AI Medical Scribes and Ambient Clinical Documentation

An ambient AI scribe listens to a clinical visit, transcribes the conversation and drafts the note. The clinician then edits the draft and signs it. The randomized trials so far find modest savings in note time that depend on the product, larger self-reported gains in burnout, and errors in a minority of notes that someone must catch before signing. Scribes and dictation both turn speech into clinical notes, but they work differently, and the best choice depends on the kind of visit. For the wider subject, see our guide to medical dictation.

An ambient AI scribe records the visit, transcribes it and drafts the note for the clinician to review and sign

An ambient AI scribe runs during the consultation and turns the conversation into a draft note. "Ambient" means the clinician does not have to narrate the note. A medical scribe was historically a person who wrote the notes during a visit. The AI version does that job from a recording, using automatic speech recognition and a language model that organizes the text into a note.

The draft goes into the electronic health record, the system where clinical notes are kept. In February 2026, Healthcare Dive reported that Epic had rolled out an AI tool that drafts clinical notes ¹.

The output is therefore a draft written by a model from a conversation, not a transcript typed by a person. Once the clinician signs, the note is the clinician's record, which is why the review step in the flowchart is not optional. Human transcriptionists have long turned dictated speech into formatted notes. The medical transcription guide covers that job.

Documentation burden is the problem scribes exist to solve: EHR and desk work takes more of the day than face time with patients

Physicians spend a large share of the office day on records and desk work. In a 2016 time-and-motion study of U.S. physicians, physicians spent 27.0 percent of their office time in direct face time with patients and 49.2 percent on EHR and desk work ². That works out at about 1.8 hours of EHR and desk work for each hour with patients ³. The practices were self-selected high performers in four specialties ², so the figure is not a national estimate. This is the basic answer to how much time doctors spend on documentation.

Nurses carry a similar load. In a Western Australian hospital, observers found nurses and midwives spent 28.1 percent of morning shifts on clinical documentation ⁴. Staff in the same study estimated 50.4 percent of morning shifts when asked ⁵. In that hospital, self-reports ran well above observed time, which is a reason to check survey measures against direct observation.

Note-writing method is also linked to different note time. In Epic's analysis of ambulatory physicians ⁶, physicians who mainly used transcription and dictation spent 4.8 minutes per visit on notes, less time than physicians who mainly used SmartTools ⁷. These are adjusted associations in observational data, not experiments.

Randomized trials show modest time savings that differ by product

The most informative test so far is a randomized trial at UCLA that compared two scribes, Microsoft DAX Copilot and Nabla. It assigned 238 outpatient physicians across 14 specialties to one of the two scribes or to usual care ⁸. Only the Nabla arm showed a statistically significant reduction in time-in-note: users cut it by 9.5 percent against the control group ⁹. DAX users showed no significant change, at -1.7 percent ¹⁰. A result that is not significant does not show that DAX has no effect. Its confidence interval runs from a 9.4 percent reduction to a 5.9 percent increase ¹⁰.

The three bars do not measure the same thing. The UCLA figures are changes in logged time-in-note against a control group. The Singapore figure is documentation time per consultation in a before-and-after study ¹¹. In that study, documentation time fell from a mean of 5.3 minutes to 4.5 minutes, and consultation length did not change ¹¹.

Most of these trials ran for weeks, not years. The Wisconsin design lasted 24 weeks ¹², and the UCLA trial ran at one academic center over two months ⁸. Short trials show the effect in the first months of use. They do not show whether the saving lasts once the novelty wears off, which is the question a clinic should ask before it signs a contract.

Other randomized designs point the same way, with smaller samples and secondary measures. A stepped-wedge trial at the University of Wisconsin randomly assigned 66 practitioners to three 6-week sequences of ambient AI ¹². Time spent on notes fell by 0.36 hours per day, a secondary outcome ¹³. A randomized crossover trial at Duke Primary Care compared two products, and product B reduced time in notes by 3.19 minutes per day more than product A ¹⁴.

The studies also differ in what they timed. The UCLA trial used logged writing time in the note, measured as change from each physician's baseline and compared with a control group ⁹. The Wisconsin trial treated time on notes as a secondary outcome, alongside documentation quality scored with the PDSQI-9 instrument ¹³. The Duke trial reported minutes in notes per day ¹⁴, and the Singapore study measured documentation time per consultation ¹¹. These are different slices of documentation, so compare them with care.

The table's numbers are cited in the text above, and the Duke comparison is a difference between two products in one primary care site. The practical reading is that a scribe can save time, but the saving varies by product and is measured in minutes, not hours.

Burnout scores fall more than documentation minutes do

Burnout measures moved more consistently than note time in these trials. In the UCLA trial, both products showed potential improvements in burnout, task load and work exhaustion, but the authors call these secondary end points that need confirmation in larger, multicenter trials ¹⁵. DAX showed no significant change in time-in-note, but the trial's scale scores still pointed toward potential improvement for users of either product.

In Wisconsin, ambient AI use significantly reduced work exhaustion and interpersonal disengagement on the Stanford Professional Fulfillment Index, by 0.44 points on a five-point scale ¹⁶. The same abstract describes the professional fulfillment gain as nonsignificant, even though it reports a P value of .04 ¹⁶, so read the full paper before quoting that outcome. In the Duke trial, both products reduced personal and work burnout scores, and the difference between them on burnout was not meaningful ¹⁷.

Two cautions apply. Burnout here is a self-reported score, not a clinical diagnosis. And these burnout outcomes were secondary in the trials that reported them. For the measured scores, the answer to whether AI scribes reduce physician burnout is that these studies report lower scores, with the caveats above.

AI scribes make errors of omission and invention, so every note still needs review

Most AI-drafted notes were free of significant errors, but a few were not. In a 2024 pilot at UC Davis, physicians reviewed 356 of 7,545 notes made with an ambient scribe, and 94.7 percent were free from significant errors ¹⁸. The remaining notes contained errors that the physicians sorted by type. Omissions were the most common:

Omissions were the most frequent error, in 18 percent of the reviewed notes ¹⁹, followed by hallucinations in 11.5 percent ²⁰ and accidental inclusions in 9.3 percent ²¹. The most serious category matters most for safety: 19 notes, or 5.3 percent, contained an error that physicians rated as posing serious or imminent risk if it was not corrected ²². The counts come from physician reports on a survey, not a line-by-line audit against the recording. An omission and an invention carry different risks. A dropped medication change or a missed symptom can alter the plan, while an invented sentence can record a finding the patient never reported. A reviewer has to check both against the conversation, not only against the draft.

A sentence-by-sentence evaluation in npj Digital Medicine of LLM-generated clinical notes reports a 1.47 percent hallucination rate and a 3.45 percent omission rate ^{23 24}. The rates come from 18 experimental configurations, with sentences annotated by clinicians against the source conversation ²⁴. The authors were employees of Tortus AI, which sells the system tested ²³. The paper declares no competing interests ²⁴, but the rates reflect the authors' own prompts and workflows, not an independent test of every commercial scribe.

Products also differed in error counts in one exploratory audit. A Rochester Regional Health audit recorded simulated visits and gave identical recordings to three commercial scribes. The vendor with the fewest errors had 3, against 8 and 9 for the others ²⁵. The audit used two simulated visits and three vendors, so its counts cannot rank products reliably. The authors also found that clinician preference did not track those counts ²⁵.

Dictation has its own error profile, measured differently. In a 2018 analysis of dictated clinical notes, 96.3 percent of speech-recognition drafts contained at least one error ²⁶. Those figures come from clinical dictation rather than ambient scribes, so they are not directly comparable with scribe audits. The point they share is that a person must check the output before it is signed. The error-rate question and the word error rate entry explain how these measures are defined.

Patient consent and recording rules come before deployment

Recording a clinical conversation raises privacy and consent questions that the software cannot answer for you. California's Penal Code 632 punishes a person who intentionally records a confidential communication without the consent of all parties ²⁷. The statute covers confidential communications, and courts look at whether a party reasonably expected privacy, so whether a given visit is covered is a legal question ²⁷. See the guide to recording consent laws for the state-by-state picture.

Vendor contracts matter too. HIPAA's Privacy Rule lets a covered entity allow a business associate to create, receive, maintain or transmit protected health information only if it obtains satisfactory assurance that the business associate will appropriately safeguard the information ²⁸. The HIPAA guide for dictation explains what to check in a vendor's terms.

This is general information, not legal advice. Ask your compliance office or counsel before you record a single visit.

Scribes suit conversational visits, while dictation still suits procedures, radiology and letters

A scribe listens to a conversation. Dictation captures a monologue, which is what a procedure, a radiology report or a letter needs. The question on scribe versus dictation turns on which input your work produces.

Dictation still has a clear role where the clinician narrates findings, such as a procedure or a report, rather than a conversation. A radiologist describing images or a surgeon narrating an operation is already speaking the record, so there is no two-person exchange for a scribe to capture. A letter to a referring doctor is written after the visit from the clinician's own summary, which a dictated draft handles well. A recording of the whole consultation can add material that the letter does not need. Dragon Medical One is sold per user per month through resellers, which quote about $79 to $99 a month ²⁹. We found no primary list price for the ambient products covered here, so we do not quote one.

Speech technology also changes the human role. The US Bureau of Labor Statistics says advances in speech recognition and natural language processing let physicians document patient encounters in real time, which reduces the need for medical transcriptionists ³⁰. For the job itself, see the medical transcription guide.

Yap, which we make, is a dictation app, not an ambient scribe. It turns speech you choose to dictate into text on your Mac, Windows PC or phone. That fits letters, reports and notes you narrate. For the workflow in this guide, the medical dictation guide explains how the two approaches compare for clinicians who dictate.

Epic now drafts notes inside its record, and scribe prices are hard to verify

The large EHR vendors are moving into this space. Epic has launched AI drafting inside its record ¹. Vendors also publish adoption figures that are hard to check, and we found no primary price list for the scribe products covered here, so ask each vendor for its terms in writing.

Start with the question that the trials cannot answer for you: whether the time saved in your clinic outweighs the review time each draft needs. Measure both in your own workflow before you commit.

Before you choose a product, take five practical steps:

1. Ask which product and version you are evaluating, since the trials above compared named products that may have changed since.
2. Ask how the vendor measured time and errors, and whether it audited drafts against the recording.
3. Pilot the tool in your own clinic with a comparison period, as the Wisconsin, Duke and UCLA designs did.
4. Set a consent step for every recorded visit, and check it against your state's rules.
5. Sign a business associate agreement and review the vendor's data retention terms before any patient audio is recorded.

Related guides and questions

• Up: dictation for work and study.
• Related guides: medical dictation, medical transcription, dictation for nurses and therapists, how much time dictation saves at work.
• Questions answered in this guide: What is an ambient AI scribe?, How much time do doctors spend on documentation?, Do AI scribes reduce physician burnout?, How accurate are AI medical scribes?, Should doctors use an AI scribe or dictation?.
• Related on other topics: speech recognition accuracy, human versus AI transcription, AI dictation.