What an AI Medical Scribe Actually Does During a Visit

Listening is not the same as charting

An AI medical scribe does not sit with a clipboard. During the visit it captures the conversation from a phone, tablet, or exam-room microphone and tries to turn that talk into a structured draft. That is a different job from speech-to-text. Speech recognition gives you a transcript of words. A scribe has to decide what belongs in history, what belongs in the exam, and what is small talk that should never land in the chart. Treat the tool like a tape recorder and you will spend the evening cleaning a transcript instead of reviewing a note.

The useful mental model is a first-pass writer sitting just outside the conversation. It hears the chief concern, the timeline, the medications named out loud, and the plan you state before you stand up. It does not perform the physical exam. It does not know a finding you never verbalized. If you murmur that the lungs are clear while you listen, that phrase can make it into the draft. If you only type the finding later, the model has silence to work with, and silence is where invented exam language usually starts.

What the model is doing while you talk

While you take a history, the system is doing three jobs at once. It is converting audio to text. It is trying to assign speakers so the patient's words do not become your assessment. And it is mapping fragments onto the sections your template expects. A good AI medical scribe is not summarizing the visit into a paragraph. It is filling a form that looks like your note: chief complaint, history, review of systems, exam, assessment, and plan. If your template is empty or generic, the draft will look generic even when the visit was specific.

That mapping is why ambient tools fail in messy rooms. Two people talking, a parent translating, a nurse giving vaccines, a device alarm in the hall: the model has to ignore noise without dropping the one sentence that changes the plan. It will not always succeed. The draft you see after the visit is a hypothesis about what happened, not a legal record. Your job is to confirm, strike, or add. Sunrise aims for drafts in about two minutes after a typical clinic visit so that review can happen before the next patient is roomed, not at 9 p.m.

What the clinician still has to do

You still own every diagnosis, every medication change, and every follow-up instruction. The scribe can place a draft plan in the right heading. It cannot know that you decided against imaging after you left the room, or that the blood pressure you trusted was the second reading. Review is not a courtesy pass. It is the moment the document becomes yours. If you sign without reading the exam, you are attesting to findings you may never have made. That is true whether the draft came from software or from a person.

The fastest reviewers develop a short ritual. Scan the chief complaint against what you remember. Check that negatives you did not ask are not sitting in the review of systems. Confirm that the assessment list matches the problems you actually addressed, not every chronic condition mentioned in passing. Then read the plan out loud in your head: medications, orders, precautions, return precautions. If a sentence would surprise the patient who was in the chair, it does not belong. That ritual is shorter than rewriting the visit from a blank screen.

Where the draft goes after the visit

Capture in the room is only half the workflow. The draft has to land where you already work: the encounter in your EHR, a holding queue, or a document that staff can paste with identifiers intact. Practices stall here more often than on the microphone. An integration that dumps an unformatted blob into a miscellaneous note is not a finished product. Ask how identifiers are matched, how addenda are handled, and who can see the audio. Details live on Sunrise EHR integrations and should be written into the vendor's process, not left as a hallway promise.

Audio and text are PHI the moment a patient is identifiable. A signed BAA should be in place before the first real visit is recorded. Ask where processing runs, who can listen to files, and whether audio is retained after the note is signed. Sunrise uses U.S.-based processing and will execute a BAA before any PHI moves. Those answers belong in writing, next to your HIPAA packet, so a new office manager is not rediscovering the workflow from a login screen six months later.

How this differs from a person in the room

A human scribe can catch a nod, a wince, or the fact that you pointed to the left knee. An AI scribe only has sound and, in some products, a limited view of the screen. That gap is not a reason to reject the tool. It is a reason to narrate findings you care about. 'Left knee, no effusion, full extension' takes three seconds and saves a later edit. Clinicians who already think out loud adapt faster than clinicians who chart in silence and expect the model to infer the exam from atmosphere.

The other difference is consistency. A person has a good day and a tired day. Software makes the same class of mistake until you correct the template or change how you speak. That can feel mechanical. It is also how you train the system without a six-week project: speak the headings you want, correct the same error twice, and stop accepting a draft that invents a complete review of systems. For clinics that still want a person on the line, a virtual medical scribe remains a parallel option, not a failed version of ambient capture.