Card 04 of 13 · during the visit
delegable draft, review required

Ambient visit transcription

The reason for this card is not efficiency. It is presence. Everything else here exists so that presence does not cost you accuracy, and so that a note you did not write does not end up carrying your signature.

The case

A patient is telling you something hard about her life. In the old workflow you would be typing through it, eyes on the screen. The scribe is listening instead, and she has your eyes.

Guiding question: Does the note say anything the visit did not?

Capturing the conversationDelegable
Structuring the noteDelegable with review
Deciding what belongsYours
Signing itYours · a legal document
01 · What it does, and what it does not
What it does well

Listens while you look at the patient, and captures more of the conversation than your typing would. The time savings are modest and real, and part of it goes back into review, which is the point: what you gain is attention, not speed.

What it does not do

Know what happened outside the audio. Nothing you did in silence exists as far as the note is concerned, and nothing the model adds gets fact-checked before your signature goes on it.

Clinical design position

The move from pen to computer created a patient who watched a doctor "who never looked at me," without knowing he was documenting. Ambient transcription gives the eye contact back, and that is worth protecting properly. One precision before anything else: you are not delegating documentation. You are delegating typing.

02 · Why this card is different
The environment decides whether this card exists for you

On every other card you write the prompt and prepare the input. Here the tool listens passively, and the input is the whole visit, including what was never meant to become a record. There is no anonymizing: the audio is identified by nature, with a name, a history, and a voice, which is biometric data.

In a tool covered by an agreement for patient data (Card 01), this is a delegable draft with review. In a consumer app with the microphone open in an exam room: do not. There is no "paste less" here. The identifiable audio is the entire product.

Consent before the first second

Explicit, documented, with refusal respected and the visit proceeding normally without the tool.

And a detail almost nobody states: recording captures everyone in the room, so consent covers the person who came with the patient too.

Before you use it, know where the audio goes, how long it stays, and whether it trains anything.

03 · Where it fails here
Documents what nobody said

Two mechanisms. In transcription, speech recognition can invent whole phrases during pauses. In note generation, the template completes: a "denies fever" nobody asked about, a "lungs clear" nobody examined. The second is the legally serious one, because it documents an act that did not happen, under your signature.

Loses what was said out of order

Real visits are non-linear; notes are linear. The "oh, and sometimes my chest gets tight going up stairs," dropped in the middle of another topic, is exactly what compression discards as noise.

Swaps who said what

The daughter's report becomes the patient's. Your hypothesis, spoken out loud, becomes an established finding. Plus the classic mishearings: sound-alike drug names, numbers, laterality.

These three link to the site's failure taxonomy, where the transcripts live.

04 · Three outputs, not one
01The transcript.Kept as the source for review, with retention following your institution's policy rather than becoming a new permanent archive by default.
02The structured note.The draft you will review and sign.
03The index of what was discussed.A plain list of topics, a receipt for the conversation, and the anti-omission device of this card. Scanning a topic list against your memory of the visit takes thirty seconds; if something you remember discussing is not there, you go to the transcript. The tool still selects, but the selection becomes visible and contestable instead of silent. Alongside it, preserve clinically meaningful patient phrasing in quotation marks, verbatim: next visit's "but you told me..." needs the exact sentence, not a paraphrase.
Console · technique

The visit, spoken for two listeners

You run the visit normally for the patient and, with small adjustments, make it legible to the machine. Nothing here is for the microphone's benefit alone: each move also improves the visit.

Speak the exam as you perform it

What you do in silence becomes either an omission or a fabricated template line.

"Blood pressure lying down, 138 over 84.
Standing, 112 over 70."

Numbers spoken out loud land in the recording correctly, and the patient hears what you found.

State negatives as conclusions

Review of systems happens in a rapid volley, and a nod does not transcribe. Close out loud instead:

"So for the record: no fever, no abdominal pain,
no urinary symptoms, sleep is fine."

That transcribes perfectly, structures the note, and confirms with the patient in one move.

The third voice

Rooms often hold three, and the tool tends to flatten them into one narrative. The clinical protocol you already practice is the one the tool has to follow: information from a companion goes in attributed, and validated with the patient, who is the primary source as long as she can speak for herself.

When accounts differ, both versions go in with attribution, never merged. A documented disagreement is clinical data, and it is your protection. Watch treatment preferences especially: the companion's worry must not become the patient's choice on paper.

Instructions for the tool
Produce three outputs: the transcript; a structured note;
and an index of every topic covered.
Preserve, in quotation marks, the patient's own clinically
meaningful phrasing.
Attribute every piece of information to its source. Anything
from a companion is recorded as "per the companion." When
accounts differ, record both; do not merge them.
Record only negatives that were spoken explicitly.
Record only exam findings the clinician stated out loud.
Hypotheses spoken aloud are recorded as hypotheses.
Console · exercise

What showed up that nobody said

This exercise is the mirror image of Card 02. There you mark what disappeared from the source; here you mark what appeared that the conversation never contained. Omission and fabrication are opposite motions of the same muscle. Synthetic case.

From the visit · transcript
PatientI get dizzy when I stand up fast, for about two weeks now.
DaughterHe's forgetful too, doctor. I think he should get an MRI.
ClinicianMr. Silva, do you feel forgetful?
PatientNormal for my age, doctor. I forget names, nothing that gets in my way.
PatientOh, and the new blood pressure pill, I started it about three weeks ago.
ClinicianI'm going to check your pressure lying down and standing. Lying, 138 over 84. Standing, 112 over 70.
ClinicianSo for the record: dizziness on standing for two weeks, started after the new blood pressure medication. No fainting, no chest pain, no palpitations.
The generated note

Patient reports postural dizziness for 15 days, beginning after a recent change in antihypertensive. Reports progressive memory loss. Denies syncope, chest pain, and palpitations. Denies headache and visual changes. BP 138/84 supine; 112/70 standing. Cardiac exam: regular rhythm, no murmurs. Assessment: medication-induced orthostatic hypotension.

Clean, plausible, and ready to sign. Three of these sentences describe things that did not happen.

Swapped attribution

"Reports progressive memory loss" did not come from the patient. It came from his daughter, and the patient minimized it when asked directly. A consequential swap: it changes who owns the complaint, and it is the sentence a future clinician will act on.

Fabricated negative

"Denies headache and visual changes" was never asked. The template completed the review of systems, and a fabricated negative is a question the note swears you asked.

Fabricated exam

"Cardiac exam: regular rhythm, no murmurs" did not happen and was not spoken. This is the legally serious item on the page: an exam documented without an exam performed.

What came through cleanly, and why

The orthostatic numbers, because they were spoken. The negatives for syncope, chest pain, and palpitations, because they were closed out loud. And the timing link to the new antihypertensive, a mid-visit digression that survived only because the clinician restated it. The techniques above, working.

05 · Before you sign

0 of 7

Signed means yours. Nothing signed unread, and AI use recorded per local rules.

Where this leads

The documented visit is raw material for the next card. With your exam done and your hypothesis formed, the question becomes: what did I fail to ask, and what would rule out the alternatives?

Two purposes, both real: patient safety, and your own documentation. Card 05 handles it.

06 · Evidence

Adoption is real but uneven, the time savings are modest, and fabrication in speech recognition is documented rather than theoretical. Those three facts together are the argument for the review step above, not against using the tool.

  1. [article] Rotenstein L, et al. Changes in clinician time expenditure and visit quantity with adoption of AI-powered scribes: a multisite study. JAMA, 2026. 8,581 clinicians across five academic systems, 21% adoption, 16 minutes of documentation time saved per 8 hours of care. doi.org/10.1001/jama.2026.2253
  2. [article] Koenecke A, et al. Careless Whisper: speech-to-text hallucination harms. ACM FAccT, 2024. About 1% of transcriptions contained entirely invented phrases, 38% of those carrying explicit harm, with pauses raising the risk. doi.org/10.1145/3630106.3658996 The authors note the system improved after the study: cite it as a demonstration of the mechanism, not as a current rate.
  3. [report] Menlo Ventures, The State of AI in Healthcare, 2025. Ambient scribe adoption estimated around 35% in large US health systems. link to complete

Further reading · Associated Press investigation, October 2024, on invented content in medical transcription. The narrative version of the mechanism above.

Educational content; synthetic visit and note. Not a substitute for clinical judgment or for the rules on recording, consent, and documentation that apply where you practice.