health

Best AI Medical Scribes

The best AI medical scribes in 2026 — Abridge, Microsoft Dragon Copilot, Ambience Healthcare, Nabla, Suki, Freed, Heidi, Lyrebird Health, DeepScribe, Commure, Doximity Scribe, Sunoh.ai and Corti compared on published pricing, EHR integration, language support and compliance, set against what the randomised trials and multi-site cohorts actually measured.

Updated September 2026

Quick answer: For enterprise health systems, the strongest AI medical scribe in September 2026 is Abridge, which won Best in KLAS 2026 for Ambient Speech with an overall performance score of 94.7/100 (Becker’s Hospital Review, 4 February 2026) — but it publishes no price and sells only to health systems. For a solo clinician or small practice, Freed is the best pick and the most transparent vendor in the category, at $39/month for 40 notes or $79/month unlimited. The best genuinely free option is Heidi, whose free tier gives unlimited standard-template notes; US clinicians can also use Doximity Scribe at no cost, with no EHR integration and notes that auto-delete after 29 days. Outside the United States, Corti is the only vendor here registered as a medical device in the EU, and Heidi is the only one running compliance programmes across the US, UK, EU, Canada and Australia/New Zealand. In Australian general practice specifically, Lyrebird Health is the strongest pick, because it is built around Australian workflows and is free for Best Practice Premier subscribers.

The caveat that governs this entire category: the measured benefits are much smaller than the marketing. The largest study to date — 8,581 ambulatory clinicians across five academic health systems — found ambient scribes saved 16.0 minutes of documentation time per 8 scheduled patient hours, generated $167.37 per clinician per month in marginal billing revenue, and produced no statistically significant reduction in after-hours “pajama time” (JAMA, 1 April 2026). In the only three-arm randomised trial, Microsoft’s DAX Copilot did not reach statistical significance against control (NEJM AI, 26 November 2025). And in controlled testing of five platforms against simulated encounters, a mean of 26.3% of key clinical elements were omitted or recorded incorrectly, with a mean of 3.0 errors per case carrying potential for moderate-to-severe harm (Mayo Clinic Proceedings: Digital Health, 2025).

Pricing, language support and compliance claims on this page were checked against vendor primary sources on 18 September 2026. Only six of the thirteen vendors covered publish a price at all. Where a figure exists only in third-party reporting, it is labelled as such; where we could not verify a figure from a credible source, we write “data not available” rather than repeat a guess.

This page covers clinical documentation — tools that listen to a patient encounter and draft a note for the medical record. That is a different job from meeting transcription, podcast transcription and general speech-to-text APIs, which are covered on best AI for transcription. The vendors barely overlap, and the buying criteria do not overlap at all: a general transcription tool is judged on word error rate and price per hour, while a clinical scribe is judged on EHR write-back, coding accuracy, recording consent and regulatory status.

The current state of AI medical scribes: September 2026

Ambient clinical documentation is the fastest-adopted clinical technology since the electronic health record itself — and 2026 is the year the evidence caught up with the claims.

Adoption is real but narrower than headlines suggest. In the American Medical Association’s physician survey fielded January 2026 and published 12 March 2026 (approximately 1,700 physicians), 29% of physicians reported using ambient or voice documentation, up from 20% in April 2025 (AMA). The widely quoted “81% of physicians use AI” figure from the same survey covers any professional AI use, including literature summarisation — it is not a scribe-adoption number, and the two are routinely conflated. In Australia, the Royal Australian College of General Practitioners recorded GP AI-scribe use rising from 22% in August 2024 to 40% by November 2025 (RACGP).

Five shifts define the current moment.

1. The evidence base matured, and it is sobering. Three serious studies landed between November 2025 and April 2026 — a randomised trial, a randomised crossover trial and a large multi-site cohort. All three converge on modest time savings, none found a reduction in after-hours EHR work, and the largest characterised the revenue effect as “nominal”. The detail is in the next section, and it is the single most useful thing on this page.

2. Pricing went dark at the top of the market. Six of the thirteen vendors here publish no price whatsoever. Both abridge.com/pricing and nabla.com/pricing return 404. Microsoft publishes no per-clinician price for any Dragon Copilot SKU. Every dollar figure circulating for Abridge, Nabla, Suki, Ambience and DeepScribe originates from reseller listings or competitor content-marketing blogs, not from the vendors.

3. Microsoft tried to go down-market and retreated in three months. The Dragon Copilot Physician Practice licence launched 1 February 2026 for practices of up to 100 clinicians through CSP partners, and reached end of sale on 1 May 2026 (Microsoft Partner Center). Microsoft never published a price for it. Any comparison still listing it as a buying option is out of date.

4. The UK reversed its regulatory position, and the litigation wave started in the US. On 29 July 2026 the MHRA ruled that AI summarisation of a clinical conversation is not a medical device — reversing NHS England’s April 2025 position. Meanwhile two US class actions now allege that ambient recording without all-party consent violates state wiretapping law. Both are covered below.

5. Accuracy failures are now documented by auditors, not just researchers. Ontario’s Auditor General reported on 12 May 2026 on government evaluations of 20 AI medical transcription programs run during a Supply Ontario procurement: 9 of 20 produced hallucinations, fabricating clinical information such as referrals and test orders that had not happened; 12 of 20 captured a different drug from the one the doctor prescribed; and 17 of 20 missed key details about patients’ mental health in at least one test (CBC).

What the evidence actually shows

This is the section most comparison pages skip. Every figure below is from a peer-reviewed source, with the study design stated, because design is what separates a real finding from a press release.

Documentation time saved

The JAMA cohort is the most generalisable figure, and it comes with a dose effect that matters commercially: clinicians who used the scribe in at least 50% of visits saved 27.3 minutes of documentation time, but only about 32% of adopters used it that heavily. The headline 16-minute average is diluted by light users. If you are building a business case, the number you should model is the one for the adoption rate you can realistically achieve, not the average.

The after-hours result that nobody markets

Three independent studies looked for a reduction in “pajama time” — EHR work done outside scheduled hours, the specific burden most often cited as the reason to buy these tools. None of them found one.

• JAMA 2026 (n=8,581): after-hours EHR time “did not change significantly.”
• JAMIA 2026 (Duke crossover, n=160): “There was no evidence of a change in average pajama time after implementation of ambient listening tools.” Both arms showed small non-significant increases.
• JAMIA 2026, same study: “neither tool was able to demonstrate a decrease in time spent outside of scheduled hours on any EHR activity.”

The most plausible reading, offered by the JAMA authors themselves, is that saved minutes get reallocated — to the inbox, to chart review, or to patient face time — rather than recovered as free time.

Note accuracy and error rates

The strongest accuracy evidence comes from controlled testing rather than vendor claims. Researchers ran five ambient scribe platforms through 14 simulated ambulatory encounters and scored the output (Mayo Clinic Proceedings: Digital Health, 2025):

• Mean share of key clinical elements omitted or recorded incorrectly, across platforms: 26.3% (95% CI 17.0–31.0)
• Mean errors per case with potential for moderate-to-severe harm: 3.0
• Proportion of correctly reported elements that were consistently correct across all platforms: 35.8%
• Errors of omission made up 76.3% of all errors

That last figure is the important one. Omissions are structurally harder to catch than hallucinations, because nothing appears on the page to flag them — a clinician reviewing a fluent, plausible note has no cue that a detail is missing. This is also what the Duke trial’s free-text safety comments described: over-summarisation and omissions in the assessment and plan, especially in subspecialty notes, and speaker-attribution failure when more than two people were in the room.

Two further findings sharpen the picture:

• AI drafts are systematically over-confident. Across 62,811 paired note sections, clinicians introduced hedging language into AI drafts more often than they removed it — the drafts stated things with more certainty than clinicians were willing to attest to (arXiv:2606.00018, 14 April 2026; preprint, not peer-reviewed).
• Clinicians edit most drafts, and edit behaviour is personal rather than specialty-driven. In a study of 23,760 notes from 225 clinicians, 84.4% of notes were edited before sign-off, and variation was driven overwhelmingly by the individual clinician rather than their specialty or their vendor (medRxiv, 13 January 2026; preprint, not peer-reviewed).

There is also an equity dimension. In simulated English and Spanish encounters, ambient scribes propagated medical-interpreter errors directly into clinical notes (JMIR Medical Informatics, 28 July 2026) — meaning accuracy risk concentrates in patients with limited English proficiency.

Burnout improves — but probably not for the reason vendors say

The strongest burnout figure in the field comes from a study of 263 clinicians across six health systems using Abridge, where burnout fell from 51.9% to 38.8% after 30 days (JAMA Network Open, 2 October 2025). Two honest caveats: it was a quality-improvement study with no control group and no randomisation, and the “74% burnout reduction” figure circulating in vendor material is a trade-press relative restatement of this result, not a number in the paper. Do not rely on it.

The controlled evidence is more modest. The UCLA randomised trial found roughly 7% improvement in burnout scores in both scribe arms, which the authors called “modest” and said required confirmation. The Duke crossover trial found no statistically significant difference between the two products on any burnout subscale, and patient-related burnout did not improve for either.

Most usefully, both sets of authors reject the mechanism the marketing implies. Rebecca Mishuris, senior author on the JAMA cohort and Chief Health Information Officer at Mass General Brigham, said the time reductions are “modest” and “unlikely to fully account for the massive drops in physician burnout often associated with these tools.” The Duke authors put it directly: the driver is “the improved experience rather than efficiency of the work.” Clinicians feel better because they are looking at the patient instead of the screen — not because they are going home earlier. That is a real benefit. It is just not the one on the ROI slide.

The ROI question is unresolved

The Peterson Health Technology Institute assessed ambient scribes and found them “potentially effective at reducing clinician documentation time and cognitive load” but identified “gaps in evidence regarding the impact of ambient scribes on productivity and financial performance” (PHTI, 25 March 2025). PHTI has not published a graded assessment of the category.

There is a live debate about where any revenue gain comes from. PHTI executive director Caroline Pearson told Advisory Board in April 2026 that scribes are “increasing coding intensity. One hundred percent” (Advisory Board). Payers have responded with outlier programmes that downgrade Level 4 and 5 visits by one level, worth roughly $50 per visit. Health economist Daniel Polsky of Johns Hopkins, quoted in the same piece, called the resulting dynamic “a zero-sum game.”

Best AI medical scribes compared

Vendors are ordered by the buyer they actually serve, because that is the first filter — six of these will not sell to a solo clinician, and three of them will not integrate with Epic.

1. Abridge — best for enterprise health systems

Price: Not published. abridge.com/pricing returns 404; all routes go to sales. EHR: Epic (deepest integration in the category — first “Pal” in Epic’s Partners and Pals programme, records in Haiku, writes into Hyperdrive), Oracle Health/Cerner, athenahealth, eClinicalWorks, MEDITECH, NextGen, Greenway Languages: 28+ claimed, though Abridge’s own science page says 14+ Availability: United States only, per its clinician terms

Abridge won Best in KLAS for Ambient Speech in both 2025 and 2026, scoring 94.7/100 in the 2026 report based on customer evaluations collected December 2024 to December 2025. It states partnerships with 300+ health systems including Duke (5,000 clinicians), Mayo, Emory and Inova.

Why it wins: It is the reference implementation for Epic-centred health systems, and its KLAS standing reflects sustained customer satisfaction rather than a single launch cycle. It is also the product used in the strongest published burnout study in the field.

Limitations: No published price and no self-serve path for small practices — you cannot evaluate cost without entering a sales process. It publishes no quantified accuracy metric, despite claiming to “set the industry standard for word error rate.” SOC 2 Type II status could not be verified from any Abridge-owned page. And it is US-only. Note also that Abridge’s product is the one named in both 2026 consent class actions (Abridge itself is not a defendant in either — see the compliance section).

Best for: Epic-based health systems with an enterprise procurement process.

2. Ambience Healthcare — best for coding and revenue-cycle ROI

Price: Not published; vendor describes fees as outcome-linked EHR: Epic (via Epic Toolbox, native in Hyperspace and Haiku since 15 August 2025), Oracle Cerner, athenahealth Compliance: SOC 2 Type I and Type II, GDPR conformant, HIPAA business associate

Ambience Healthcare raised a $243M Series C at a $1.25 billion valuation, co-led by Oak HC/FT and Andreessen Horowitz (Fierce Healthcare). It differentiates on coding rather than documentation alone, claiming 95% coding accuracy verified by AAPC, with real-time suggestions for HCC, E/M, ICD-10 and CPT codes.

Why it wins: It is the only vendor here with a KLAS-validated ROI study attached to a named customer — at St. Luke’s, a 41% reduction in chart closure time and 22% increase in patient face time, both measured via Epic User Action Log, with $13,000 generated per clinician per year through HCC and E/M coding. It was also one of the three products in the JAMA multi-site cohort.

Limitations: Enterprise only, no free tier, no self-serve. Its language handling translates any language into English — it does not produce notes in other languages. Given the coding-intensity debate above, treat the revenue half of its ROI case as contested rather than settled.

Best for: Health systems whose business case rests on coding accuracy and documentation integrity, not just clinician time.

3. Microsoft Dragon Copilot — best if you are already a Microsoft and Epic shop

Price: No published per-clinician price. The only Microsoft-published figure is $0.25 per AI-Assisted Session on pay-as-you-go, effective 4 May 2026 EHR: Epic (including Epic Rover for nurses), athenahealth, MEDITECH Languages: Captures conversation in 58 languages; the note is written in the primary language of the country Availability: Physicians in the US, Canada, UK, Ireland, France, Germany, Austria, Belgium, Netherlands and Switzerland; nurses US only

Dragon Copilot is the successor to Nuance DAX Copilot and reports 100,000+ clinicians using it. Intermountain Health reported a 27% reduction in time in notes per appointment across 2,500+ active clinician users.

Why it wins: The broadest international physician footprint of any vendor here, the widest conversation-language support, and nurse-specific deployment through Epic Rover — a genuinely differentiated capability, since almost nobody else ships a nursing product.

Limitations: Two serious ones. First, DAX Copilot was the arm that failed to reach statistical significance in the UCLA randomised trial, while Nabla succeeded — the only head-to-head randomised comparison against a control group that exists, and it did not go Microsoft’s way. Second, pricing is opaque and unstable: the Physician Practice SKU launched 1 February 2026 and was withdrawn from sale on 1 May 2026. Microsoft publishes no compliance attestations on any Dragon Copilot-specific page, and places patient-consent responsibility on the customer.

Best for: Multinational health systems already standardised on Microsoft and Epic, and organisations that need nursing documentation.

4. Nabla — best published evidence, strongest compliance stack

Price: Not published; nabla.com/pricing returns 404. Terms specify per-authorised-user, invoiced monthly in advance, quoted in EUR excluding VAT EHR: Epic, athenahealth, Oracle Health, NextGen, Greenway, Altera, plus Nabla Connect, an iFrame that works with any EHR Languages: 35+, including bilingual encounters Compliance: SOC 2 Type 2, ISO 27001, HIPAA with published BAA, GDPR, CCPA, Texas RAMP Level 2, EU AI Act, PIPEDA

Nabla reports 85,000+ clinicians across 130+ health organisations, including CVS Health, Children’s Hospital Los Angeles and Denver Health.

Why it wins: It is the only ambient scribe to have met its primary endpoint in a randomised controlled trial against a control group. In the UCLA trial, Nabla users cut writing time per note by 41 seconds against 18 seconds in control — a statistically significant 9.5% net reduction. It also publishes the most complete compliance documentation of any US-facing vendor here, including an actual BAA, a 14-day default note retention policy, and an explicit commitment not to train models on customer data.

Limitations: No published pricing. The ~$119/month figure in circulation comes entirely from third-party blogs with no Nabla corroboration. A free trial exists but its limits are not published.

Best for: Buyers who want the strongest published efficacy evidence and the most auditable compliance posture, and who can tolerate a sales-led purchase.

5. Suki — best multi-EHR coverage outside Epic

Price: Not published. Third-party sources report roughly $299/month for Suki Compose and $399/month for the full assistant; none is vendor-corroborated EHR: Epic, Oracle Health/Cerner, athenahealth, MEDITECH — dedicated integrations for each Languages: 80+ supported for input, but the note is always generated in English, with no language selector Compliance: SOC 2 Type II, HIPAA, 42 CFR Part 2, TX-RAMP, NIST AI RMF

Suki reports a 72% reduction in median documentation time per note and 3.3 hours saved per clinician per week from an American Academy of Family Physicians Innovation Lab pilot — though that was 132 physicians over 30 days, not a controlled study.

Why it wins: The most even coverage across the four major EHRs, including MEDITECH, where it is deployed at 12+ health systems. The 42 CFR Part 2 certification matters specifically for substance-use disorder treatment settings, where the confidentiality rules are stricter than HIPAA.

Limitations: English-only note output is a real constraint for multilingual practices. No published pricing, no free tier, no self-serve trial.

Best for: Multi-site organisations running MEDITECH or a mixed EHR estate.

6. Freed — best for solo clinicians and small practices

Price: $39/month Starter (40 notes), $79/month Core (unlimited), $119/month Premier (unlimited plus EHR push and coding), $104/month Premier billed annually EHR: Chrome-extension push to browser-based EHRs only, Premier tier — SimplePractice, Practice Fusion, Tebra, Elation, TherapyNotes, athenahealth, DrChrono and others. No Epic, no Oracle Health, no eClinicalWorks Languages: 14 named; note always translated to English Compliance: SOC 2 Type I and Type II, HIPAA, HITECH, BAAs signed with organisations

Freed reports 26,000+ clinicians and is the most transparently priced vendor in the category — the only one whose full rate card is published without a sales conversation.

Why it wins: Price transparency, a genuine 7-day trial with no credit card required, and a published pilot giving students, residents and trainees free access to the Core plan, alongside a 50% student discount on the monthly individual plan. For a solo clinician, this is the shortest path from interest to working product.

Limitations: Its “EHR integration” is a browser extension that pastes into web-based EHRs — categorically different from the FHIR and API integrations the enterprise vendors offer, and it does not reach Epic. It publishes no accuracy metric. Its “5 to 15 hours back per week” claim is self-reported by users, not measured.

Best for: Solo clinicians, therapists and small practices on a browser-based EHR.

7. Heidi — best free tier and best international coverage

Price: Plan structure published (Free, Evidence Plus, Clinician, Practice, Enterprise) but the dollar figures render client-side and could not be verified. Third-party sources report roughly $150/month for Clinician billed annually — treat as unverified EHR: Published as two classes rather than a named list — “Embedded” (widget inside the EHR) and “Connected” (requires Practice tier or above). Named EHR partners: data not available Compliance: Dedicated compliance programmes for HIPAA, UK, EU GDPR, Canada, and Australia/New Zealand APP

Heidi operates the only credible free-to-enterprise ladder in the category. Its free tier is permanent, not a trial: unlimited transcription and standard-template notes, unlimited evidence queries with citations, plus 10 paid credits per month for advanced templates and agentic features. EHR integration is excluded from free.

Why it wins: It is the only vendor here running compliance programmes across five jurisdictions, and it serves clinician types the US enterprise vendors ignore — psychologists, dietitians, allied health, podiatrists, dentists and veterinarians. New Zealand’s national AI advisory group approved Heidi and iMed for clinician use (ABC News, 14 August 2026).

Limitations: The February 2026 rebrand is genuinely confusing — Heidi Pro became Scribe Plus, the old Practice plan became Scribe Team, and new Scribe Plus and Scribe Team subscriptions are only available in the UK and EU. In-session Evidence is unavailable in the UK and EU. It publishes no accuracy or time-saving figure, and its SOC 2 Type II status could not be verified.

Best for: Clinicians outside the United States, allied health professionals, and anyone who wants to evaluate a scribe properly before paying.

8. Lyrebird Health — best for Australian general practice

Price: Free plan at $0 (50 transcribe and dictate actions per month, 10 documents, no EMR integration); Bp Free at $0 for Best Practice Premier subscribers (unlimited consult notes); Pro at A$160 per month billed annually; Enterprise custom EMR: Best Practice (Bp Premier), Genie and Gentu, Oracle Cerner, FHIR, plus Cubiko and BetterConsult Data residency: All data processed and stored onshore in Australia, aligned with the Australian Privacy Principles Regions: Australia, United Kingdom, United States

Lyrebird Health is the Melbourne-founded scribe that leads Australian general practice, and it is built around the Australian workflow rather than translated into it — it generates chronic condition management plans, mental health treatment plans, and Centrelink, WorkCover and NDIS PDFs, and integrates with PBS authority scripts. It raised US$12 million at a US$50 million valuation in June 2025, led by Five V Capital and UK-based Octopus Ventures (Herbert Smith Freehills Kramer), and reported powering documentation for 30,000 consultations per day in Australia (MobiHealthNews).

Why it wins: Two things no other vendor here offers. First, the Bp Free tier — Best Practice Premier subscribers get unlimited consult notes at no additional cost, and Best Practice is among the most widely used GP clinical systems in Australia, which makes this the cheapest credible route to ambient documentation for a large share of Australian GPs. Second, Lyrebird prompts the clinician for patient consent and timestamps it for medicolegal purposes. Given that both live US class actions turn on consent capture, and that AHPRA and the RACGP both require documented consent, this is the most underrated feature in the category and almost nobody else ships it.

Its Gold Coast Health case study, covering 100+ clinicians across 21 specialties, reports 88% of clinicians rating note quality improved, with notes scoring higher on the PDQI-9 structured quality instrument, and 84% reporting improved workflow efficiency. Its headline “80% reduction in post-consult documentation time” is qualified on Lyrebird’s own page as applying to “some clinicians” — read it as a best case, not an average.

Limitations: Its strength is also its constraint — the deep integrations are with Australian systems, so outside Australia it is a more ordinary product. Pricing is published in Australian dollars and is not directly comparable to the USD figures elsewhere on this page. Its compliance badges read “aligned with” the Australian Privacy Principles, HIPAA and GDPR rather than naming certifications, and SOC 2 Type II status could not be verified. Its claim to meet “requirements set by the Therapeutic Goods Administration” is a statement about compliance obligations, not about ARTG inclusion — as of August 2026 no AI scribe was included in the ARTG.

Best for: Australian GPs and specialists, especially any practice running Best Practice or Genie/Gentu.

9. Corti — best for EU deployments and developers

Price: Corti Assistant Pro $99 per clinician per month; Assistant Free $0 for 10 sessions per week; speech-to-text API $0.0065 per audio minute EHR: No individual EHR named; EHR connectivity is an Enterprise feature. Corti sells primarily to EHR vendors, including Dedalus and CompuGroup Medical Languages: 10+ for Assistant, 14+ for the API Compliance: The most extensive set here — EU MDR medical device registration (September 2025), UK device registration, SOC 2 Type 2, ISO 27001/27017/27018/42001/13485/14971, NHS DSPT and DTAC, EU AI Act, DORA, NIS2

Corti is a Copenhagen-based healthcare AI infrastructure company whose Corti Assistant is the clinician-facing scribe. It is the only vendor here publishing a quantified accuracy figure: 98.6% word accuracy, or 1.4% word error rate, on realtime English. That benchmark is Corti-run, though the company publishes an open-source evaluation tool for independent replication. Note one internal inconsistency: the same page cites both 4.6% and 4.1% WER against Dragon Medical One.

Why it wins: EU and US data hosting with no cross-border transfer, sovereign EU cloud, and formal medical device registration — which matters if your product does anything beyond transcription and summarisation. It also has a real free tier and the lowest-priced paid individual plan here.

Limitations: No named EHR integrations at the Assistant tier, so note transfer is copy-and-paste unless you buy Enterprise. The business is really an API sold to builders; the scribe is a secondary product.

Best for: European health organisations, and companies embedding clinical documentation into their own product.

10. DeepScribe — best for oncology

Price: Not published; no pricing page and no self-serve signup EHR: Epic, iKnowMed Generation 2 and OncoEMR only — the vendor describes the strategy as “depth over breadth”

DeepScribe has repositioned as oncology-first, with a Flatiron Health partnership covering 4,200+ providers. It reports that organisations using it see roughly 40% of all US cancer visits. It scored 98.8/100 in a KLAS Emerging Company Spotlight Report published 7 January 2025 — note this is a different, smaller-sample report than the Best in KLAS ranking Abridge won.

Limitations: Its most recent finalised SOC 2 Type II report covers calendar year 2024, with a bridge letter dated October 2025 — the oldest attestation of any vendor here. It publishes no language support information at all. Stale marketing pages for eClinicalWorks, AdvancedMD and NextGen still exist but those EHRs no longer appear on its integrations page; any review describing DeepScribe as a broad multi-specialty scribe is out of date.

Best for: Oncology practices and cancer networks on iKnowMed or OncoEMR.

11. Commure Scribe — cheapest paid plan with published pricing

Price: ScribePro $708/year billed annually, an effective $59/month, against a $1,068 list. Monthly billing lists at $89/month. Enterprise is custom EHR: athenahealth, AdvancedMD, eClinicalWorks, Elation, MEDITECH, Practice Fusion, SimplePractice, Tebra, WebPT and others — Epic and Oracle Health are not named on the Scribe integrations page. EHR sync is gated to the Enterprise tier

Commure offers a 7-day trial with no credit card, then requires an upgrade. Its enterprise arm claims deployment across large US health systems supporting 40 million appointments annually.

Limitations: EHR sync being Enterprise-only makes the $59/month tier a copy-and-paste product in practice. Language claims conflict between Commure’s own pages — 60+ on the Scribe site, 100+ on the enterprise page. No refunds, including on annual plans. A “99.4% transcription accuracy” figure appears in search results but could not be verified on any Commure page.

Best for: Small practices wanting the lowest published annual price, if you do not need EHR write-back.

12. Doximity Scribe — best free option for US clinicians

Price: $0 EHR: None. Manual copy and paste Availability: US physicians, nurse practitioners, physician assistants, CRNAs and students with a verified Doximity account

Doximity Scribe is genuinely free, funded by Doximity’s advertising network model rather than clinician subscriptions. Two constraints define it: a 140-minute per-session recording cap, and notes that permanently auto-delete after 29 days.

Doximity is unusually candid about its own limitations, publishing this in its help centre: Scribe runs on GPT-4o, which “has limited knowledge of clinical data and world events after October 2023… Think of Scribe as a helpful 4th-year med student — trust but verify.” No other vendor here publishes a caveat of that kind. Doximity holds a BAA with OpenAI, which provides the underlying transcription and summarisation.

Limitations: No EHR integration at all, no language list beyond a binary English/Non-English setting, and no published accuracy or time-saving claim. Whether Doximity signs a BAA with the individual practice is not stated.

Best for: US clinicians evaluating whether ambient documentation suits them at all, before spending anything.

13. Sunoh.ai — best for eClinicalWorks practices

Price: $149 per user per month, described as a limited-time rate against a $199 list. No contracts. A $1.25-per-visit option was published by eClinicalWorks in October 2024 but no longer appears on Sunoh’s own pricing page; whether it remains orderable is data not available EHR: eClinicalWorks (native), plus athenahealth, Cerner, Epic, ModMed, Practice Fusion and Tebra. A free Chrome extension pushes notes to any browser-based EHR Compliance: BAA included in the subscription; runs on Microsoft Azure. SOC 2 Type II is not claimed anywhere on the vendor site

Sunoh.ai is the eClinicalWorks-affiliated scribe, and its advantage is distribution: native integration into the eCW desktop EHR, eClinicalTouch and eClinicalMobile.

Limitations: Its headline claim that 60% of providers save one to four hours daily comes from an internal vendor survey. No published supported-language count and no published accuracy figure. The absence of any SOC 2 claim is notable in a field where most competitors publish one.

Best for: eClinicalWorks practices, where the native integration outweighs everything else.

Feature comparison: the full matrix

One footnote this table needs: “Epic” means three different things across these rows. Abridge, Ambience, Dragon Copilot, Nabla, Suki and DeepScribe have genuine bi-directional API or FHIR integrations. Freed and Sunoh offer a Chrome extension that pastes into browser-based EHRs. Doximity and Corti Assistant have nothing — you copy the note yourself. Do not compare these as though they are the same product feature.

Pricing comparison: what you’ll actually pay

Lyrebird’s figures are in Australian dollars, taken from its Australian pricing page, and are not directly comparable to the USD rows above.

The structural fact about this market: the six vendors that integrate properly with enterprise EHRs are the six that will not tell you what they cost. If you are a small practice, your real choice set is the bottom half of this table, and your EHR determines it more than your budget does.

Use-case specific recommendations

For a solo clinician or a practice under ten people

Freed ($39–$119/month), because it is the only vendor with a complete published rate card, a no-credit-card trial and a self-serve path. If your EHR is browser-based, Premier’s Chrome extension handles note transfer. Alternative: Doximity Scribe at $0 if you can live with copy-and-paste and 29-day note retention.

For an Epic-based health system

Abridge, on the strength of Best in KLAS 2026 at 94.7/100 and the deepest Epic integration in the category. Alternative: Ambience Healthcare if your business case depends on coding accuracy rather than clinician time.

For the strongest published efficacy evidence

Nabla — the only product to have met its primary endpoint against a control group in a randomised trial (NEJM AI, November 2025). Weigh this against the fact that it was a single-centre trial over two months.

For nursing documentation

Microsoft Dragon Copilot, the only vendor here shipping a nurse-specific product with Epic Rover integration. Important caveat: there is no published randomised trial of ambient AI for routine nursing documentation. Nursing charting is structurally different from an episodic encounter, so physician evidence does not transfer, and vendor testimonials about hours saved per shift are uncontrolled.

For Australian general practice

Lyrebird Health, which generates chronic condition management plans, mental health treatment plans and Centrelink, WorkCover and NDIS documents natively, integrates with Best Practice and Genie/Gentu, keeps all data onshore, and timestamps patient consent. If your practice runs Best Practice Premier, the Bp Free tier gives unlimited consult notes at no additional cost, which makes the decision close to automatic. Alternative: Heidi, if you also need New Zealand coverage or work in allied health.

For practices elsewhere outside the United States

Heidi for New Zealand, Canada and the UK — the only vendor running compliance programmes across all of them, and one of two approved by New Zealand’s national AI advisory group. Alternative: Corti for EU deployments needing medical device registration and sovereign data residency.

For oncology

DeepScribe, whose oncology-first repositioning gives it iKnowMed Generation 2 and OncoEMR integrations that nobody else has.

For multilingual and interpreter-mediated encounters

No clear winner, and this deserves caution rather than a recommendation. Suki supports 80+ input languages and Nabla 35+, but Suki, Freed and Ambience all output the note in English only. More importantly, ambient scribes have been shown to propagate interpreter errors directly into notes (JMIR Medical Informatics, July 2026). Review these notes more carefully, not less.

For evaluating whether to adopt at all

Heidi Free or Doximity Scribe, both permanently free. Run 20 encounters, read every note against what you remember of the visit, and count the omissions. Given that a measured 26.3% of key clinical elements were omitted or recorded incorrectly across five tested platforms, this is worth doing before you sign anything.

Compliance, consent and the 2026 litigation

This is the section that most vendor comparisons omit, and in 2026 it is the one most likely to cost you money.

”HIPAA compliant” does not answer the consent question

Every vendor here asserts HIPAA compliance, and most will sign a business associate agreement. Neither addresses whether you may lawfully record the encounter. A BAA governs how a vendor handles protected health information. It does not touch state wiretapping law, state medical confidentiality statutes, or patient consent to being recorded — and those are the theories on which the current litigation proceeds.

Two class actions are now live

• Sharp HealthCare (November 2025, San Diego Superior Court). A putative class action alleges Sharp used Abridge’s ambient tool to record encounters without patient consent under the California Invasion of Privacy Act. Its most distinctive allegation is a “false consent” theory: that the patient portal recorded the patient as having been advised of recording and having consented, when the complaint alleges no reliable workflow existed to ensure that actually happened (KPBS).
• Sutter Health and MemorialCare (Washington v. Sutter Health, filed 8 April 2026, N.D. Cal., case no. 4:26-cv-03012). Claims under CIPA, the Federal Wiretap Act and the California Confidentiality of Medical Information Act (TechTarget).

Abridge is not a defendant in either case. The exposure alleged runs to the health systems, not the vendor. As of 18 September 2026 there is no judgment, settlement or class certification ruling in either, and the allegations have not been tested on the merits.

Why this structure matters, per analysis from Alston & Bird (27 April 2026): the alleged violation completes at the moment of interception — when the conversation is recorded — not later at storage or disclosure. That defeats downstream data-handling defences, statutory damages scale per encounter, and exposure extends to everyone present at the visit, including spouses and carers.

Almost no vendor helps you here, which is the buying insight. Consent is the provider’s obligation, not the vendor’s, and most products simply start recording when you press the button. The exception in this comparison is Lyrebird Health, which prompts the clinician to obtain consent and timestamps it in the record for medicolegal purposes. Given that the Sharp complaint’s distinctive allegation is precisely that a consent attestation was recorded without a reliable workflow behind it, a product that captures when consent was actually sought is doing something materially useful. If you are deploying in an all-party consent state, or in Australia where AHPRA requires informed consent before an AI tool processes personal information, ask every shortlisted vendor how their product evidences consent — and treat “we are HIPAA compliant” as a non-answer.

All-party consent states

Recording consent is a state-law question in the US. States commonly classified as all-party consent for at least some categories of recording are California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania and Washington.

Do not treat that as a flat list. Some of these states distinguish in-person from telephonic recording, and several condition liability on a reasonable expectation of privacy (Reporters Committee for Freedom of the Press). Counts vary between eleven and twelve depending on how Connecticut is treated. The reason this doctrine bites for ambient scribes specifically is that an exam room is a strong candidate for reasonable expectation of privacy in every state.

State AI disclosure laws

• Texas SB 1188, effective 1 September 2025, makes review of AI-created records a legal duty for diagnostic use and requires disclosure of AI use, with civil penalties of $5,000 to $250,000 per violation.
• California AB 3030, effective 1 January 2025, requires GenAI disclaimers on patient communications concerning clinical information — but exempts communications reviewed by a licensed human provider, which is the exemption most clinician-reviewed scribe workflows rely on. A claim circulating in vendor material that AB 3030 requires a disclosure inside the body of the signed clinical note is not supported by the statute text.

Certification does not cover the product

The Joint Commission launched voluntary Responsible Use of AI in Healthcare certification on 1 June 2026, following joint guidance with the Coalition for Health AI in September 2025. It recognises organisations with sound AI governance and explicitly does not validate or certify individual AI products (The Joint Commission). No vendor can be “Joint Commission certified” as a scribe.

Regulatory status by country

United Kingdom — the position reversed in July 2026

The MHRA published Ambient voice technology-enabled products on 29 July 2026, and NHS England adopted it the same day. This reversed NHS England’s April 2025 position, which had treated generative summarisation as likely making a product a medical device.

Not medical devices under the current MHRA position: transcribing a conversation for clinician review; summarising it for review and edit; structuring content into problem or medication lists; suggesting clinical codes matched to terms explicitly stated in the conversation; and drafting a discharge summary or letter for clinician review.

Are medical devices: products offering suggested diagnoses or treatment options; products marketed with claims that they guide diagnosis or improve outcomes; and agents that finalise notes or order tests without clinician review. These are typically Class IIa where they allow direct diagnosis, requiring UKCA or CE certification.

Two things to know. NHS England’s own adoption guidance page had not been reconciled with the MHRA position as of 18 September 2026 and still contains the superseded summarisation language — cite the MHRA document for qualification questions. And the Health Services Safety Investigations Body launched an investigation into AVT use in hospitals on 6 August 2026, reporting in 2027.

Australia — compliance action is underway

The TGA’s position is that a scribe intended only to transcribe or translate, without analysis or interpretation, is not a medical device. On 3 August 2026 at the HIC2026 conference in Sydney, the TGA said its digital-scribe review had moved into compliance action, citing scope creep, AI functions influencing clinician decisions, and gaps in post-deployment monitoring (The Medical Republic). Software as a Medical Device is a TGA enforcement priority for 2026–27.

As of August 2026, no AI scribe was included in the ARTG — around a dozen products were available in Australia and none was approved by the regulator (ABC News, 14 August 2026).

AHPRA requires clinicians to check all scribe output for accuracy and to obtain informed consent before an AI tool processes a patient’s personal information. The RACGP requires documented consent, a patient’s right to decline without affecting care quality, and warns against over-reliance: a scribe “should produce a draft, not an autopilot chart.”

United States — no scribe-specific rule

The FDA has no guidance naming ambient scribes as a regulatory category. The operative framework is the Clinical Decision Support Software guidance, finalised January 2026: transcription and summarisation for clinician review generally sit outside the device definition, while outputs recommending a diagnosis or treatment engage the CDS criteria. The FDA published a discussion paper on generative AI-enabled medical devices on 18 August 2026, with comments due 19 October 2026 — the agency states explicitly that it is not draft or final guidance and proposes no policy change.

European Union

A pure ambient scribe is not listed in Annex III of the EU AI Act, so it is not high-risk on that route. Article 50 transparency obligations have applied since 2 August 2026. A scribe that is an MDR device requiring notified-body assessment becomes high-risk via Article 6(1), with those obligations deferred to 2 August 2028 under the Digital Omnibus that entered into force on 27 July 2026. GDPR applies independently and is not deferred.

Canada

Health Canada has no AI-scribe-specific guidance. The CMPA’s position is that physicians must obtain consent before any recording, document that discussion, and “retain ultimate responsibility for all chart entries, even those generated by AI scribes.”

What clinicians actually report

Automation bias is the recurring theme. Elizabeth Deveny, chief executive of the Consumers Health Forum of Australia, described the pattern: “There’s evidence that for the first few weeks of using AI, [people] check things pretty carefully and then at a point they stop checking because they assume it will be right.”

The error mode is specific and predictable. Dr Sean Stevens, chair of the RACGP’s Digital Health and Innovation committee, whom the ABC notes is also an adviser to one of the largest AI scribe vendors, described it: “The classic is, it will get the side of the body incorrect and say right when you’ve said left … so you need to check things pretty closely, especially things like drug doses.”

Documented harm exists. An Australian patient, Rebecca Green, consented to AI transcription of a urology appointment. The scribe fabricated a claim that she micro-dosed psychedelic mushrooms, which appeared in the post-operative letter sent to her GP as a possible cause of prior bleeding. She had never used them, and discovered it only on reading the letter after surgery. Her surgeon apologised in writing and corrected the correspondence. A Digital Rights Watch report documented further cases, including a scribe recording the wrong breast in a breast cancer diagnosis, and recording that a patient had epilepsy when they did not (ABC News, 14 August 2026).

Adoption is uneven even where the tool is available. In an emergency department deployment, the scribe was used in only 11.2% of eligible encounters, and 9.8% of physicians produced 70.5% of all scribe-assisted notes (Annals of Emergency Medicine, 2026). Physicians self-selected lower-acuity, non-interpreted encounters. Licence utilisation, not licence count, is what determines whether a deployment pays for itself.

Frequently asked questions

What is the best AI medical scribe in 2026?

For enterprise health systems, Abridge, which won Best in KLAS 2026 for Ambient Speech with an overall performance score of 94.7 out of 100 based on customer evaluations collected between December 2024 and December 2025. For solo clinicians and small practices, Freed, which publishes a complete rate card starting at $39 per month and offers a seven-day trial with no credit card. For clinicians outside the United States, Heidi, which runs compliance programmes across the US, UK, EU, Canada and Australia and New Zealand, and which New Zealand’s national AI advisory group approved for clinician use.

How much does an AI medical scribe cost?

Published prices range from $0 to $149 per clinician per month. Doximity Scribe is free for verified US clinicians, Heidi, Corti and Lyrebird all offer permanently free tiers, Commure ScribePro is $708 per year billed annually, Freed runs $39 to $119 per month, Corti Assistant Pro is $99 per month and Sunoh.ai is $149 per month. Lyrebird’s Pro plan is A$160 per month billed annually, in Australian dollars. The six vendors that integrate deeply with enterprise EHRs — Abridge, Ambience, Microsoft Dragon Copilot, Nabla, Suki and DeepScribe — publish no price at all and sell only through a sales process.

Do AI scribes actually save doctors time?

Yes, but less than commonly claimed. The largest study, covering 8,581 ambulatory clinicians across five academic health systems, measured 16.0 minutes of documentation time saved per eight scheduled patient hours. The only three-arm randomised controlled trial measured 41 seconds saved per note for Nabla, against 18 seconds in the control group. Clinicians who used a scribe in at least half their visits saved considerably more, 27.3 minutes, but only about 32% of adopters used it that frequently.

Do AI scribes reduce after-hours “pajama time”?

No. Three independent studies published between February and April 2026 looked for a reduction in EHR work outside scheduled hours and none found one. The JAMA cohort of 8,581 clinicians reported that after-hours EHR time did not change significantly, and the Duke randomised crossover trial found no evidence of a change in average pajama time, with both arms showing small non-significant increases. The most likely explanation is that saved minutes are reallocated to inbox work, chart review or patient face time rather than recovered as free time.

How accurate are AI medical scribes?

Less accurate than the category’s marketing implies. In controlled testing of five platforms against 14 simulated ambulatory encounters, a mean of 26.3% of key clinical elements were omitted or recorded incorrectly, with a mean of 3.0 errors per case carrying potential for moderate-to-severe harm. Errors of omission made up 76.3% of all errors, which makes them harder to catch than fabrications because nothing appears on the page to flag them. Ontario’s Auditor General separately found that 9 of 20 AI transcription programs produced hallucinations and 17 of 20 missed key mental health details.

Do I need patient consent to use an AI scribe?

In many jurisdictions, yes, and a HIPAA business associate agreement does not substitute for it. Twelve US states are commonly classified as all-party consent for at least some categories of recording, including California, Florida, Illinois, Pennsylvania and Washington, and two class actions filed in November 2025 and April 2026 allege that health systems recorded encounters without valid consent under state wiretapping and medical confidentiality law. In Australia, AHPRA requires informed consent before an AI tool processes a patient’s personal information, and the RACGP requires that consent be documented and that patients be able to decline without affecting their care.

Are AI medical scribes regulated as medical devices?

It depends on what the product does and where you are. The MHRA ruled on 29 July 2026 that transcription, summarisation, structuring and code suggestion are not medical devices in Great Britain, but that products offering suggested diagnoses or taking autonomous action are, typically as Class IIa. Australia’s TGA takes a similar transcribe-versus-interpret line and moved into compliance action in August 2026, with no AI scribe included in the ARTG at that point. The FDA has no scribe-specific rule and regulates through its Clinical Decision Support Software guidance. Corti is the only vendor covered here registered as a medical device in the EU.

Is Abridge or Microsoft Dragon Copilot better?

On the only randomised evidence available, Nabla outperformed both: in the three-arm UCLA trial, Microsoft’s DAX Copilot did not reach statistical significance against the control group while Nabla did. On customer satisfaction, Abridge is clearly ahead, having won Best in KLAS for Ambient Speech in both 2025 and 2026. A separate single-site crossover study of 18 emergency physicians found DAX scored better than Abridge on perceived work burden and note quality, which points in the opposite direction from the UCLA result — small, perception-based and not generalisable, but worth knowing. Dragon Copilot’s clearest advantages are international availability across ten countries and a nurse-specific product.

Can an AI scribe write the note without me reviewing it?

No, and every professional body that has addressed the question says the same thing. The MHRA states that clinician responsibility for reviewing and verifying AI-generated output before use in patient care is unchanged. Canada’s CMPA states that physicians retain ultimate responsibility for all chart entries, including those generated by AI scribes. Texas SB 1188 makes review of AI-created records a legal duty for diagnostic use, backed by penalties of $5,000 to $250,000 per violation. A product that finalises notes without clinician review is also, per the MHRA, a medical device requiring certification.

What is the best free AI medical scribe?

Heidi’s free tier is the most capable, offering unlimited transcription and unlimited standard-template notes permanently, plus 10 paid credits per month for advanced features, though EHR integration is excluded. Doximity Scribe is free for verified US physicians, nurse practitioners, physician assistants, CRNAs and students, capped at 140 minutes per session with notes auto-deleting after 29 days and no EHR integration. Corti Assistant Free allows 10 sessions per week, and Lyrebird’s free plan allows 50 actions per month. The most generous free offer in the category is Lyrebird’s Bp Free tier, which gives unlimited consult notes with Best Practice integration at no cost, but it requires an existing Best Practice Premier subscription. Commure and Freed advertise free tiers that are actually seven-day trials.

What is the best AI medical scribe in Australia?

Lyrebird Health, for most Australian GPs and specialists. It is built around Australian workflows rather than adapted to them, generating chronic condition management plans, mental health treatment plans and Centrelink, WorkCover and NDIS documents, integrating with Best Practice and Genie/Gentu, and storing all data onshore in Australia. Best Practice Premier subscribers get unlimited consult notes free through its Bp Free tier, and its Pro plan is A$160 per month billed annually. Heidi is the strongest alternative, particularly for allied health and for practices that also operate in New Zealand. Note that as of August 2026 no AI scribe was included in the Australian Register of Therapeutic Goods, and the TGA had moved into compliance action over products whose functions had drifted beyond transcription.

Is an AI medical scribe the same as a transcription tool?

No. General transcription tools such as Deepgram, AssemblyAI and Whisper convert speech to text and are judged on word error rate and cost per audio hour — those are covered on best AI for transcription. A clinical scribe additionally structures the output into a clinical note, writes it back into the electronic health record, often suggests billing codes, and operates under HIPAA, recording-consent law and in some jurisdictions medical device regulation. The buying criteria barely overlap.

Conclusion: how to choose in September 2026

The category works, but not in the way it is sold. Ambient scribes reliably improve how documentation feels — clinicians look at patients instead of screens, and the burnout data is real even if the mechanism is not the one on the ROI slide. What they do not reliably do is send anyone home earlier.

• Enterprise, Epic: Abridge (Best in KLAS 2026, 94.7/100).
• Enterprise, coding-led business case: Ambience Healthcare (KLAS-validated $13,000 per clinician per year at St. Luke’s).
• Strongest randomised evidence: Nabla — the only product to beat a control group on its primary endpoint.
• Solo and small practice: Freed ($39–$119/month, published pricing, no-card trial).
• Free: Heidi, or Doximity Scribe for US clinicians.
• Australian general practice: Lyrebird Health, and free if you run Best Practice Premier.
• Elsewhere outside the US: Heidi for New Zealand, Canada and the UK; Corti for the EU.
• Oncology: DeepScribe.
• eClinicalWorks: Sunoh.ai.

Three things to do before you sign. Model your real utilisation, not your licence count — the measured benefit roughly doubles for clinicians who use the tool in over half their visits, and only about a third do. Settle consent before deployment, not after — two class actions in ten months turn on exactly this, the alleged violation completes at the moment of recording, and a BAA does not help you. And run the tool against 20 of your own encounters and count the omissions — at a measured 26.3% of key clinical elements omitted or recorded incorrectly, dominated by things the note fails to say, that hour is the cheapest diligence available.

Prices, availability and regulatory positions on this page were verified against primary sources on 18 September 2026 and change frequently. Only six of the thirteen vendors covered publish a price; figures sourced from resellers or third-party reporting are labelled as such in the text. Study findings are reported with their design, because design determines how much weight a figure carries.