Where an AI Scribe Belongs in the Note, and Where It Should Stay Out

The note has jobs. The scribe should only do some of them.

A clinical note is not one blob of prose. It is a set of jobs that happen to share a screen. One job is to record what was said and done. Another is to state what you think. A third is to commit the chart to an action: a prescription, a referral, a hold on a blood thinner. An AI scribe is built to do the first job from audio. Vendors keep sliding it into the second and third because those sections are what buyers notice in a demo. That slide is where clinics get hurt.

Treat the draft as a stack of trays. History and the spoken exam can arrive full. Assessment should arrive as headings you supplied, with blank lines under them if you did not dictate the reasoning. Orders should not arrive at all unless you said the order in a way your staff could have taken it down by hand. If your product cannot be configured to that split, you do not have a scribe. You have a ghostwriter.

Physicians love a note that looks done. That feeling is the hazard. A finished-looking assessment saves you four minutes and can smuggle a diagnosis you mentioned only to rule out, or a plan you discussed as an option and did not choose. Read the assessment as if a colleague covering your patient tonight has only those lines. If you would call them to correct it, it is not ready to sign.

History is where the draft earns its keep

The history of present illness is mostly chronology, and chronology is what a listener can capture. Who the patient is, why they came, how long this has been going on, what they already tried, and what they are worried about: if you and the patient said those things, a competent scribe should get them onto the page in the patient's order, not in a textbook order. Do not let it rearrange the story into a 'classic presentation.' Classic presentations are how real atypical stories get falsified.

Past history, medications, and allergies are a different problem. The patient may be wrong, the chart may be wrong, and the model may blend the two into a third list that matches neither. Prefer the medication list already in the EHR, reconciled by a person, over a list the scribe 'heard.' If the visit changed a dose, that change belongs in the plan as an explicit sentence you said, not as a silent edit to the list. Silent edits are how tomorrow's covering doctor gives the old dose or the new one without knowing there was a decision.

Social history and the patient's own words deserve quotation more often than paraphrase. A model asked to be concise will turn 'I am not taking it because it makes me dizzy at work' into 'noncompliant with antihypertensive.' You did not say that, and it is a worse clinical fact. Instruct the product, and your own edit habit, to keep motive. Motive is why the plan will succeed or fail. A cleaned-up history that flatters the clinician is a worse history.

The exam should be what you did, not what the template expects

If you say 'lungs clear, no peripheral edema, abdomen soft,' the scribe may write that. If you do not examine the abdomen and the template always includes an abdominal line, a careless configuration will still print a normal abdomen. That line is not a style issue. It is a false statement in the legal record. Turn off exam autofill. If the product cannot leave a system blank, do not use it for the exam.

Speak the abnormal findings out loud, including the ones that are merely absent. 'No focal neuro deficit' is worth saying if you looked. It is not worth having the software infer because the visit was a blood pressure check. The edit of the exam should be a hunt for systems you did not touch. Cross them out. A shorter exam is more credible than a complete one you did not perform, and open-notes patients can tell the difference when they were in the room.

Procedures are not an exam paragraph. A joint injection, a skin biopsy, a pelvic exam with a chaperone: the consent, the timeout, the specimen, and who was present should be dictated by you in the procedure note, not summarized from room noise. If the scribe captured chatter during the procedure, treat that text as scrap. Rebuild the procedure note from what you would have dictated on a phone. Ambient audio is a poor witness for a sterile field.

Assessment and plan stay in your voice

The assessment is the sentence another doctor will act on. It should sound like you, because it is you. A model that writes 'likely viral syndrome, low suspicion for bacterial pneumonia, supportive care' may be a fair paraphrase of a long conversation, or it may be a generic ending it has seen a thousand times. You cannot tell from the prose. You can tell by asking: did I say the suspicion out loud, including how low? If you only thought it, it is not in the audio, and it must not appear as if you said it.

Dictate the plan in short operational lines after the patient leaves if you need to. 'Chest X-ray today. No antibiotics. Return if fever past Friday. Work note for two days.' Those lines are easy for a scribe and hard for a model to improve without adding theater. Refuse theatrical plans. 'Patient educated extensively on red flags' is a phrase that means you are afraid. Write the red flags you named, or do not claim the education.

Differentials you discarded should appear only when you would want the next clinician to know you discarded them. A scribe that adds three rule-outs 'for completeness' creates problems for coders, for the patient reading the note, and for you if one of those rule-outs was never in your head. Completeness is not a clinical virtue when it is fictional. Your edit of the assessment is deletion more than addition. If you are adding the real plan by hand every time, move that work to a thirty-second dictation and stop letting the model draft it.

Orders, referrals, and the sentences that move money or drugs

An order is not a sentence in a note. It is an instruction to a pharmacy, a lab, or a colleague. The path from 'we might start a statin' to a prescription must include your confirmation on the order screen, not a checkbox someone else prepared from the draft. If staff are entering orders from the AI note before you sign, stop that this week. The note can remind them what you said. It cannot authorize.

Referrals fail when the reason is vague and succeed when the reason is specific. A scribe can capture 'refer to cardiology for exertional chest pain, abnormal ECG in the chart, not a current emergency' if you said it. It should not upgrade that to 'unstable angina' because the words are neighbors in its training data. Read every referral diagnosis before it is sent. The receiving physician will trust the label more than the paragraph, and the patient will see the label in the portal.

Prior authorization language is a special temptation. Models have read thousands of letters that say what payers want to hear. If your scribe starts producing those sentences inside the visit note, you are mixing a coverage argument with a clinical record. Keep the clinical facts exact. Let a human, using those facts, write the letter. A note written to win an authorization is a note that will read badly in any other setting, including a deposition.

Sensitive visits are not a configuration edge case

Psychiatry, reproductive health, substance use, HIV, domestic violence, and any visit where the patient lowers their voice are poor candidates for a default-on recorder. The clinical value of a verbatim history is real, and so is the harm of a transcript that lives in a chart more people can open than the patient understands. Have a written pause. The pause is a sentence you say and a control you hit, not a policy PDF.

Adolescents with a parent in the room, patients with a guardian, and patients who do not share a language with you need an extra rule. The scribe will capture whoever talks most. If the parent talks most, the note becomes the parent's history. If an interpreter talks, the note may attribute the interpreter's summary to the patient. Look at one such note from last month if you already use a scribe. You will see the problem immediately, and you should change the workflow before you change the vendor.

Turn the tool off for part of a visit without apology. 'I am going to pause the note so this part stays between us, and I will write a short line myself' is a competent sentence. Patients who have been surprised by a recording do not become more trusting when you explain the business associate agreement. They become more trusting when they see you can stop it. A product that cannot be paused quickly does not belong in an exam room.

Editing is the product. Time it.

Do not ask whether people 'like' the scribe. Ask how many minutes sit between the end of the visit and the signature, and how many of those minutes are repair. Repair means deleting invented detail, restoring a negation, and fixing a medication. If repair exceeds a couple of minutes on a routine visit, the scribe is writing too much. Shrink its territory before you buy another training session.

Edit in a fixed order so you do not get lost in prose. Medications and allergies first. Then anything with a side or a number. Then the assessment, looking only for claims you did not make. Then the exam, looking only for systems you did not touch. Then stop. A full literary rewrite means you are treating the draft as a student note. It is not a student. It does not learn from your elegant sentence unless the vendor has a feedback path you have actually tested.

Share the edit rules with everyone who signs, including locums. A locum who trusts the draft because it looks like the group's usual note is the person most likely to sign an invention. Give them the same one-page rule: history may stand if it matches the visit, assessment must match what you said, orders are confirmed on the order screen, sensitive portions are paused. That page is the implementation. The rest is software.

A line you can defend later

Someone may read this note in a year: a partner, a patient, a reviewer, a lawyer. They will not care which model version wrote the first draft. They will care whether the signed document is what you did and what you decided. Build the scribe's permissions backward from that reader. Give it the history. Give it the spoken exam. Keep the decision, the order, and the sensitive aside in your hands.

If you want a single sentence for the medical staff: the AI scribe may draft what was said and done; the signing clinician owns what was decided; nothing is ordered, coded as a diagnosis, or left in a sensitive history unless a person checked it against the visit. Post it. Configure to it. Sample notes against it. That is a scribe a physician can live with, because it does not pretend to practice.

Sunrise's work has always ended at a document a clinician can sign without reconstructing the visit from memory at 9 p.m. An ambient tool can help that happen if it stays in its lane. The moment it starts improving your medicine, turn that part off. Improvement you did not ask for is not a feature. It is an unsigned coauthor, and coauthors do not belong in a chart that has only one signature line.

Review that sentence with the people who configure the tool, not only with the clinicians who use it. Configuration is where the lane is actually set: which sections autofill, whether a normal exam is assumed, whether a problem list is rewritten, whether orders are suggested in the same window as the note. If the config screen can do something your one-sentence rule forbids, turn that control off and remove permission to turn it back on without the medical director. A rule that the software can silently violate is a wish.

Two visits that show the lane clearly

A routine hypertension follow-up. The patient talks about home readings, dizziness on the current dose, and a new knee pain they do not want imaged today. The scribe should capture that story, the readings they recited, and the exam you spoke. You should dictate or type the decision: lower the dose, why, when to check labs, and that the knee is not being worked up today. If the draft already contains a knee MRI and a lifestyle essay you did not deliver, you are looking at a tool that has left the history and started practicing. Delete both. Do not 'tidy' them into softer language. Absence is the correct record.

A visit that turns. You brought the patient in for diabetes and they tell you about chest pressure. The history of the pressure belongs in the scribe's draft if it was said. The decision — ECG now, aspirin or not, emergency department or not — belongs in sentences you say after you have decided, not in a paragraph the model writes because chest pressure 'usually' leads somewhere. Pause the scribe if the room gets crowded and the story is coming from three people. Resume only when you are ready to state what happened and what you are doing. The note of a turn in the visit should read like a turn, not like a smooth narrative with a moral.

After either visit, the test is the same. Hand the signed note to a colleague who was not there and ask what they would do tonight if they had only the assessment and the orders. If they would do what you intend, the scribe stayed in its lane. If they would do something you mentioned and rejected, the draft won. Track those misses for a month. They are a better quality metric than minutes saved, because minutes saved can be produced by signing faster and reading less.

Write the lane down where the people who train new clinicians will see it. One paragraph is enough: the scribe may draft the history and the exam you performed; the assessment, the plan, the orders, and any sentence that changes a drug or a disposition are written or dictated by the clinician after the decision; sensitive portions of a visit are paused; a rejected recommendation is deleted, not softened. Put that paragraph in the onboarding note and in the vendor configuration ticket. New hires will follow the screen in front of them. If the screen offers a one-click normal exam and a suggested order set inside the note, your paragraph lost. Take those controls off the screen, then teach the paragraph.

Revisit the lane when the model updates. Vendors ship 'smarter' drafts that reach further into the assessment because that is what the demo rewards. Treat an update like a new device: turn it on for a few visits with a reader who is not in a hurry, compare the draft to what was said, and only then allow it for the group. If the update starts proposing plans, roll it back or turn that section off. The lane is a clinical policy. A release note does not amend it.