The best ambient AI scribe for most health systems is the one your clinicians still use in month six — on Epic, that means running Epic AI Charting and Abridge head-to-head before you sign. The largest multi-site study to date found ambient scribes save about 16 minutes of documentation time per eight-hour clinical day, a 10% cut, with no change in after-hours charting. That is real, but it is not the hour a vendor demo implies. The purchase you are making is mostly about adoption, note quality, and who holds the liability when the draft is wrong.
My recommendation, for a 1,000-clinician ambulatory group on Epic seeing roughly 16 visits a day per clinician:
| Product | Buy it if | Do NOT buy it if | Epic depth | Published price (checked 2026-09-26) |
|---|---|---|---|---|
| Abridge | You are on Epic and want the most mature embedded third-party workflow today | You need a self-serve solo contract, or your consent process is not airtight | Embedded in Haiku and Hyperdrive | Contact sales |
| Epic AI Charting | You are on Epic and can wait for the feature set to settle | You are not on Epic, or you need proof before you pilot | Native | Not disclosed |
| Ambience Healthcare | Coding and CDI revenue is the real business case | You only want notes | Integrated (per health-system deployments) | Contact sales |
| Microsoft Dragon Copilot | You already have an enterprise Microsoft deal and want dictation plus ambient in one contract | Your case rests on measured time savings | Integrated; its ambient tech also underpins Epic AI Charting | Contact sales |
| Freed | You are a small, non-Epic practice buying per clinician | You are a health system needing embedded EHR workflow | "EHR push" on Premier and Groups | $39 / $79 / $119 per clinician per month |
The loser for an Epic shop is standalone Microsoft Dragon Copilot — not because the product is bad, but because the published randomized trial that tested it showed no significant time effect, and Epic is now building its native scribe on Microsoft's ambient technology anyway. Paying separately for the same engine needs a reason other than the notes.
How much time do ambient AI scribes actually save?
Ambient AI scribes save roughly 1-2 minutes of note-writing per visit, or 13-16 minutes per clinical day — not the hour-plus that vendor marketing implies. An ambient AI scribe is software that records the clinician–patient conversation and drafts the clinical note for the clinician to review and sign.
The measured evidence, in order of rigour:
- The randomized trial. A three-arm pragmatic RCT of 238 outpatient physicians published in NEJM AI compared Microsoft's DAX Copilot, Nabla, and usual care. Nabla cut time-in-note by 9.5% relative to control; DAX showed no statistically significant effect. Both scribes were used in only about a third of visits — 33.5% for DAX, 29.5% for Nabla.
- The largest observational study. Across five academic medical centres, 1,800+ adopters versus 6,770 controls, scribes were associated with 13 fewer minutes in the EHR and 16 fewer minutes of documentation per eight hours of patient care, and roughly 0.5 extra visits a week. Only 32% of users used the tool in more than half their visits — and those users got two to three times the savings. STAT's report on the same study notes primary care clinicians benefited more than others.
- A single-site Abridge cohort. At the University of Chicago, Abridge users cut note-writing time by 1.8 minutes per appointment (15.9%) versus matched controls, with no significant change in after-hours documentation.
- The largest deployment. The Permanente Medical Group estimated 15,791 hours saved over 63 weeks across 7,260 physicians and 2.5 million encounters. By our arithmetic that is about 2.2 hours per physician over the whole period across everyone who touched it — the savings concentrate in the heavy users.
The steel-man for buying anyway: the same studies consistently show burnout and task-load improvements that outrun the minutes. That is a legitimate retention case. But write it down as a retention case, not a productivity case, or your CFO will measure the wrong thing in month nine.
Which ambient scribe integrates best with Epic?
On Epic, the choice is between native Epic AI Charting and Abridge's embedded workflow; everything else is a step further from the chart. Epic holds over 42% of the acute-care hospital market and nearly 55% of beds, so for most readers this is the whole decision.
Epic AI Charting rolled out on February 4, 2026. It listens during the visit, drafts the note, and suggests orders; Epic's CMO told STAT it will keep adding "orders" and "diagnoses". When Epic first announced it in August 2025, it said it would use Microsoft's Dragon Ambient AI technology for transcription while Epic generates the note. Epic has not published pricing. Who should not pick it yet: any system that needs published outcome data before a pilot — there is none for AI Charting specifically.
Abridge launched "Abridge Inside" in 2024, embedding recording in Haiku and the draft note in Hyperdrive so clinicians never leave Epic, and has since extended it to inpatient and outpatient orders and emergency medicine. It has the one peer-reviewed time-savings result tied to a named product in the table above (the Chicago cohort). Who should not pick it: a small practice wanting a self-serve contract, and any system whose recording-consent workflow is not already airtight (see the liability section).
Microsoft Dragon Copilot combines dictation and ambient capture. Mount Sinai chose it after a bake-off in which its digital chief said Microsoft and Abridge were "relatively close" and Microsoft won on business terms — while openly reserving the right to evaluate Epic's tool. That is the honest shape of most Dragon wins: the contract, not the note.
Ambience Healthcare is the pick when the business case is revenue. Cleveland Clinic rolled it out for documentation, clinical documentation integrity and point-of-care coding after a 2024 pilot across 80+ specialties. If you do not have a coding or CDI owner ready to measure capture, you are paying for a capability nobody will use.
Freed is built for individual clinicians and small groups. Its pricing page lists "EHR push integration" only on Premier and the custom Groups plan. Check exactly what that does in your EHR before assuming it matches an embedded Epic workflow — for a health system, it will not.
How good are the notes, and how much do clinicians edit them?
Expect clinicians to change about a tenth of each draft, and expect the dangerous errors to be omissions rather than inventions. Omission errors — something said in the room that never reaches the note — are the failure mode to test for, because a reviewing clinician cannot see what is missing.
- A pragmatic pilot of 31 physicians and 7,545 notes in JMIR Medical Informatics found a median 9.0% of AI-generated words were changed, 14.9% of notes were left entirely unedited, and 94.7% of reviewed notes were free of significant errors. Editing ranged from 1.9% to 69.3% per physician.
- An OHSU study of five scribe platforms on identical simulated encounters, in Mayo Clinic Proceedings: Digital Health, found note error rates of 16.7-23.3% for the four commercial platforms and 46.1% for a free one. About 76% of errors were omissions. The platforms were anonymised — which is precisely why you must run your own.
- Kaiser Permanente's quality review of its rollout rated notes 4.35 of 5 on a modified PDQI, with thoroughness scoring lowest and multi-speaker tracking a named weakness.
The 14.9% unedited figure is the one to worry about. A clinician who signs an unedited draft is either looking at a perfect note or not reading it, and you cannot tell which from the edit log alone. Epic says it and its customers track the number of changes made to drafts as a guardrail. Whatever you buy, demand that metric per clinician — and audit a sample of zero-edit notes against the audio before it is deleted. We covered what happens when nobody does in NHS Scribes Dropped 'Null.' Patients Caught It, Not GPs.
Who carries the liability: BAA, retention and training terms
The vendor contracts put the liability for the final note — and for anything lost when audio is deleted — on you. Read the retention clause before the accuracy slide.
Microsoft's own documentation is the clearest statement of where the industry stands. Dragon Copilot is "not an electronic health record system"; the customer is "solely responsible" for filing data into the EHR, and Microsoft disclaims liability for "any deletion by Dragon Copilot following its standard data retention periods". Outside Great Britain, it is not offered as a medical device. None of that is unusual. It is simply written down.
Abridge's public privacy policy says customer data processing is "governed exclusively" by customer agreements, including BAAs, and retained "as described in our agreement with that customer". The retention period and any training rights live in your contract, not on a web page — so negotiate them explicitly: audio retention in days, whether de-identified data may train models, and deletion certification.
Consent is where this becomes a lawsuit. A proposed class action filed in December 2025 alleges Sharp HealthCare recorded more than 100,000 patients with Abridge without consent, in a state that requires all-party consent, and that notes stated patients had consented when they had not. These are allegations, not findings — but the pattern is the risk. Auto-inserted consent language is a template setting you control.
Freed, at the small-practice end, states audio is "temporarily saved... then automatically deleted" and offers an organization-wide BAA on its Groups plan. Confirm individual plans are covered by a BAA before a single visit is recorded. For the wider questions, use the checklist in AI Vendor Security Review: 6 Questions That Change the Answer.
What does an ambient scribe cost per clinician?
Only Freed publishes prices; every enterprise vendor is "contact sales", and Epic has not disclosed AI Charting pricing. As of September 26, 2026, per Freed's pricing page: Starter $39 per clinician per month (40 notes), Core $79 (unlimited), Premier $119 monthly or $104 billed annually (adds EHR push integration). Abridge, Ambience and Dragon Copilot do not publish list prices, and we will not estimate them.
What actually drives your per-clinician number:
- Paid seats versus active users. The five-centre study found only 32% of users were heavy users. If you license every eligible clinician, most of your spend buys minimal savings. Negotiate on active use — the same trap we measured in M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Setting. Emergency and inpatient modules can be separate products; ambulatory is the baseline.
- Coding and CDI add-ons. These are where revenue claims come from, and where price rises.
- Term and exclusivity. Cleveland Clinic signed a five-year exclusive partnership with Ambience. With Epic's native tool maturing, a long exclusive now carries more switching risk than it did in 2025.
The decision: what predicts regret
The purchase you regret is the one chosen on a demo and licensed for everyone. The criteria that actually predict outcomes:
- Adoption at day 90, by specialty. The RCT's scribes were used in about a third of visits. Pilot with a control group and measure time-in-note from EHR audit logs, not surveys.
- Zero-edit rate and omission audits. Compare notes against audio on a sample every month, weighted to the clinicians who never edit.
- Consent workflow and retention. Written, all-party where state law requires, and never auto-asserted by a template.
- Exit terms. Can you move to Epic's native tool in 18 months without a penalty?
What changes the answer: if Epic publishes AI Charting pricing that bundles into your existing licence, the default for Epic shops flips toward native. If you are on Oracle Health or a smaller EHR, Epic drops out and Abridge or Dragon Copilot become the default on integration depth. If your case is revenue, Ambience leads regardless.
This Week: Pull your EHR audit-log baseline for time-in-note and after-hours time, by clinician. Without it, no vendor claim is checkable. This Month: Run a paired pilot — Epic AI Charting and one third party — with a usual-care control arm, as in Hospitals Test Vendor AI. Fewer Than Half Have a Sandbox.. Before Contract: Get audio retention, training rights, consent language and active-user pricing in writing.
The Bottom Line
Ambient scribes are the first generative AI tool clinicians have asked for rather than had imposed on them — and the evidence says they deliver a modest, real time saving and a larger morale one. Buy them like that. Pilot against a control, pay for the clinicians who use them, and own the consent and review process, because the contract already says you do.
Sixteen minutes a day is worth having. It is not worth a five-year lock-in.
Continue Reading
- NHS Scribes Dropped 'Null.' Patients Caught It, Not GPs.
- Hospitals Test Vendor AI. Fewer Than Half Have a Sandbox.
- FDA Cleared 1,357 AI Devices. Three Were Tested on Patients.
- AI Vendor Security Review: 6 Questions That Change the Answer
- M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Lung Cancer AI: Chart Data Held Up. Scans and Slides Didn't.