AI Medical Scribes for Healthcare Providers

Four steps, and each one is a place a deployment can fail.

1. Capture. A phone, a room microphone or a telehealth stream records the encounter after the patient has been told and has agreed. Audio quality does more to determine note quality than any model choice.
2. Transcription. Speech is converted to text, with speaker separation so the patient's words and the clinician's words are not blended into one voice.
3. Drafting. A language model turns the transcript into a structured note, usually in the specialty template already in use, and drops the small talk that does not belong in a chart.
4. Review and sign. The clinician reads the draft, corrects it and signs. Everything before this step is a suggestion, and every deployment that treats it as more than that will eventually put something wrong into a permanent record.

The interesting engineering is at the ends, not the middle. Getting clean audio in a busy room and getting the signed note back into the right encounter are the two problems that decide whether the thing is usable, and the second one is really an EHR integration problem.

Not typing speed. The problem is that documentation is deferred, and deferred work accumulates interest. A note written six hours after the visit takes longer to write, contains less detail, and is more likely to be the reason a clinician is still at a laptop at nine in the evening.

An ambient scribe moves the draft to the moment the conversation ends. The clinician still owns the note, but they are editing something rather than starting from an empty template. Practices that measure it well tend to see the benefit land in three places: time to close an encounter, the number of open charts at the end of the week, and how clinicians describe their day.

What it does not solve is understaffing. If the schedule is already full to the minute, returned documentation time will be absorbed by more visits rather than a shorter day, and that is a decision to make deliberately rather than discover afterwards. We work through that trade-off explicitly during an AI readiness audit.

There are three common integration depths, and the difference between them is felt every single visit.

• Native or embedded. The scribe runs inside the EHR's own mobile or desktop client. The note lands in the right encounter with no copying. Best experience, longest procurement.
• API based. The scribe writes back through a documented interface, commonly using FHIR resources or the vendor's app platform. Good experience, needs an integration project and a sponsor in your IT team.
• Copy and paste. The draft appears in a separate app and the clinician pastes it in. It works on day one and it is how most pilots start, but the friction is real and adoption decays.

Which of these is available to you depends on your system. The details for the largest one are on the Epic integration page, and the same questions apply to every other vendor: what is the write back path, who approves it, and how long does approval take.

An ambient scribe processes protected health information from the first second of the first recording, so the compliance work happens before the pilot, not after it.

• A signed business associate agreement with the vendor, and with any subprocessor that touches the audio.
• A clear answer on retention: how long raw audio is kept, whether it is kept at all after the note is drafted, and who can retrieve it.
• A clear answer on model training: whether your data is used to train shared models, and whether you can decline.
• Patient notification, and consent where your state requires it. Recording law is a state matter and it is not uniform.
• Access controls and audit logging that match the rest of your estate, not a weaker standard because the tool is new.

The HIPAA and AI compliance page works through each of these against the rule text, and the procurement checklist turns them into questions you can send a vendor before a demo.

Measure two things before the pilot starts, because you cannot reconstruct them afterwards: minutes of documentation per encounter, and the share of encounters closed the same day. Everything else is a proxy.

Then model the money honestly. The saving is clinician time, and clinician time only converts to revenue if it is used for something billable or is genuinely given back as shorter days. Both are legitimate outcomes. Confusing them is how business cases fall apart at renewal.

Our AI scribe ROI calculator takes your visit volume, documentation minutes and cost per clinician hour and shows the range rather than a single flattering number. Its assumptions are printed on the page so you can argue with them.

The market splits by buyer rather than by feature. Vendors selling to health systems lead with EHR depth, coding support and enterprise security review. Vendors selling to individual clinicians and small practices lead with self serve signup and published pricing, and several of them are genuinely good.

Buying the enterprise product for a four clinician practice is a common and expensive mistake, and so is buying the self serve product for a health system that will need a write back integration in month three. Match the vendor to the buying entity first, then compare features.

We keep an evaluated shortlist on the best AI medical scribes page, with the method and the gaps in our knowledge stated up front. We take no vendor commissions, which is why some pages say we could not verify something rather than filling the cell in.

Nine ambient documentation vendors appear in our registry, and every one of them signs a business associate agreement. That is the entry condition, not a differentiator. What separates them is who they are built to sell to, how deep they sit inside the EHR, and whether you can find out the price without a sales call. The table below is the shortlist we would start from, with the reasoning in the linked profiles. It is checked against vendor sites in September 2026 and it will age, so treat it as a starting point rather than a verdict.

Read the table by row, not by column. A health system with an Epic analyst and a security review process is choosing among the first four. An independent practice with no IT department is choosing among Freed, Heidi, Nabla and, if it runs eClinicalWorks, Sunoh. DeepScribe belongs on an oncology shortlist and rarely on a general one. The full grading, including what we could not verify for each vendor, is on the best AI medical scribes comparison.

Two things the table cannot show. First, "native" means the vendor and the EHR publish an integration, not that it is live at a site like yours; ask for a reference customer on your version. Second, the vendors that sell to health systems will not quote a price until they have qualified you, so if you need a number for a budget cycle this month, the published tier at the bottom of the table is the only one you can plan against today.

Between nothing and a figure the vendor will not put in writing until the third call. The spread is real and it maps almost exactly onto who the product is built for.

• Freed publishes individual plans from 39 to 119 US dollars per clinician per month: a capped Starter tier at 39, Core at 79 with unlimited notes, and Premier at 119, as listed on the vendor's site in August 2026. Group pricing is on request.
• Sunoh.ai lists 199 US dollars per user per month, with a limited time offer of 149, as listed on the vendor's site in August 2026.
• Heidi Health publishes a free tier and names paid Clinician, Practice and Enterprise tiers, with no dollar amounts on the public page in August 2026.
• Nabla publishes an individual clinician plan; organisation pricing is on request.
• Abridge, Dragon Copilot, Ambience, Suki and DeepScribe publish nothing. Pricing is on request, negotiated per contract, and in Dragon Copilot's case often bundled into an existing Microsoft or Dragon agreement, which can make the unit cost hard to see.

The licence is the smaller half of the cost at the enterprise end. Integration work, security review time, training and the internal owner who keeps the templates right are real money that never appears on a price list. At the self serve end the licence is most of the cost, which is why a practice can model the decision in an afternoon with the ROI calculator.

Where a vendor says pricing is on request, ask for four things in writing: the per clinician monthly rate at your volume, the minimum term, what happens to the rate at renewal, and whether integration and onboarding are included or invoiced separately. A vendor that will not put the renewal uplift in writing is telling you what it will be. We keep the published figures and the questions to ask side by side on the AI scribe pricing comparison, and each vendor has its own page under pricing.

One more thing to plan for: the monthly subscription model that makes the self serve products easy to start also makes them easy to stop, and that is a feature. If you are asked to sign a multi year commitment for a product you have not run in your own clinic for a quarter, the correct answer is no.

The ones with long conversational encounters and a note format the vendor has already trained on. That is a description of primary care, and it is why most vendor case studies are primary care. It does not mean the category stops there, but it does mean you should not assume a primary care result carries to your specialty.

• Primary care and family medicine is the reference case. Broad differential, a lot of history, a note that follows a predictable shape. If a vendor cannot demonstrate a good note here, it cannot demonstrate one anywhere. See the primary care page.
• Emergency medicine is harder: interrupted encounters, several patients open at once, noisy rooms, and a note that must support both clinical handover and a high acuity coding level. Ask specifically how the product handles an encounter that pauses and resumes. See emergency medicine.
• Fertility clinics have a repeating cycle of structured consultations, a great deal of numeric data, and patients who are highly engaged with their own record. Search demand for an AI scribe for IVF has grown sharply, and the vendors' specialty templates have not all kept up. See fertility clinics.
• Neurology is the test of a long history of present illness. A tool that summarises aggressively will lose the detail that a neurologist put ten minutes into eliciting. Pilot with your longest HPI, not your shortest. See neurology.
• Behavioral health sits at the intersection of the greatest documentation benefit and the strictest consent regime, including 42 CFR Part 2 for substance use disorder records. The recording conversation with the patient is different here, and some clinicians will decline outright. See behavioral health.

Procedural specialties with short visits and heavy use of templates tend to get less back, because there was less free text to begin with. Dermatology and ophthalmology often find the win is in the counselling visit rather than the procedure note. The rule is the same in every case: pilot in the specialty you intend to scale to, with the clinicians who see the hardest encounters, and read the notes against the audio for the first month.

The note has to end up in the right encounter, signed, with the right author and the right timestamp, and every EHR has a different route for getting it there. This is the part of the purchase that most often turns a two week decision into a six month project, so it is worth knowing your own system's paths before the first demo.

• Epic. The deepest integrations in the category. Abridge, Dragon Copilot and Ambience run embedded, including in Haiku on the clinician's phone. Suki, Nabla and DeepScribe build on Epic's published interfaces. Freed and Heidi have no documented write back and rely on the clinician copying the note across. Approval runs through your own Epic team.
• Oracle Health. Ambience, Suki and Nabla are recorded as native on the Millennium platform, with Dragon Copilot and Abridge on the API path. Launch inside the chart uses SMART on FHIR.
• athenahealth. Abridge, Suki and Dragon Copilot sit inside the Marketplace and athenahealth also ships its own Ambient Notes. This is the most integration friendly platform for a practice with no IT staff.
• eClinicalWorks. Sunoh.ai is eClinicalWorks' own scribe and is the path of least resistance. Ambience is native; Dragon Copilot, DeepScribe and Nabla build on the certified FHIR R4 API.
• MEDITECH. Suki and Dragon Copilot are native in Expanse. Fewer choices, but the choices that exist are mature.
• NextGen. Nabla is native and NextGen also markets its own Ambient Assist. DeepScribe builds on the API developer programme.
• Veradigm. The thinnest coverage. Heidi builds on the API; Dragon Copilot is recorded as a workaround. Expect a copy step.

Whatever the platform, ask the same four questions. What exactly is written back, into which note type, and who is recorded as the author. Who on our side approves the connection, and how long is their queue. What happens when the EHR is upgraded. And can we see it working at a site on our version, not on the vendor's demo tenant. Each pairing has its own page under integrations with what is documented and what is not.

HIPAA has no AI clause. The scribe vendor is a business associate because it creates, receives and maintains protected health information on your behalf, so the Privacy Rule, the Security Rule and the Breach Notification Rule apply exactly as they would to a transcription company. The difference in practice is the volume and sensitivity of what the vendor holds: audio of a clinical conversation is the most complete record of an encounter that exists, and it is held by a third party before the note is even drafted.

The federal floor is a signed BAA that says what the vendor may do with the data, flows the obligations down to the model and cloud providers behind it, states retention and destruction terms for audio and transcripts, and sets breach notification timing you can meet your own deadlines with. Model training on your patients' data is a permitted use only if the agreement says so. If the agreement is silent, do not assume the answer you would like. The HIPAA and AI compliance page works through each clause, and the HIPAA compliant AI scribe checklist turns it into questions to send before the demo.

State law adds two layers HIPAA does not. Recording consent is a state matter: a number of states require every party to a conversation to consent before it is recorded, so a notification script that is sufficient in a one party state is not sufficient across the border. And a growing set of states now regulate AI in healthcare directly. Texas's Responsible AI Governance Act sets disclosure duties for AI used in care. California has enacted several healthcare AI laws, including a requirement that patients be told when generative AI produces clinical communications without clinician review. Neither is triggered by a scribe whose output a clinician reviews and signs, but both are triggered by the moment someone decides the review step can be skipped.

The practical position for a multi state organisation is one policy set to the strictest state you operate in: tell every patient in plain words, record the notification, offer an opt out that does not degrade their care, and keep the clinician's signature as the only thing that makes the note real. The healthcare AI law checker shows which state rules apply to your footprint.

Five failures account for most of the disappointment we see discussed publicly and hear about in first calls.

• Review discipline decays. Week one, every note is read carefully. Week six, some are signed unread. This is the risk that matters most, and it is a management problem rather than a product one.
• The note is longer, not better. Ambient notes can be verbose. A longer note that nobody reads is a downgrade, and it is worth setting a house style early.
• Specialty fit is uneven. Performance in a long primary care conversation is not performance in a rapid procedural clinic. Pilot in the specialty you intend to scale to.
• The audio is bad. Shared rooms, background noise and a phone left in a coat pocket produce drafts that cost more to fix than to write.
• No owner. Tools without a named internal owner stop being configured, and then stop being used.

None of these are reasons to avoid the category. They are reasons to run the pilot as a project with a sponsor, which is the substance of an AI readiness audit.

If your documentation burden is modest but your front desk is drowning in phone calls and forms, an ambient scribe is the wrong first project. The load is at intake, and that is where the first agent should go.

If claim denials are the thing keeping the practice awake, start there instead. And if your EHR contract is up for renewal inside a year, sequence the scribe after the migration decision rather than integrating twice.

The right first project is the one where the pain is measurable, the workflow is owned by someone who will show up to the weekly call, and the failure mode is recoverable. For smaller organisations we set out how we sequence that on the private practice page.

Eight to twelve weeks, six to ten clinicians who volunteered, one specialty, and a baseline measured before anyone installs anything. Volunteers matter: a pilot staffed by conscripts measures resistance rather than the product.

Define the stopping rule in advance. Write down what result would make you buy, what result would make you walk away, and who decides. Pilots without a stopping rule become permanent, unbudgeted and unevaluated.

Review a sample of signed notes weekly against the audio for the first month, and keep a log of what the clinician had to change. That log is the most useful artefact the pilot produces. It is what tells you whether the tool is safe at scale, and it is the input we use to build an implementation plan once a pilot succeeds.

What an independent review adds at this point is a shortlist built from your EHR, your specialties and your state footprint rather than from a vendor's pipeline, a contract read against the BAA clauses above, and a pilot design with a stopping rule you set in advance. That is the substance of our vendor selection engagement. We take no vendor commissions, so the recommendation can be that you should wait. Book a call if you want that view before you sign.