AI Medical Scribes: How Ambient Documentation Works

Ambient clinical documentation is the fastest-adopted AI technology in medicine, and the gap between what it is sold as and what the published evidence shows is wider here than almost anywhere else in enterprise AI. This page is a technology review: how the pipeline works, what the independent studies found, where it fails, and the contractual rule that decides which tools a provider can lawfully use at all.

A boundary, stated once and meant literally. This is a technology overview for people evaluating or procuring ambient documentation software — practice managers, clinical informatics leads, physicians deciding whether to adopt. It is not clinical guidance, not legal advice, and not a compliance opinion. It does not tell any clinician how to practice. Nothing here substitutes for your own counsel, your compliance office, or your medical judgment.

The Challenge

• The note is a legal record, not a summary. It has a required structure — SOAP, or an HPI with review of systems — and it carries billing, liability and continuity-of-care weight.
• It gets written after hours. "Pajama time" — EHR work outside scheduled clinical hours — is the burden clinicians actually complain about, and the outcome most heavily marketed against.
• Typing during the visit competes with the patient. A clinician looking at a screen is not looking at a person.
• And the room is acoustically hostile. A clinician, a patient, sometimes a family member, sometimes an interpreter. Crosstalk, a rustling paper gown, a hallway door. Drug names that sound like other drug names. Accents. Code-switching mid-sentence.

That last constraint is the one this page is really about. Ambient documentation is not a writing problem. It is a hearing problem with a writing problem bolted onto the end of it, and the errors compound in that order.

Adoption has run well ahead of the evidence. The AMA's 2026 Physician Survey on Augmented Intelligence (n = 1,692, fielded January–February 2026) found over 80% of physicians now use AI professionally — double the 2023 share — with 28% using it for documentation of billing codes, charts or visit notes. In the same survey, privacy is the only dimension on which more physicians expect harm than help, and concern runs at 71% for non-institutional tools versus 42% for institutional ones. That gap is the shape of the problem: clinicians know the tool they reach for personally is not the tool their institution vetted.

How AI Solves It

The pipeline has four stages, and each fails differently. Keeping them separate is the whole of understanding this category.

1. Capture. An ambient microphone — a phone, a badge, a room device — records the encounter. This is where the recording becomes protected health information, and everything downstream inherits that status.

2. Speech recognition (ASR). Audio becomes text. This is the stage that mishears hydralazine as hydroxyzine, and where the most dangerous errors are born, because a fluent mistake does not look like a mistake. See voice recognition and audio processing.

3. Speaker diarization. Who said each thing — the clinician, the patient, the daughter who drove them to the appointment. This is a separate, weaker model bolted onto the ASR, and it is the stage that can attribute a symptom to the patient that the family member actually reported.

4. Note generation. A large language model reads the diarized transcript and drafts a structured note. It has no ears. It never touched the audio. It will confidently smooth over stage 2's errors rather than flag them.

Then the stage that is not a stage:

The clinician reviews, corrects and signs. The AI drafts; the human attests.

This is the central fact of the entire category, and it is not a disclaimer — it is the load-bearing architecture. The signature is a legal attestation that the note is accurate. No ambient documentation vendor assumes that responsibility, and none claims to. It is why these tools sit outside FDA device regulation, and why the review step cannot be optimized away no matter how good the models get. Every honest evaluation of this technology is, at bottom, an evaluation of how much work the review step is doing.

What the technology genuinely delivers is narrower and more interesting than "it writes your notes": it moves the work from composition to verification. Reading and correcting a draft is a different cognitive task from writing from scratch, and clinicians overwhelmingly prefer it — which is exactly what the burnout data shows, and exactly what the time data does not.

Recommended Tools & Models

Two things first. None of the consumer AI tools in our catalog is a medical scribe, and none may lawfully be used as one on a consumer account — that is the next section, and it is the gate everything else passes through. And the underlying models are almost entirely undisclosed: vendors do not publish which ASR or which LLM they run, so a model-by-model comparison is not available to anyone outside the vendors.

Pricing verified July 2026. Most of this market does not publish prices at all, which is itself a procurement fact.

The structural fact that outranks all of them: Epic now ships its own. On 4 February 2026 Epic released AI Charting — native ambient note and order drafting inside the EHR. Epic's CMO Jackie Gerhart described the intent plainly: "It's a scribe, but that's passive. We really want this to be active." It is Epic-native but Microsoft-powered, running Microsoft's Dragon ambient AI underneath. Epic continues to integrate Abridge, Nabla, Suki and Ambience while now competing with them. If you are procuring in 2026, that is the dynamic you are procuring into.

The rule that comes before any tool: HIPAA and the BAA

A recording of a patient encounter is protected health information. PHI under 45 CFR 160.103 is individually identifiable health information transmitted or maintained "in any other form or medium" — the definition is medium-agnostic, and audio is squarely inside it.

A vendor that "creates, receives, maintains, or transmits" PHI on a covered entity's behalf is a business associate. Under 45 CFR 164.502(e), a covered entity may only let a business associate touch PHI if it obtains satisfactory assurances that the PHI will be safeguarded — assurances that "must be documented through a written contract." That contract is the Business Associate Agreement (BAA).

The gate is therefore contractual, not technical. A consumer chatbot is not forbidden because it is insecure. It is forbidden because no BAA covers it, which makes the disclosure an unauthorized disclosure of PHI by the provider.

And the trap is not the free tier. It is the paid one. Every major vendor operates a closed allowlist: a BAA covers named services only, and everything else is excluded by omission rather than by any warning label. As of July 2026:

Read the middle column again. ChatGPT Business and Claude Team are paid business tiers with no BAA. A clinician who upgrades has bought speed and higher limits — not coverage. That is the mistake this category actually produces, and a pricing page will never warn you about it. (One wrinkle for procurement: Claude accessed through Amazon Bedrock or Google Cloud is covered by the cloud provider's BAA, not Anthropic's, because there the cloud provider is the data processor.)

Recording consent is a separate question, and HIPAA does not answer it

HIPAA permits uses of PHI for treatment, payment and healthcare operations without patient authorization (45 CFR 164.506(a)), and disclosure to a business associate under a BAA needs no authorization either. So HIPAA is not the source of any duty to ask permission before recording.

That duty comes from state recording law, independently. The federal floor is one-party consent (18 U.S.C. § 2511(2)(d)), but stricter state law governs where it exists — and most "all-party consent state" lists circulating online are built from telephone statutes that misclassify states in both directions for an in-person encounter, which is what an exam room is. The details are genuinely contested in several states and are a question for your counsel, not a web page.

The engineering conclusion, though, is simple and robust: an explicit, audible, documented notice-and-consent step before recording starts satisfies the strictest in-person regimes and moots most of the contested analysis. Build it into the workflow and the jurisdictional question largely stops mattering.

Where the FDA line actually is — and it is not where most people think

Ambient documentation tools are generally not FDA-regulated medical devices, and the reason is more specific than "there's a human in the loop."

Section 3060 of the 21st Century Cures Act added § 520(o) to the Food, Drug & Cosmetic Act, excluding several software categories from the device definition — including software for "administrative support of a health care facility" and software intended "to serve as electronic patient records." A tool that transcribes an encounter and drafts a note is recordkeeping and administrative software. It never reaches the Clinical Decision Support analysis at all.

The evidence is concrete. FDA's AI-Enabled Medical Device List, updated March 2026 and current through 30 December 2025, covering 1,451 authorized devices, contains no matches for "scribe," "speech," "dictation," or "transcription," and no entries for Abridge, Nabla, Suki, Ambience, Dragon Copilot, Freed or Heidi. No ambient documentation product has ever been FDA-cleared — because none has needed to be. (FDA's Clinical Decision Support guidance was revised in January 2026, superseding the 2022 version. The words "scribe," "transcription" and "documentation" appear nowhere in it. It is simply not about these products.)

Two consequences a procurement lead should internalize:

Billing and ICD-10 coding suggestions do not push a tool toward device status. Coding sits squarely inside the administrative-support exclusion, which names "claims or billing information" explicitly. A revenue-cycle feature is an accuracy and compliance question, not an FDA question.

Specific clinical directives are where the line lives. Non-device status depends on the software not providing a specific diagnostic or treatment directive, not replacing the clinician's judgment, and letting the clinician independently review the basis for its output. Which is why Epic's stated direction for AI Charting — expanding from drafting the note to drafting the orders, and beyond that toward diagnoses — is the thing to watch in this entire category. A tool that drafts a note is administrative. A tool that decides what to order is walking toward a different regulatory regime, and the vendors know it.

Step-by-Step Implementation

An evaluation and deployment runbook for the person doing the procuring — not a clinical protocol.

1. Get the BAA before you get the demo

Ask for the executed BAA and which named services it covers. "We're HIPAA compliant" is a marketing sentence; the BAA is a document with a service list in it. If a vendor cannot produce one, the evaluation is over — not because the product is bad, but because you cannot lawfully pilot it with real PHI.

Then ask the three questions the BAA does not answer:

• Is the source audio retained, and for how long? This one has teeth. During the 2024 reporting on Whisper hallucinations (below), Nabla's CTO confirmed the tool erases the original audio for data-safety reasons — which means an AI transcript cannot be checked against the recording it came from. That is a defensible privacy design and a real loss of auditability. Make the trade-off deliberately rather than discover it later.
• Is our data used to train the vendor's models? Get the answer in the contract, not the sales deck.
• Who are the subcontractors? The vendor's ASR may be someone else's API. Business-associate obligations flow down, but you should know who is in the chain.

2. Build the consent step into the workflow, not into a waiting-room poster

An audible, scripted notice before recording begins, the patient's response documented, and a genuinely easy way to decline. In the UCLA randomized trial, 6.4–7.2% of patients declined to be recorded. That is a real number and it needs a real path — a workflow that makes declining awkward is one that will quietly stop offering the choice.

3. Pilot with a control group, or do not bother

The single highest-value instruction on this page, and the one most pilots skip.

The same tool, DAX Copilot, produced a 20.4% reduction in time-in-notes in a 46-clinician uncontrolled pilot and a statistically insignificant 1.7% in a 238-physician randomized trial. Both are peer-reviewed. Both are honest. Only one is an estimate of the tool's effect.

If your pilot has no control arm, it will measure your enthusiasts, your novelty effect and your seasonal EHR load. Randomize, use a matched comparison group, or accept that you have run a satisfaction survey rather than an evaluation.

4. Measure adoption before you measure benefit

Adoption is the buried story of this literature, and it will dominate your ROI:

• UCLA RCT: the tools were used in 33.5% and 29.5% of eligible visits, and roughly 15% of physicians assigned a tool never used it once.
• JAMA multisite (8,581 clinicians): only 32% of adopters used it in half or more of their visits.
• Emergency department: 11.2% of encounters, with 38% of attendings ever using it — and use clustered in the easy cases (low-acuity, telemedicine, no interpreter).
• Stanford: 55% utilization, with a subset of users who spent more time per note than at baseline.

A license nobody opens has an ROI of exactly negative one license. Track per-clinician utilization weekly from day one, and treat a non-adopting clinician as information — about specialty fit, room acoustics, or patient population — rather than as a training failure.

5. Audit notes against the encounter, with a physician reviewer

Sample and grade. Not for "quality" in the abstract, but for omissions, fabrications and medication errors specifically — the three categories the published safety research says actually occur, in that order of frequency.

The benchmark to beat is uncomfortable: the independent NIH/AHRQ-funded evaluation below found 26.3% of key clinical elements omitted or erroneous across five platforms. If your audit finds a rate dramatically better than that, check your audit before you celebrate.

6. Watch the review step, because it is the whole product

Track how much of the draft clinicians actually change. At UC Davis the median editing rate was 9.0% of AI-generated words — and 14.9% of notes were signed with no edits at all. A signed, unedited AI note is the failure mode this technology must be managed against. Not because unedited is always wrong, but because "I didn't change anything" and "I read it carefully" are indistinguishable in the audit log, and only one of them is safe.

Real-World Examples

These are documented findings from published evaluations. No patient scenarios have been invented for this page.

Example 1: The same tool, two study designs, two different worlds

The uncontrolled pilot. Duggan et al., JAMA Network Open, 2025 — a single-group before-and-after study of 46 clinicians across 17 specialties running DAX Copilot in Epic for seven weeks.

• Time in notes per appointment: 10.3 → 8.2 minutes (−20.4%, P < .001)
• After-hours work: 50.6 → 35.4 minutes/day (−30.0%, P = .02)
• Same-day chart closure: 66.2% → 72.4%

Those are the numbers that got repeated. (They are also, incidentally, the numbers most often misattributed to Kaiser Permanente's much larger deployment in secondary write-ups. They are not Kaiser's.)

The randomized trial. Lukac et al., NEJM AI, 2025 — a three-arm pragmatic RCT, 238 outpatient physicians across 14 specialties, DAX vs Nabla vs usual care.

• DAX Copilot: −1.7% time in notes (95% CI −9.4 to +5.9, P = 0.66). Null.
• Nabla: −9.5% (P = 0.02) — an absolute saving of about 41 seconds per note
• After-hours EHR time: no significant difference for any group
• Burnout, however, improved for both tools (Mini-Z 2.0 +2.83 and +2.69)

What changed: not the tool. The presence of a control group. The pilot measured 46 volunteers who chose to be there during a seven-week novelty window; the RCT measured what the tool does. This is the most instructive comparison in the ambient-scribe literature, and the fact that both studies are legitimate is the point: an uncontrolled pilot is not evidence of effect, no matter how good the journal.

Example 2: What the notes actually get wrong

The independent safety evaluation. Anderson et al., Mayo Clinic Proceedings: Digital Health, 2025 — NIH- and AHRQ-funded, no vendor money, no competing interests declared. Five ambient scribe platforms (four commercial, one free) were run against 14 simulated ambulatory encounters and graded by physicians.

• Mean 26.3% of key clinical elements omitted or erroneously captured (95% CI 17.0–31.0%)
• Only 35.8% ± 11.3% of elements were captured correctly across all five platforms
• A mean of 3.0 errors per case carried potential for moderate-to-severe harm (range 0–21). One error was graded as carrying a death risk.
• Transcripts averaged 13.9 errors per case, and 19.5% of transcript errors propagated into the final note — the language model filters some ASR noise, and nowhere near all of it
• Omissions were 76.3% of all errors. The note leaves things out far more often than it makes things up
• Qualitatively: medication-related errors were the most common, and one platform produced substitution errors "in which a clinical term or phrase was replaced with an unrelated but contextually plausible alternative"

That last phrase describes the most dangerous failure mode in the category. The error is fluent. It does not look like an error. It looks like a note.

A larger data point. Taylor et al., JMIR Medical Informatics, 2026 — a UC Davis pilot of 7,545 AI-generated notes, 356 formally quality-assessed: hallucinations in 11.5%, accidental omissions in 18%, and 2.5% of notes containing errors posing a serious or imminent risk of harm. Over 80% were free of significant errors — which is either reassuring or alarming, depending entirely on how many notes you sign per week.

And the vendor-funded study agrees. Palm et al., Frontiers in Artificial Intelligence, 2025 reports a 31% note-level hallucination rate for the ambient tool versus 20% for physician-written comparison notes (P = 0.01). All authors disclose commercial relationships with Suki AI, which funded the publication fees. A vendor-affiliated paper reporting a 31% hallucination rate in its own category is not a number anyone had an incentive to inflate.

Example 3: The speech layer hallucinates too — and it reached hospitals

Stage 2 has its own fabrication problem, and it is the best-documented one in the field.

Koenecke et al., "Careless Whisper: Speech-to-Text Hallucination Harms," ACM FAccT 2024. The researchers ran 13,140 audio segments from AphasiaBank through OpenAI's Whisper API:

• 1.4% of transcriptions contained hallucinations — text with no counterpart anywhere in the audio
• 38% of hallucinated transcriptions contained at least one identified harm: perpetuation of violence (19%), inaccurate associations such as invented names, relationships or health status (13%), and false authority (8%)
• Hallucinations were significantly more likely for speakers with aphasia (1.7% vs 1.2%, P = 0.019) and correlated with non-vocal duration — silences and disfluencies. Aphasia speakers' audio was 41% non-vocal versus 15% for controls.
• Five competing ASR systems on the same segments produced zero comparable hallucinations. This was Whisper-specific.
• In fairness: on a December 2023 re-run, only 12 of the 187 hallucination-triggering segments still hallucinated, which the authors attribute to Whisper updates released that November. The problem was substantially mitigated. It was not proven absent.

Why this reached medicine. An Associated Press investigation (Burke and Schellmann, 26 October 2024) reported that Nabla's ambient tool was built on Whisper and was in use by over 30,000 clinicians across 40 health systems, having transcribed an estimated 7 million medical visits — and that Nabla deletes the source audio, so transcripts could not be checked against what was actually said. Fabrications documented by researchers included an invented medication, "hyperactivated antibiotics," and unprompted racial commentary added to speech where race was never mentioned. OpenAI's own documentation warns Whisper should not be used in "high-risk domains."

Nabla's response (28 October 2024) was that its model is based on Whisper but "contains many improvements specifically developed to suppress hallucinations," trained on 7,000 hours of annotated medical audio, with raw ASR output never written directly to the record. That is a substantive answer. It is also their last public statement about their speech stack — so what Nabla runs in 2026 is not something we can tell you, and neither can anyone else outside the company.

The transferable lesson is not "avoid Nabla." It is that word error rate does not measure hallucination. A 2025 ACL paper testing 20+ ASR models found the two are effectively decoupled: a low WER can conceal a high fabrication rate. You cannot certify a medical scribe by its accuracy score, because the accuracy score is not measuring the thing that will hurt you. This is hallucination at the acoustic layer, beneath the language model, where almost nobody is looking for it.

Industry-Specific Applications

Primary care and outpatient specialties. The core case, and what every study above measures. Structured encounters, two speakers, a note format the model has seen a million times. If it works anywhere it works here — and here the effect is roughly one minute per note.

Emergency medicine. Adoption collapsed to 11.2% of encounters in a published academic ED evaluation, clustering in low-acuity, telemedicine and non-interpreter cases. Read that as the technology telling you where it is comfortable.

Encounters with a family member present. Diarization's hardest case, and not hypothetical. A benchmark of state-of-the-art end-to-end diarization found error rising from 8.86% with one speaker to 14.02% with four and 18.96% with eight. The failure is specific and clinically meaningful: a symptom reported by the daughter gets attributed to the patient.

Patients with accented speech, or who code-switch. The most under-discussed equity problem in this category. Koenecke et al. (PNAS, 2020) found commercial ASR systems produced an aggregate word error rate of 0.35 for Black speakers versus 0.19 for white speakers — and on identical phrases spoken by both groups the gap persisted at roughly 2×, confirming the cause is acoustic rather than lexical. More than 20% of snippets from Black speakers had an error rate above 0.5; fewer than 2% of white speakers' did.

It worsens when the vocabulary is clinical. A 2025 evaluation across three regions found the medical-concept error rate was 4.5% for US speakers and 23.8% for Nigerian speakers — note the asymmetry: for US speech, medical terms were transcribed better than average speech; for Nigerian speech, worse. An African-accented clinical dialogue benchmark found diarization error nearly tripled on the medical subset (34.64%) versus general conversation (12.28%). And a 2024 Interspeech benchmark found Whisper-large recovered only 42% of spoken medication names in accented clinical speech while its headline WER looked respectable.

The practical consequence: an ambient scribe's accuracy is not a single number. It is a distribution across your patient population, and it is worst for the patients your health system probably already serves least well. Ask a vendor for accuracy stratified by patient demographics. Most cannot provide it. That answer is itself informative.

Non-English and interpreted visits. Treat as unvalidated unless the vendor shows data for that specific language pair. A third voice, a second language and consecutive interpretation stack every failure mode on this page at once.

Best Practices

• The clinician signs, therefore the clinician owns it. Every workflow decision follows from this. Any process that makes signing easier than reading is built backwards.
• Get the BAA and its service list before the pilot. Not the "HIPAA compliant" badge — the executed contract, with the covered services named in it.
• Paid ≠ covered. ChatGPT Business and Claude Team are paid tiers with no BAA. The single most common misunderstanding in the category.
• Never paste PHI into a consumer AI account. Not for a note, not for a summary, not "just to see if it works." See privacy.
• Randomize the pilot, or call it a satisfaction survey. Uncontrolled before-and-after pilots reliably overstate effect by an order of magnitude.
• Instrument adoption from day one. A tool used in 30% of visits cannot deliver an ROI modeled at 100%.
• Audit for omissions first. They are 76.3% of errors, and nothing in the note tells you what is not in the note.
• Ask what happens to the source audio. Deletion protects privacy and destroys auditability. Choose knowingly.
• Ask for accuracy stratified by accent and language. The aggregate number hides the patients most at risk.

Common Pitfalls

Believing the pilot. The most expensive mistake available. The identical tool measured 20.4% uncontrolled and 1.7%, non-significant, randomized. Almost every number in a vendor deck comes from a study of the first kind.

Believing after-hours time will fall. "Take back your evenings" is the category's central marketing promise and the claim the evidence most consistently fails to support. The JAMA 2026 multisite study (8,581 clinicians): no significant change in after-hours EHR time. The UCLA RCT: no significant difference. Sutter Health: P = .14, and directionally worse. The University of Wisconsin RCT found a reduction that became non-significant once the top 3% of outliers were removed. And Haberle et al. (JAMIA, 2024), evaluating an ambient scribe at Intermountain Health, reported after-hours EHR time going the wrong way — a significant increase relative to baseline. Not every deployment gets a null. Some get worse.

Confusing "feels better" with "is faster." Genuinely different findings, and both are real. The burnout and cognitive-load improvements are large, consistent and replicated. The time savings are small. Selling the second to justify the first is how a successful deployment becomes a broken promise.

Trusting word error rate. WER and hallucination rate are decoupled. A vendor's accuracy figure is not a safety figure, and no vendor publishes a hallucination rate.

Assuming the LLM will catch the ASR's mistakes. It will not — 19.5% of transcript errors propagated straight into the final note in the Mayo evaluation. Stage 4 has no ears. It cannot know that hydralazine was said and hydroxyzine was heard, and it will write a fluent, coherent note around the wrong drug. Medication errors were the most common error type in that study, and drug names are where a homophone stops being a curiosity.

Missing the omissions. Fabrications are what people fear; omissions are what actually happen — 76.3% of errors, 18% of notes at UC Davis. A missing finding leaves no trace. There is nothing to notice.

Signing without reading. 14.9% of AI-generated notes at UC Davis went out entirely unedited. With 2.5% of notes carrying a serious-harm risk, unedited signing is where a technology problem becomes a patient-safety problem — and it is a workflow failure, not a model failure.

Note bloat. Multiple studies found AI notes got longer: +258.8 characters at Sutter, +20.6% in the Duggan pilot. Longer notes take longer to read, which quietly eats the time the tool saved and makes the record harder for the next clinician to use.

Believing the FDA vetted it. Nobody did. These tools are non-devices under the administrative-support exclusion, which means no regulator has reviewed the accuracy of any of them. Your audit is the only audit.

Measuring Success

Take your benchmarks from the independent literature, not the sales deck.

Burnout is where the technology genuinely delivers, and where you should set your expectation of value — the UCLA RCT found significant burnout improvement for both tools even where time savings were null. Use a validated instrument (Mini-Z, or the Professional Fulfillment Index) and measure before you deploy, because you cannot reconstruct a baseline afterwards.

Two caveats, though. The largest burnout numbers come from studies with response rates of 11–30% — serious non-response bias in a self-selected population. And the review literature is genuinely split. A 2025 systematic review (Sasseville et al.) concluded ambient scribes have "limited impact on reducing burnout" — though note it performed no meta-analysis, and the non-significant p-value often quoted alongside that phrase comes from a single included pilot with about nine participants, not a pooled estimate. A Yale rapid review called the evidence base "sparse": six real-world studies out of 1,450 screened. Pulling the other way, a 2025 meta-analysis (Zhao et al.) that did pool its studies reports a moderate reduction in documentation burden (SMD −0.71). The burnout effect is the most replicated finding in this field and the reviewers disagree about it. Both are true, and anyone quoting you one side has not read the other.

Edit rate is the metric nobody tracks and everybody needs: what percentage of the draft do clinicians change, and how many notes are signed untouched? The second number is a safety metric wearing a productivity metric's clothes.

Mass General Brigham's own analysis, worth keeping in view against a per-seat license, put the revenue effect at $167/month per adopting clinician — "statistically significant but nominal."

Cost Analysis

License. Mostly unquotable, which is itself the finding. Of the major enterprise vendors — Abridge, Nabla, Suki, Ambience, Commure — none publishes a price. Microsoft publishes a partial rate ($0.25 per AI-assisted session in consumption units; the per-user subscription price is not public). Small-practice tools do publish: Freed at $59–$119/month, Heidi's Clinician tier listed at $150/user/month. Infrastructure is cheapest and least finished: AWS HealthScribe at $0.10/minute of audio, Corti's STT at $0.0065/minute.

Integration. The EHR integration is the real cost and the real timeline. Epic's "Pal" and "Toolbox" designations exist because this is hard. Budget for it in months, not weeks.

The review time you did not eliminate. If a clinician spends four minutes reviewing a draft instead of five minutes writing a note, the license has to be worth one minute per encounter. Run that arithmetic against a real quote using the independent benchmark (−0.5 to −1 minute per note), not the vendor's.

What it buys that is real. Attention during the visit, and a lower cognitive load — a clinician not typing is a clinician listening. In the Sutter study, clinicians reporting "undivided attention" to the patient went from 57.9% to 93.0%. Two disclosures that belong next to that number: it is a self-reported survey item, not measured attention, and Sutter Health is a strategic investor in Abridge, the tool it evaluated. Take it as a large effect that points the right way, not as an independent measurement. It is still a legitimate reason to buy this technology. It is simply not a time-savings reason, and confusing the two is how deployments fail politically even when they succeed clinically.

What it does not buy. The signature. The responsibility. The audit. Those stay with the clinician, permanently, by design.

Related Use Cases

• AI for Podcast Show Notes — the same ASR-and-diarization pipeline in a domain where the failure modes are embarrassing rather than dangerous. The clearest way to build intuition for stages 2 and 3.
• AI for Lesson Plans — the same contractual gate in a different regulatory regime: FERPA is to student records what HIPAA is to PHI, and the consumer-account trap is identical.

Related Tools

None of these is a medical scribe. They are here because they are what clinicians reach for when no approved tool exists — and, on a consumer account, they are exactly what must not be used with PHI.

• ChatGPT — Enterprise and the API can be BAA-covered; Free, Plus, Business and Team cannot
• Claude — the API under a HIPAA-ready organization and Claude Enterprise can be covered; Pro, Max and Team cannot
• Google Gemini — coverage runs through Google Cloud's enterprise platform, not the consumer app
• Microsoft Copilot — Microsoft 365 Copilot carries a default-on BAA; preview features are contractually excluded
• Descript and ElevenLabs — transcription and diarization tooling; useful for understanding the pipeline, not for PHI
• NotebookLM — grounds answers in uploaded documents and cites them; the enterprise edition is BAA-eligible, the consumer one is not

Related Models

The models inside commercial ambient scribes are not disclosed. These are the general-purpose models that make stage 4 possible, and that vendors are widely believed to build on.

• Claude Sonnet 5 — long, messy transcripts; literal instruction-following
• GPT-5.5 — strong general summarization and structured output
• Gemini 3.5 — large context for long encounters

Further reading here: AI healthcare for the wider clinical landscape, hallucinations for why fluent output is the dangerous kind, accountability and trust for the framework that makes the signature the whole point, and VibeVoice-ASR for where long-form speech recognition and diarization are heading.