Definition
Ambient clinical documentation is software that listens to a doctor–patient consultation through a microphone in the room, transcribes it, and has a large language model draft the clinical note in the structure the medical record expects — which the clinician then edits and signs. On the two questions clinicians and health-system buyers actually arrive with: it does save time, but far less per visit than the marketing implies, and the largest deployment ever published works out to roughly 22 seconds per encounter; and it is permitted almost everywhere because it is generally not a regulated medical device at all, which is precisely why it spread faster than any diagnostic AI in medicine.
Everything else follows from one boundary. The software drafts; the clinician attests. It does not diagnose, it does not recommend treatment, and it writes nothing into the record on its own authority. That boundary is what keeps it outside the device pathway, and it is also what leaves every legal consequence of the note exactly where it was before the microphone arrived. It is the fastest-moving corner of AI healthcare, and the only major one that is administrative rather than clinical.
How It Works
The pipeline has four stages, and the interesting difficulty is not where people expect it.
First, capture and speech recognition. This is a much harder problem than medical dictation, which is a solved one. A dictating clinician is a single cooperative speaker at a close microphone, using a controlled vocabulary, self-correcting in real time, and consciously producing text. A consultation is two to five people at an open microphone: crosstalk, interruptions, half-finished sentences, an interpreter, a family member answering for the patient, an accent the model has seen little of, a rustling gown and a corridor door. The system must also attribute speech, not merely transcribe it — "the chest pain comes on when I climb stairs" is a symptom when the patient says it and a clinician's confirming paraphrase when the doctor says it, and the two land in different parts of the note.
Second, the draft. A large language model reads the transcript and produces a structured note — SOAP, or a history of present illness with a review of systems — in the format the specialty and the record system expect. The essential thing to understand about this stage is that it is mostly a discarding operation. A fifteen-minute consultation is a couple of thousand spoken words; the note is a few hundred. The model's real job is deciding what does not belong in the record, and that decision is clinical judgement being performed by a system that has none.
Third, review and signature. The clinician reads the draft, corrects it, and signs. Fourth, it posts to the record. Only at the signature does the draft become a legal document.
The error modes are asymmetric, and this is the part that most first-time buyers misjudge. Invented content — a review of systems that never happened, a medication dose nobody said — is at least the kind of error a reader can catch, because the wrong thing is on the page. Omission cannot be caught the same way. Spotting that the patient's mention of new shortness of breath is missing requires reconstructing the encounter from memory and noticing an absence, which is a fundamentally harder cognitive task than proofreading. A blank note demands recall; a fluent draft invites assent. That asymmetry is the mechanism behind every clinical concern on this page.
Real-World Applications
Kaiser Permanente ran the largest published measurement of clinical AI use of any kind. Across 7,260 Permanente Medical Group physicians and 2,576,627 patient encounters at 17 Northern California medical centres, between October 2023 and December 2024, the estimated documentation time saved relative to non-users was more than 15,700 hours — 15,791, about 1,794 working days. 3,447 physicians used it in at least 100 encounters each. Adoption clustered in mental health, emergency medicine, primary care, allergy and cardiology, and was lowest in infectious disease, obstetrics and gynaecology, and urology. The findings were published in NEJM Catalyst as "Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses".
Now do the division, because it is the most important arithmetic on this page. 15,791 hours is 56,847,600 seconds. Spread over 2,576,627 encounters, that is about 22 seconds saved per encounter — call it seven minutes across a twenty-patient clinic day. Sit with how unimpressive that is, and then with the fact that the technology is winning anyway. The reported benefit did not show up as shorter visits. It showed up in after-hours charting and in physician-reported burnout: seven minutes recovered at 9pm from a task the clinician resents is not worth the same as seven minutes recovered at 11am, and the adoption curve is measuring the difference.
The randomised evidence says the product matters more than the category. A three-group pragmatic randomised clinical trial published in NEJM AI in 2025 assigned 238 outpatient physicians to Microsoft DAX Copilot, to Nabla, or to usual care, from 4 November 2024 to 3 January 2025. Nabla users showed a 9.5% decrease in time-in-note against control. DAX users showed no significant change against control (p = 0.66). Nabla came in 7.8% below DAX. Both scribe arms improved physician well-being, task load and work exhaustion.
That result is the single most decision-useful thing a buyer can know. Two vendors, the same trial, the same clinicians, the same weeks — and significantly different documentation-time outcomes. "Ambient AI scribe" is a category, not a product with an effect size. A study of one tool is not evidence about another, and a vendor citing someone else's trial, or its own internal numbers, is citing something that does not transfer. The only figure that describes what a given tool will do in a given health system is one measured in that health system.
Independent multi-system measurement points the same way on magnitude. A 2026 JAMA study of 8,581 clinicians across five health systems found documentation time fell by roughly 16 minutes per 8 scheduled patient hours — about 10% — with no significant change in after-hours EHR time. Set that against Kaiser's account of where the benefit landed and the honest summary is that the effect on the clock is small and contested, while the effect on how clinicians feel about the work replicates consistently. For a health system deciding whether burnout and retention justify the licence cost, that may be the right trade. It is a different case from the one usually made.
Key Concepts
The regulatory gap explains the adoption curve. Transcribing and summarising a consultation for a clinician to edit and sign does not diagnose or recommend treatment, so this software generally falls outside the medical-device pathway: no De Novo, no pivotal trial, no lengthy pre-market review. That is the structural reason ambient documentation reached millions of encounters while diagnostic AI was still accumulating authorisations. The flip side is unavoidable and should be stated as plainly: no device pathway also means no required pre-market evidence, which is why independent randomised results arrived years after mass deployment rather than before it. Kaiser's 2.5 million encounters preceded the trial that compared two vendors head to head.
Time-in-note, after-hours EHR time and burnout are three different outcomes, and they do not move together. A tool can reduce one, leave the second flat and improve the third, which is roughly what the published evidence shows. Deciding which of the three you are buying — before the pilot, not after — is what separates a procurement that can be evaluated from one that will be argued about.
Challenges
Automation bias is the central clinical risk, and it is a property of the reviewer, not the model. A fluent, well-formatted, plausible draft receives less scrutiny than a blank page would, because approving looks like a smaller act than writing. The clinician who signs remains legally responsible for every word, so the safety of the whole system rests on the one step most likely to be performed quickly at the end of a long day. Add the omission asymmetry described above and the risk is concrete: an inaccurate medical record is a patient-safety and liability problem, not a typo. Whatever a vendor's hallucination rate is, the review step is the only control, and its effectiveness degrades exactly when clinic runs late.
Consent and privacy are procurement questions with contractual answers. Recording a consultation places audio containing protected health information into a third party's pipeline, engaging HIPAA in the United States and GDPR in the EU, and requiring that patients be told and agree. The questions that decide whether a tool is usable are not technical:
- Where is the audio processed, and by whom — including any subprocessor the vendor uses for speech recognition.
- How long is audio and transcript retained after the note is drafted, and can retention be set to zero.
- May the recording be used to train models, and is that default opt-in or opt-out.
- Is the vendor contractually a business associate, with the paperwork actually signed — consumer tiers of general-purpose assistants are not.
Note bloat is the quiet one. Because generating text is now free, the drafted note gets longer, and length is not quality. The note is simultaneously a clinical handover, a billing document and a medico-legal artefact, so a fuller record can raise billing-compliance exposure while making the chart slower to read for the next clinician who opens it. Documentation time saved by the author can reappear as reading time spent by everyone downstream, and no vendor metric measures that.
Future Trends
The change worth watching is scope creep past the note. Vendors are moving from drafting the record toward drafting orders, billing codes and referrals — and the moment a generated draft shapes a clinical order rather than describing what already happened, the software starts to look like decision support. The regulatory exemption that made this category spread was earned by not touching treatment decisions; reopening that question on unfavourable terms is a real possibility rather than a distant one.
The second is whether buyers start demanding product-specific evidence. The NEJM AI trial established that two tools in the same study can differ significantly, which makes the category-level claim ("ambient scribes save time") unusable for procurement. The mature version of this market is one where a health system runs its own measured pilot with a control arm and compares time-in-note, after-hours time and clinician-reported burden separately — because the published record already shows those three can disagree. For the practical, tool-by-tool version of that evaluation, see the guide to AI medical scribes.