actual clinical reasoning, not the model's interpretation of what that reasoning probably was. For common, uncomplicated visit types, the difference is likely negligible. For complex, ambiguous presentations, it matters considerably.
The Plan: Highest Fidelity When Narrated Explicitly
Plan documentation benefits from the same principle as the Assessment: explicit narration yields better results than inference. A physician who states orders, referrals, prescriptions, follow-up timing, and patient education content aloud during the encounter — even briefly — gives the ambient system the material it needs to generate an accurate Plan section.
Consider a plausible outpatient scenario: an internist seeing a 58-year-old with poorly controlled hypertension and a new complaint of bilateral lower extremity edema. The Plan might include: uptitrate lisinopril, order BMP and urine microalbumin, add a loop diuretic, review sodium intake with the patient, and follow up in four weeks. If the physician narrates these decisions — either to the patient or as a verbal summary at the close of the encounter — the ambient system captures them with good fidelity. If the physician makes those decisions internally and places orders in the EHR without narrating them, the Plan section may be sparse or inaccurate.
This represents a genuine behavioral shift that ambient documentation asks of clinicians. The workflow change is not just adopting a tool — it involves developing a habit of thinking aloud, at least partially, so the documentation layer can follow along.
What SOAP Structure Gets Right That AI Should Preserve
The SOAP format has endured for decades not because it's the only way to organize clinical information, but because it maps well to how clinical care is reviewed and coordinated. A covering physician picking up an unfamiliar patient's chart can navigate a SOAP note quickly — the structure tells her where to look for the presenting problem, where the objective findings are, what the attending thought the diagnosis was, and what the current management plan is. Continuity of care depends on that navigability.
We're not suggesting that ambient AI is going to replace SOAP or reorganize clinical documentation around fundamentally different structures — at least not in outpatient medicine at current levels of adoption. What ambient tools change is the input method for each section, not the architecture of the note itself. The sections remain; the work of composing them is redistributed from post-encounter typing to ambient capture with physician review and correction.
What should change for clinicians adopting ambient documentation is the review workflow. Reading a pre-written SOAP draft requires active engagement with each section — confirming that the HPI reflects the actual encounter, verifying that the Assessment represents your clinical reasoning and not the model's interpretation of it, checking that the Plan matches the orders you actually placed. A sign-without-reading workflow is not appropriate for ambient-generated notes, and practices that adopt these tools without establishing clear review expectations create documentation risk. The structure of SOAP is a helpful scaffold for that review — each section has a defined scope, and working through the note section by section is a natural framework for systematic verification.
The note that emerges from a well-reviewed ambient-generated draft should look like a note the physician would have written — because the physician shaped it at every stage that mattered. The AI is a production layer, not a clinical author. That distinction is worth keeping clear, both for the quality of individual notes and for the longer-term question of how clinical documentation serves patient care.