Ambient AI Scribes in Clinical Practice: What the Evidence, the Regulators, and the Balance Sheet Now Say

Ambient artificial intelligence documentation — software that listens to a clinical conversation and drafts the resulting note — has become the most widely deployed generative AI application in healthcare delivery. It reached that position unusually quickly, and largely ahead of the evidence base that normally accompanies a clinical technology of comparable reach.

That gap is now closing. Between late 2025 and mid-2026, the field acquired its first randomised controlled trials, its first large-scale economic analyses, its first note-quality benchmarking studies, and — in Great Britain, on 29 July 2026 — its first dedicated regulatory guidance on whether these products are medical devices at all. The picture that emerges is more interesting than either the marketing or the backlash suggests: consistent and measurable improvement in how clinicians feel, inconsistent and product-dependent improvement in how much time they save, and a set of accuracy and equity failure modes that are real but manageable with the right controls.

This article is written for professionals who must make or defend decisions about these systems: clinical leaders evaluating adoption, informatics and IT teams responsible for deployment, quality and safety officers designing oversight, procurement and finance staff modelling cost, and researchers and policy specialists tracking the field. It explains how the technology works, what the trial evidence does and does not establish, how the economics behave, where the systems fail, how regulators in the United States, Great Britain, and the European Union are treating them, and what a defensible governance model looks like.

---

Executive summary

• Adoption is now mainstream. The American Medical Association's 2026 physician survey found 81% of responding physicians using AI tools in practice, against 38% in 2023, with visit documentation among the leading use cases (American Medical Association [AMA], 2026).
• Randomised evidence supports well-being benefits more strongly than time savings. In a three-arm trial at UCLA Health, both tested products improved clinician-reported burnout, task load, and work exhaustion measures, but only one produced a statistically significant reduction in time spent writing notes (Lukac et al., 2025).
• Effect sizes are modest and product-specific. A stepped-wedge pragmatic trial found a significant reduction in work exhaustion and roughly 0.36 fewer hours per day on notes; the reduction in after-hours work was sensitive to outlier removal (Afshar et al., 2025).
• The economic case is real but not yet settled. A UCSF cohort study of about 1.2 million ambulatory encounters found adopters generated 5.8% more work relative value units per week with no rise in claim denials — though the authors caution that this may partly reflect changed coding behaviour rather than additional care delivered (Holmgren et al., 2026).
• Accuracy is good but not benign. Contemporary large language model scribes are associated with error rates around 1–3%, substantially better than legacy speech recognition, but they introduce distinctive failure modes: fabricated detail, critical omission, and misattribution of who said what (Chin et al., 2025).
• Equity risks are under-measured. Speech recognition performance disparities and interpreter-mediated encounters remain the least rigorously evaluated part of the stack.
• Regulatory clarity arrived in 2026. Britain's MHRA confirmed on 29 July 2026 that AVT products limited to transcription, summarisation, correspondence drafting, and code suggestion for clinician review are not medical devices in Great Britain; products supporting diagnosis or acting autonomously are (Medicines and Healthcare products Regulatory Agency [MHRA], 2026).
• Accreditation is becoming the practical control point. The Joint Commission and the Coalition for Health AI published governance guidance in September 2025, detailed playbooks in May 2026, and a voluntary certification programme thereafter — a faster-moving lever than device regulation for most delivery organisations.
• The binding constraint is governance, not model quality. Local validation, specialty-specific error surveillance, consent practice, and attribution of clinical responsibility for the final note determine whether deployment is safe.

---

Why documentation became generative AI's first scaled clinical use case

Clinical documentation burden is a well-characterised problem with a long empirical literature. Electronic health record adoption in the 2000s and 2010s shifted clerical work onto clinicians, and after-hours documentation — often labelled "pyjama time" — became a reproducible correlate of exhaustion and attrition.

The intermediate solutions each had structural limits. Human scribes improved note quality and satisfaction but cost too much to scale across a workforce. Speech recognition dictation removed typing but not composition, and carried error rates in the region of 7–11% because of medical vocabulary and accent variability (Chin et al., 2025). Earlier "digital scribes" that combined automation with vendor-side human editing produced inconsistent efficiency gains.

Two developments changed the calculus. Large language models made it feasible to convert unstructured, multi-speaker conversation into a structured note without a template-driven interface. Simultaneously, the risk profile of the task was attractive: drafting an administrative artefact that a licensed clinician reviews and signs sits outside the diagnostic and therapeutic decision path, which lowers both the regulatory and the clinical-risk barrier to entry. By the time the first randomised trials were designed, investigators noted that more than 50 vendors were offering functional products (Lukac et al., 2025).

Callout — why this matters for evaluation design Because ambient documentation is administrative rather than diagnostic, it largely escaped premarket regulatory evaluation. Responsibility for establishing safety and effectiveness therefore falls on the deploying organisation, not the regulator. That inversion is the single most important structural fact about this technology.

---

How ambient documentation systems work

A modern ambient scribe is a pipeline, not a model. Understanding the stages matters because each contributes a distinct failure mode.

Capture. Audio is collected from a clinician's mobile device, a room microphone array, or a telehealth stream. Capture quality — background noise, distance, overlapping speech — sets a ceiling on everything downstream.

Transcription and diarisation. Automatic speech recognition converts audio to text; diarisation attributes utterances to speakers. Diarisation errors are the origin of misattribution problems, where a patient's speculation is recorded as a clinician's assessment, or vice versa.

Contextual grounding. Better systems retrieve patient context from the EHR — problem list, medications, recent results — so that the draft is consistent with the record rather than with the conversation alone.

Generation. A large language model produces a structured draft in the requested format, commonly a SOAP or specialty-specific template, sometimes with suggested diagnosis codes or orders staged for review.

Clinician review and attestation. The clinician edits and signs. This is the safety-critical step, and the one most vulnerable to erosion over time as trust in the draft grows.

Retention and feedback. Audio and transcript retention policies, edit-distance telemetry, and error reporting channels determine whether the system can be monitored at all.

Figure 1 — proposed original diagram Title: "Ambient Clinical Documentation Pipeline and Its Failure Points" Purpose: show the end-to-end data flow and locate the characteristic error mode introduced at each stage. Layout: horizontal left-to-right flow, six primary boxes connected by solid arrows: Encounter audio capture → Speech recognition + speaker diarisation → EHR context retrieval → LLM note generation → Clinician review & attestation → Signed note in EHR. Secondary elements: beneath each of the first four boxes, a downward dashed arrow to a red-outlined "failure mode" label: Noise / overlapping speech, Misattribution / accent-related error, Stale or missing context, Fabrication, omission, over-confident phrasing. Beneath the review box, an amber label: Automation bias / declining edit rate. Right-hand vertical lane: a shaded governance column spanning the full width, labelled "Monitoring layer", with upward dotted arrows from each stage feeding three boxes: Edit-distance telemetry, Specialty error sampling, Clinician incident reporting. Visual hierarchy: primary flow in solid dark strokes; failure modes de-emphasised in outline only; monitoring layer as a background band to signal that it wraps the whole pipeline. Caption: "Each stage of an ambient documentation pipeline contributes a distinct error mode; monitoring must span the pipeline rather than sample only the final note."

---

What the randomised evidence actually shows

Until late 2025, the evidence base consisted mainly of pilots, pre-post studies, and satisfaction surveys — designs that are highly susceptible to enthusiasm effects among voluntary early adopters. Three categories of stronger evidence have since appeared.

Randomised trials

In a parallel three-group pragmatic randomised trial at UCLA Health, 238 outpatient physicians across 14 specialties were assigned to one of two commercial scribes or to usual care between 4 November 2024 and 3 January 2025. Only one of the two products produced a statistically significant reduction in time-in-note against control — a decrease of 9.5% (95% CI, −17.2% to −1.8%; P = 0.02) — while the other showed no significant change on the primary outcome. Both, however, were associated with improvements in the Mini-Z work-life measure, physician task load, and work exhaustion. Participants reported clinically significant inaccuracies "occasionally" on a five-point scale, and the authors explicitly flagged that the secondary findings require confirmation in larger multicentre trials (Lukac et al., 2025).

A separate 24-week stepped-wedge pragmatic trial randomised 66 practitioners across ambulatory clinics in two US states to three six-week sequences. Ambient AI use was associated with a significant reduction in work exhaustion and interpersonal disengagement (−0.44 points; 95% CI, −0.62 to −0.25; P < 0.001) and a decrease of 0.36 hours per day in time spent on notes. The reduction in work-outside-work did not survive removal of the top 3% of daily observations — a useful reminder that aggregate time metrics can be driven by a small number of extreme days (Afshar et al., 2025).

A randomised crossover trial of two ambient products among 160 outpatient clinicians at a tertiary academic centre has since been published in the Journal of the American Medical Informatics Association, extending the comparative literature (Chowdhury et al., 2026).

Large-scale real-world deployment

The Permanente Medical Group's Northern California rollout remains the largest published operational account: 7,260 physicians used ambient scribes across roughly 2.58 million encounters between 16 October 2023 and 28 December 2024, with 3,447 physicians using the tool in at least 100 encounters (Tierney et al., 2025). This is observational and organisation-reported rather than controlled, but it establishes that sustained voluntary use at scale is achievable — a non-trivial finding given the abandonment rates of many clinical technologies.

Note-quality benchmarking

A blinded comparison of 97 outpatient encounters across five specialties scored AI-generated drafts against physician-authored reference notes using an adapted Physician Documentation Quality Instrument. Reference notes scored marginally higher overall (4.25 versus 4.20 out of 5; P = 0.04) and on accuracy, succinctness, and internal consistency; AI drafts scored higher on thoroughness and organisation. Hallucinations were identified in 31% of AI notes and — notably — in 20% of physician-authored notes (P = 0.01). Reviewers nonetheless preferred the AI drafts more often than the human ones (Palm et al., 2025). Readers should weigh that several authors were affiliated with a scribe vendor, which is disclosed in the paper and is a material consideration when interpreting the result.

Table 1 — Evidence map for ambient clinical documentation (original synthesis)

---

The economics: a defensible case with an open question

Subscription pricing for ambient documentation typically falls in the range of roughly US$200–600 per clinician per month, which makes the return-on-investment question unavoidable at scale (Rotenstein & Melnick, 2026).

The most substantial economic analysis to date examined 1,565 physicians and nearly 1.2 million ambulatory encounters at a single academic health system between January 2023 and April 2025. Among the 698 physicians (44.6%) who adopted the tool, weekly work relative value units rose by 1.81 — a 5.8% increase, equivalent to roughly US$3,044 in additional annual revenue per physician at 2025 Medicare payment rates — alongside a 2.8% increase in weekly encounters and no increase in claim denials (Holmgren et al., 2026).

The authors are careful about what this does and does not demonstrate. It is not established whether the RVU increase reflects additional clinical services delivered, more complete capture of services already delivered, or a shift in coding intensity driven by more thorough AI-generated documentation. Those three explanations have very different implications for payers, for compliance functions, and for the aggregate cost of care. This is the most consequential open empirical question in the field, and organisations deploying at scale should expect payer scrutiny of documentation-driven coding shifts.

A second economic caution: the well-being benefits found in trials are real but are not automatically monetisable. Retention effects, if they exist, will take years to demonstrate and are confounded by everything else happening in a workforce.

---

Where these systems fail

Accuracy failure modes

LLM-based scribes report overall error rates in the region of 1–3%, well below the 7–11% typical of older dictation systems, but the errors differ in kind rather than only in frequency (Chin et al., 2025). The characteristic modes are:

• Fabrication — plausible clinical detail that was never discussed.
• Critical omission — a symptom, allergy, or safety-netting instruction dropped from the draft.
• Misattribution — a statement assigned to the wrong speaker, which can convert patient speculation into clinician assessment.
• False negation — documenting that something was denied when it was never asked.
• Confidence inflation — verbal uncertainty rendered as declarative statement.

The last two are the most dangerous because they are the hardest to detect on review: a fluent, well-organised note containing a confidently phrased false negative reads as correct.

Equity and language

This is the least well-evaluated dimension. Speech recognition systems have documented performance disparities across speaker groups, including higher error rates for African American speakers (Chin et al., 2025). In a study of simulated English–Spanish encounters, ambient scribes propagated interpreter errors into the resulting notes in the majority of tested cases across two vendors — a small, simulated study, but one that identifies a systemic vulnerability in interpreter-mediated care rather than a product-specific defect (Rodriguez & Ali, 2026). Code-switching, dialect variation, and mixed-language consultations remain poorly characterised.

The review paradox

Ambient scribes shift work from composition to verification. Verification is cognitively different: it is easier to perform badly without noticing. As draft quality improves, the expected value of careful review falls in the clinician's subjective assessment while the consequence of the residual error stays constant. Declining edit rates over time should therefore be monitored as a potential safety signal, not celebrated as an efficiency gain.

The AMA's 2026 survey found that 88% of responding physicians expressed concern about skill erosion, particularly among clinicians with fewer than ten years of experience (AMA, 2026). Whether documentation-specific skills atrophy in a way that matters clinically is currently a hypothesis, not a finding.

---

The regulatory and accreditation landscape

Great Britain: the clearest position

On 29 July 2026, the MHRA, working with NHS England, published guidance on the regulatory status of ambient voice technology products. It confirms that AVT products intended solely to transcribe, summarise clinical conversations, draft correspondence, or suggest clinical codes for a clinician to review are not medical devices under the current Great Britain framework, while products intended to support diagnosis, treatment, or prevention — or to take automated action such as placing orders without clinician review — are regulated as medical devices (MHRA, 2026; NHS England, 2026). The guidance interprets existing law rather than changing it, applies to England, Wales, and Scotland but not Northern Ireland, and is accompanied by illustrative paired examples that clarify where the line falls. NHS England separately maintains a self-certification registry of AVT suppliers to support consistent adoption.

The practical consequence is important: a product can cross into device territory through feature expansion. Governance must therefore track intended purpose over time, not only at procurement.

United States: device law plus accreditation

The FDA's existing framework is device-centric. Its final guidance on predetermined change control plans (December 2024) established a mechanism for pre-authorising specified model modifications, and its January 2025 draft guidance on AI-enabled device software functions set out total product lifecycle expectations; the amended Quality Management System Regulation, aligning 21 CFR Part 820 with ISO 13485:2016, took effect on 2 February 2026 (US Food and Drug Administration [FDA], 2025). Documentation-only scribes generally sit outside device jurisdiction on the same intended-use logic the MHRA articulates.

That leaves accreditation and voluntary standards as the operative control. The Joint Commission and the Coalition for Health AI published Responsible Use of AI in Healthcare guidance on 17 September 2025, covering AI policies, local validation, ongoing monitoring, transparency, bias mitigation, and workforce training; CHAI released detailed governance playbooks on 27 May 2026, developed with more than 100 healthcare organisations; and the Joint Commission subsequently opened a voluntary Responsible Use of AI in Healthcare certification, available to organisations whether or not they hold Joint Commission accreditation (Joint Commission & Coalition for Health AI [CHAI], 2025; CHAI, 2026).

European Union: a moved deadline that is not a reprieve

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 with staggered application. The Digital Omnibus on AI, proposed on 19 November 2025, reached provisional political agreement on 7 May 2026, was endorsed by the European Parliament on 16 June 2026, and received Council approval on 29 June 2026. It defers obligations for stand-alone Annex III high-risk systems to 2 December 2027 and for AI embedded in products regulated under Annex I — which includes medical devices — to 2 August 2028. Article 50 transparency obligations were not deferred and apply from 2 August 2026, with a narrow transition to 2 December 2026 for certain systems already on the market (European Commission, 2026).

Table 2 — How three jurisdictions treat documentation-only ambient AI (original synthesis, current as of 1 August 2026)

---

Latest developments (dated)

• 17 September 2025 — Joint Commission and CHAI publish initial Responsible Use of AI in Healthcare guidance.
• 22 October 2025 — First blinded PDQI-based comparison of ambient versus physician-authored notes published in Frontiers in Artificial Intelligence.
• 19 November 2025 — European Commission proposes the Digital Omnibus on AI.
• 26 November 2025 — First randomised trial of ambient scribes published in NEJM AI.
• 2 January 2026 — Cohort study of scribe adoption and physician financial productivity published in JAMA Network Open.
• 2 February 2026 — FDA's amended Quality Management System Regulation takes effect.
• 12 March 2026 — AMA publishes its 2026 Physician Survey on Augmented Intelligence (fielded 15 January–2 February 2026; n = 1,692).
• 27 May 2026 — CHAI releases eight governance playbooks; Joint Commission announces voluntary AI certification.
• 29 June 2026 — Council of the EU gives final approval to the Digital Omnibus on AI.
• 29 July 2026 — MHRA publishes ambient voice technology guidance for Great Britain.
• 2 August 2026 — EU AI Act Article 50 transparency obligations become applicable.

Status note: items above are documented publications or formal regulatory actions. Claims about long-term retention benefits, clinical outcome improvement, or skill erosion remain hypotheses without adequate evidence either way.

---

Implementation: governance that survives contact with a clinic

Figure 2 — proposed original diagram Title: "Ambient Documentation Governance Lifecycle" Purpose: give deploying organisations a repeatable oversight cycle rather than a one-off procurement checklist. Layout: a closed circular flow of six nodes, arrows clockwise: 1. Intended-use definition → 2. Local validation by specialty → 3. Consent & privacy design → 4. Controlled rollout → 5. Continuous monitoring → 6. Periodic re-review → back to node 1. Node annotations (small text beside each): 1 — what the tool must not do; 2 — sampled note audit against audio, per specialty; 3 — recording notice, retention, opt-out path; 4 — cohort-based, with control comparison where feasible; 5 — edit distance, error reports, denial rates, equity stratification; 6 — re-validate after any model or vendor change. Central element: a labelled hub reading "Named clinical owner" with thin lines to all six nodes, indicating accountability rather than process flow. Caption: "Vendor model updates reset validity; governance must be cyclical rather than terminal."

Four controls distinguish organisations that deploy safely from those that deploy quickly.

Define intended use narrowly and in writing. The regulatory line in Great Britain, and the analogous intended-use logic in the United States and European Union, turns on what the product is for. Document what the tool is permitted to do, and treat vendor feature expansion into order suggestion or diagnostic support as a change requiring fresh review.

Validate locally, by specialty. Performance varies by encounter type. Narrative, longitudinal consultations transcribe better than fast, interrupted, or procedural encounters. A sampled audit comparing signed notes against source audio, stratified by specialty and by patient language, is the minimum credible validation. Local validation is explicitly among the elements the Joint Commission and CHAI identify (Joint Commission & CHAI, 2025).

Design consent and retention deliberately. Patients should be told that the conversation is being recorded, by whom, for what purpose, and for how long the audio is retained. Recording-consent law varies by jurisdiction — including between US states — and a single national policy is unlikely to be adequate for a multi-state system. Retention defaults should be set to the shortest period compatible with quality assurance.

Monitor the review step, not just the model. Track edit distance between draft and signed note over time, per clinician and per specialty; run periodic blinded audits; maintain a low-friction error reporting channel; and stratify equity-relevant metrics by patient language and interpreter use. A falling edit rate combined with a stable error rate is a warning, not a win.

Table 3 — A minimum monitoring set for ambient documentation (original framework)

---

Frequently misunderstood, and demonstrably wrong

---

Practical takeaways

For clinical and departmental leaders

1. Insist on a specialty-specific pilot with a control comparison before enterprise rollout; borrowed evidence from another specialty is weak evidence.
2. Set an explicit review standard — the signed note is the clinician's professional statement regardless of who drafted it — and reinforce it in training rather than assuming it.
3. Watch for confidence inflation and false negation specifically; generic "check the note" instructions do not surface these.

For informatics and IT teams 4. Instrument edit distance and note-versus-audio auditing from day one; retrofitting telemetry after rollout is significantly harder. 5. Treat every vendor model update as a revalidation trigger and negotiate advance notice of material model changes into the contract. 6. Test performance explicitly on interpreter-mediated and accented-speech encounters before assuming parity.

For quality, safety, and compliance functions 7. Map the deployment against the Joint Commission/CHAI governance elements or an equivalent framework, and record where the organisation is non-compliant rather than leaving gaps implicit. 8. Monitor coding intensity distributions proactively; documentation-driven billing shifts are foreseeable and should be explainable before a payer asks.

For procurement and finance 9. Model total cost including subscription, integration, training, monitoring staff time, and revalidation — not licence cost alone. 10. Treat productivity gains as a hypothesis to be tested locally, given that the published RVU findings come from voluntary adopters at a single system.

---

Key insights

1. Ambient documentation scaled faster than it was evaluated; the evidence is now catching up, and it is more equivocal than early enthusiasm suggested.
2. Well-being improvements are the most reproducible finding across randomised trials.
3. Time savings are real but modest, and vary meaningfully between products.
4. Product differences within the category are large enough that "ambient scribes work" is not a useful procurement conclusion.
5. The financial productivity signal is positive but its mechanism — more care versus more complete coding — is unresolved.
6. Error rates are lower than legacy dictation, but the error types are more insidious.
7. Equity performance is the least-studied and most consequential evidence gap.
8. Regulators in Great Britain have drawn a clear, intended-use-based line; feature expansion can cross it.
9. Where products are not devices, accreditation frameworks and local governance become the operative safety mechanism.
10. The review step is where safety actually lives, and it degrades quietly.

---

Frequently asked questions

What is an ambient AI scribe? Software that passively records a clinical encounter, transcribes it, and uses a language model to draft a structured clinical note for the clinician to review, edit, and sign.

Are ambient AI scribes regulated as medical devices? In Great Britain, products limited to transcription, summarisation, correspondence drafting, or code suggestion for clinician review are not medical devices; products supporting diagnosis or treatment, or acting without clinician review, are (MHRA, 2026). Similar intended-use logic applies in the United States and European Union.

Do AI scribes reduce physician burnout? Randomised trials report consistent improvements in burnout-related measures such as work exhaustion and task load. The effects are self-reported and measured over weeks to months, so durability is not yet established (Lukac et al., 2025; Afshar et al., 2025).

How much documentation time do they actually save? Published randomised estimates cluster around a 9.5% reduction in time-in-note for one product and roughly 0.36 fewer hours per day on notes in a separate trial, with one product showing no significant saving. Expect tens of minutes per day, not hours.

Do AI scribes increase revenue? One cohort study found 5.8% higher weekly work RVUs among adopters with no rise in claim denials, worth roughly US$3,044 per physician annually at 2025 Medicare rates — but whether this reflects more care or more complete coding is unresolved (Holmgren et al., 2026).

What do they cost? Subscription pricing commonly falls between roughly US$200 and US$600 per clinician per month, before integration, training, and monitoring costs.

How accurate are AI-generated clinical notes? Contemporary LLM-based systems report error rates around 1–3%, compared with 7–11% for legacy speech recognition. One blinded study detected hallucinations in 31% of AI drafts and 20% of physician-written comparison notes.

Can the AI-generated note be signed without review? No. Clinician review and attestation is both the safety control and, in most frameworks, the basis on which the product avoids classification as a medical device.

Who is liable if an AI-drafted note contains an error? The signing clinician remains professionally responsible for the content of the record. Organisational and vendor liability depends on contract terms and jurisdiction; specific arrangements should be reviewed with legal counsel.

Do patients need to consent to being recorded? Practice should be to inform patients and offer a means of declining. Recording-consent law varies by jurisdiction, including between US states, and organisational policy should reflect local requirements.

Do AI scribes work equally well for all patients? Not demonstrably. Speech recognition disparities across speaker groups are documented, and a simulated study found ambient scribes propagated interpreter errors into notes in the majority of tested cases. This is an active evidence gap.

How should a health system pilot an ambient scribe? Define intended use narrowly, pilot by specialty with a comparison group, audit signed notes against source audio, instrument edit-distance telemetry, and pre-commit to a decision rule — including a structured exit if the tool does not fit.

Will ambient AI make junior clinicians worse at documentation? This is a widely voiced concern — 88% of physicians in the AMA's 2026 survey expressed worry about skill erosion — but it is currently a hypothesis without longitudinal evidence.

What changes on 2 August 2026 in the EU? The AI Act's Article 50 transparency obligations become applicable. High-risk obligations were deferred: Annex III stand-alone systems to 2 December 2027, and AI embedded in Annex I regulated products including medical devices to 2 August 2028.

Are ambient scribes the same as clinical decision support? No. Decision support influences a clinical judgment; documentation tools record one. The distinction is both the regulatory boundary and the clinical risk boundary, and it erodes when scribes begin suggesting orders or assessments.

What single metric best indicates a deployment is going wrong? A falling edit rate between draft and signed note that is not accompanied by an independently measured fall in error rate.

---

Glossary

Ambient voice technology (AVT) — the regulatory term used in Great Britain for products that capture and process clinical conversations, including AI scribes.

Automatic speech recognition (ASR) — conversion of spoken audio into text; the first processing stage of an ambient scribe.

Automation bias — the tendency to accept machine-generated output with insufficient scrutiny.

Diarisation — attribution of transcribed speech to individual speakers.

Edit distance — a quantitative measure of how much a clinician changed the AI draft before signing; a proxy for review effort.

Hallucination — content generated by a model that is fluent and plausible but unsupported by the source material.

PDQI-9 — Physician Documentation Quality Instrument, a validated nine-item framework for scoring clinical note quality.

PCCP (predetermined change control plan) — an FDA mechanism allowing pre-authorised modifications to an AI-enabled device without a new marketing submission.

QMSR — the FDA's Quality Management System Regulation, which aligned 21 CFR Part 820 with ISO 13485:2016 effective 2 February 2026.

RUAIH — Responsible Use of AI in Healthcare, the Joint Commission/CHAI guidance and associated voluntary certification.

SaMD (software as a medical device) — software intended for a medical purpose that is not part of a hardware device.

Stepped-wedge design — a trial design in which participants cross over from control to intervention in randomised sequence.

Time-in-note — EHR-derived measure of active time spent authoring a note; a common primary outcome in documentation studies.

wRVU (work relative value unit) — a standardised measure of physician work used in US payment and productivity accounting.

Work outside work (WoW) — EHR time recorded outside scheduled clinical hours.

---

References

Academic papers

Afshar, M., et al. (2025). A pragmatic randomized controlled trial of ambient artificial intelligence to improve health practitioner well-being. NEJM AI, 2(12). https://doi.org/10.1056/AIoa2500945

Chin, A. T., et al. (2025). Beyond human ears: Navigating the uncharted risks of AI scribes in clinical practice. npj Digital Medicine, 8. https://doi.org/10.1038/s41746-025-01895-6

Chowdhury, A., Casey, M., Wilson, J., Pollak, K. I., Goldstein, B. A., Bedoya, A., & Poon, E. G. (2026). Comparing ambient scribes: A randomized crossover clinical trial addressing ambient scribe technologies' impact on physician burnout. Journal of the American Medical Informatics Association, 33(5), 990–999. https://doi.org/10.1093/jamia/ocag018

Holmgren, A. J., Fenton, C. L., Thombley, R., Soleimani, H., Croci, R., DeMasi, O., Byron, M. E., Murray, S. G., Adler-Milstein, J., & Yazdany, J. (2026). Ambient artificial intelligence scribes and physician financial productivity. JAMA Network Open, 9(1), e2553233. https://doi.org/10.1001/jamanetworkopen.2025.53233

Lukac, P. J., Turner, W., Vangala, S., Chin, A. T., Khalili, J., Shih, Y.-C. T., Sarkisian, C., Cheng, E. M., & Mafi, J. N. (2025). Ambient AI scribes in clinical practice: A randomized trial. NEJM AI, 2(12). https://doi.org/10.1056/AIoa2501000

Palm, E., Manikantan, A., Mahal, H., Belwadi, S. S., & Pepin, M. E. (2025). Assessing the quality of AI-generated clinical notes: Validated evaluation of a large language model ambient scribe. Frontiers in Artificial Intelligence, 8, 1691499. https://doi.org/10.3389/frai.2025.1691499

Rodriguez, A., & Ali, E. (2026). Propagation of interpreter errors by ambient AI scribes. JMIR Medical Informatics, 14, e88734. https://doi.org/10.2196/88734

Rotenstein, L. S., & Melnick, E. R. (2026). Ambient AI scribes — What is the return on investment? JAMA Network Open, 9(1), e2553238. https://doi.org/10.1001/jamanetworkopen.2025.53238

Shah, S. J., Devon-Sand, A., Ma, S. P., et al. (2025). Ambient artificial intelligence scribes: Physician burnout and perspectives on usability and documentation burden. Journal of the American Medical Informatics Association, 32(2), 375–380. https://doi.org/10.1093/jamia/ocae295

Tierney, A. A., Gayre, G., Hoberman, B., Mattern, B., Ballesca, M., Wilson Hannay, S., Castilla, K., Lau, C., Kipnis, P., Liu, V., & Lee, K. (2025). Ambient artificial intelligence scribes: Learnings after 1 year and over 2.5 million uses. NEJM Catalyst Innovations in Care Delivery. https://doi.org/10.1056/CAT.25.0040

Tierney, A. A., Gayre, G., Hoberman, B., Mattern, B., Ballesca, M., Kipnis, P., Liu, V., & Lee, K. (2024). Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catalyst Innovations in Care Delivery, 5(3). https://doi.org/10.1056/CAT.23.0404

Government sources and regulatory guidance

European Commission. (2026). Digital Omnibus on AI: Targeted amendments to Regulation (EU) 2024/1689. Adopted by the European Parliament, 16 June 2026, and approved by the Council of the European Union, 29 June 2026.

Medicines and Healthcare products Regulatory Agency. (2026, July 29). Ambient voice technology-enabled products [Guidance]. GOV.UK. https://www.gov.uk/government/publications/ambient-voice-technology-enabled-products

Medicines and Healthcare products Regulatory Agency. (2026, July 29). MHRA clarifies regulatory status of ambient voice technologies used in the NHS [Press release]. GOV.UK. https://www.gov.uk/government/news/mhra-clarifies-regulatory-status-of-ambient-voice-technologies-used-in-the-nhs

NHS England. (2026). Medical device regulation for ambient voice technology products. https://www.england.nhs.uk/long-read/medical-device-regulation-for-ambient-voice-technology-products/

US Food and Drug Administration. (2025, January 6). Artificial intelligence-enabled device software functions: Lifecycle management and marketing submission recommendations [Draft guidance]. https://www.fda.gov/media/184856/download

US Food and Drug Administration. (n.d.). Artificial intelligence in software as a medical device. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-software-medical-device

Standards and accreditation frameworks

Coalition for Health AI. (2026, May 27). Responsible AI governance playbooks. https://www.chai.org

International Organization for Standardization. (2016). ISO 13485:2016 — Medical devices: Quality management systems — Requirements for regulatory purposes.

Joint Commission & Coalition for Health AI. (2025, September 17). Responsible use of AI in healthcare (RUAIH). https://www.jointcommission.org

Joint Commission. (2026, May). Responsible Use of AI in Healthcare certification. https://www.jointcommission.org/en-us/certification/responsible-use-of-ai-in-healthcare

Industry and professional association reports

American Medical Association. (2026, March 12). 2026 physician survey on augmented intelligence. AMA Center for Digital Health and AI. https://www.ama-assn.org/practice-management/digital-health/physician-survey-augmented-intelligence

---

Editorial note: Statistics, dates, and regulatory positions in this article were verified against primary sources as of 1 August 2026. Regulatory guidance in this field is changing rapidly; readers should confirm the current status of the FDA draft guidance on AI-enabled device software functions and of EU AI Act implementing measures before relying on them for compliance decisions. This article is informational and does not constitute legal, regulatory, or clinical advice.