Is an AI Medical Scribe Worth the Cost for a Small Practice

The quote is somewhere between $39 and $119 per clinician per month, the deck says “get two hours back every day,” and the rep has offered a free two-week trial. You have three clinicians, a 3% operating margin and a nurse leaving in November. The question is not whether ambient documentation works. It is whether this is where the next $3,600 a year should go.

Short answer. The business case for an AI medical scribe almost never closes on billing more. It closes on reclaimed documentation time after hours, and whether that is worth the subscription depends on what you would otherwise do with the hour. If the honest answer is “sleep,” the tool can be excellent value and you should stop trying to justify it on revenue. If the honest answer is “nothing, I would still be here because the phones and the prior auths are the real bottleneck,” you are buying a fix for your fourth-biggest problem.

That makes the decision personal rather than financial, which is unsatisfying, and it is also the only defensible reading of the evidence.

General information for practice owners, not legal, clinical or coding advice. Your compliance officer, coder and attorney get the final word.

What the studies actually measured, and where perception and stopwatch part ways

Two February 2025 papers from the same Stanford Health Care pilot make the point better than any vendor case study.

Shah and colleagues surveyed 48 physicians across a three-month pilot (paired analysis n=38) and found a statistically significant 1.94-point reduction on their burnout measure, a 24.42-point reduction in task load and a 10.9-point improvement in usability, all at p<.001 (JAMIA 32(2):375–380). Large, consistent, exactly the sort of result that ends up on a slide.

In the companion paper, Ma and colleagues measured what happened in the EHR for 45 physicians across eight ambulatory disciplines over the same three months, covering 17,428 encounters with the scribe used in 9,629 of them, a 55% utilisation rate (JAMIA 32(2):381–385). The median reduction in time per note was 0.57 minutes. Thirty-four seconds. Median daily documentation time fell 6.89 minutes, after-hours EHR time 5.17 minutes, total daily EHR time 19.95 minutes. All statistically significant, and all a long way from two hours a day.

A 24-point drop in perceived task load alongside a 34-second drop in measured time per note. Both findings are real. They measure different things, and vendors quote the first.

The larger measured effects come from Duggan and colleagues at a Philadelphia academic health system, in JAMA Network Open on 19 February 2025: a prospective single-group pre-post study of 46 clinicians across 17 specialties and 22 outpatient sites, April to June 2024. Time in notes per appointment fell 20.4%, from 10.3 to 8.2 minutes. After-hours work time, the pajama time that actually matters to you, fell 30.0%, from 50.6 to 35.4 minutes a workday (P=.02). Same-day appointment closure rose from 66.2% to 72.4%. Note length rose 15.5%.

Read that study’s last line too. Usability scored a respectable 76.6 on the System Usability Scale, and the Net Promoter Score was zero: 35.1% promoters, 35.1% detractors. A third of the clinicians who used it for three months would not recommend it. A UCI Health pilot in JAMIA (February 2026, 167 physicians) found the same shape: longer notes, less note-writing time, better survey scores.

None of these had a control group, and all used self-selected volunteers in large systems with implementation support you will not have. Treat them as the optimistic end of the range, not the midpoint.

Three ways it could pay for itself. Usually only one of them does

1. Reclaimed after-hours time. The real one. The evidence band is roughly 5 to 15 minutes a day of after-hours EHR time. Not nothing if you are finishing notes at 9:40pm four nights a week. Also not revenue: it becomes money only if you were paying someone for that hour or would otherwise have cut a session.

2. Extra patients per session. Usually a fantasy. The arithmetic is tempting: 2 minutes saved across 20 appointments is 40 minutes, so add two slots. But most of the saving is after-hours minutes, which open no slot, and the in-session minutes arrive scattered in ones and twos and cannot be booked. Converting them means changing the schedule template, which means betting the saving holds every day, including the days the tool mangles a complex visit. Test that deliberately after three months if you have a waiting list. Do not put it in the purchase justification.

3. Coding and documentation completeness. Handle carefully. The legitimate version: notes that capture the counselling, severity indicators and medical decision making you actually did but were too tired to write at 10pm. Guo and colleagues interviewed 30 clinicians at UCI Health about why they edit ambient drafts (JAMIA 33(7):1345–1353) and found that supporting coding and evidentiary accuracy was one of the four main reasons, precisely because drafts omitted decision making, counselling detail and severity of illness. The tool tends to under-document the part that supports a higher level of service, not over-document it.

There is also an illegitimate version, and it is the fastest way to turn a $100 subscription into a payer audit. A longer note is not a better-supported code. CMS’s fact sheet on office and outpatient E/M payment (11 January 2021) states that practitioners report the visit level based on either medical decision making as set out in the AMA CPT guidance or total practitioner time on the day of the visit, and that reviewers “use the medical record documentation to objectively determine the medical necessity of the visit and accuracy of the documentation of the time spent.” Note length is not in that sentence. If a pitch carries a level-of-service lift as a line item, ask which documentation element changed, and have your coder audit 20 notes. Work from the CPT E/M guidelines and CMS guidance, not a vendor ROI calculator.

The cost stack that is not in the quote

Anchor on published pricing. As of this writing, one widely used vendor lists $39 a month for up to 40 notes, $79 for unlimited notes, and $104 billed annually ($119 monthly) for the tier that pushes the note into the EHR and writes patient instructions and referral letters, with group pricing on request. Note what sits behind the top tier: EHR write-back matters more than the price gap.

The rest of the stack:

EHR integration. Copy-and-paste costs nothing, and it also means every note passes through a clipboard until someone pastes into the wrong chart. A real write-back may need an interface or API access from your EHR vendor, who may charge for it. Get their number before you sign with the scribe vendor.
Template and macro rebuild. Your phrases, macros and note templates were tuned over years. Ambient output does not slot into them cleanly, and month one goes on rebuilding what you already had.
Training and the trust period. The biggest hidden cost, and nobody prices it. For two to four weeks you will read every line of every note carefully, which is slower than writing the note yourself was. Budget for a net loss in month one.
Administration. Provisioning, deprovisioning when someone leaves, the BAA, and someone checking usage so you are not paying for a licence abandoned in week three.

Break-even in hours, not dollars

Stop computing an ROI percentage. Compute the price of an hour of your own evening, then decide whether it is fair. With the published figures, substituting your own:

Take $100 per clinician per month, so $1,200 a year, and 220 clinical days.

Optimistic case (Duggan figures): 15.2 fewer after-hours minutes a day is 55.7 hours a year. $1,200 ÷ 55.7 = about $22 per hour of evening reclaimed.
Conservative case (Ma figures): 5.17 fewer after-hours minutes a day is 19 hours a year. $1,200 ÷ 19 = about $63 per hour.
Your trial shows 2 minutes a day: 7.3 hours a year, or about $164 per hour. Decline.

Now answer the question that decides it. At $22 an hour, almost every physician-owner should buy: you would not get out of bed for $22, and the hour is worth more than that to you. At $63, it depends on whether that evening hour is the one where you read to your kids or the one where you stare at a wall. At $164, no.

Run that per clinician before you buy three licences. The physician doing 24 narrative visits a day and the one doing procedures with a four-line note land in different places.

Accuracy, attestation, and the review habit that protects you

You sign the note. That never moves. No licensed second professional stands behind an ambient tool the way one stands behind a human scribe, and no vendor indemnity changes who authenticated the record.

Which means the tool cannot be managed like a scribe who is accountable. It is a draft generator, and you are the only reviewer.

The error pattern is documented and it is not random. Biro and colleagues ran simulated encounters through two ambient products, analysed 44 draft notes (JAMIA 32(5):928–931, May 2025) and found the products failed differently: one produced drafts that took significantly longer to edit, with fewer omissions but more additions and misplaced text. The UCI Health interviews found clinicians editing for misheard medications, content attributed to the wrong speaker, discussion of people who were not the patient, missing pertinent negatives, and overcertainty, where probabilistic reasoning was flattened into definitive statements. One clinician described saying “possibly this” and the AI writing it as fact.

That last one is the clinically dangerous failure, and skimming will not catch it. A workflow that holds:

Read the assessment and plan, the medication list and every number in full, every note, forever. Skim the narrative if you must, never these.
Check your differential still reads as a differential and your hedges survived.
Sign in the room or straight after, while you remember the visit. Batch-signing eleven notes at 8pm from memory is how a fabricated detail gets attested to.
For the first month, tally every correction on paper: omissions in one column, invented or misattributed content in the other. That tally is your accuracy data, and it beats any vendor benchmark.

The BAA, the recording, and the FDA question everyone gets wrong

The vendor is a business associate and you need an executed BAA before the first patient is recorded. HHS’s sample business associate agreement provisions set the elements to check: permitted uses and disclosures defined specifically, subcontractors bound to the same restrictions, an obligation to report any use or disclosure not provided for by the agreement including breaches of unsecured PHI, and return or destruction of PHI at termination with no copies retained. That last clause is the one people skip and the one that matters when you switch.

Five questions the BAA will not answer:

Is audio retained, and for how long? Get a number in writing. One vendor publishes that audio is held temporarily until the note and quality checks are complete, then automatically deleted. That is the specificity you want. “We take security seriously” is not.
Is your data used to train or improve their models, and how do you switch that off? The answer belongs in the contract, not on a marketing page that can be edited next Tuesday.
Consent to recording. Recording-consent law is state law and it varies, with some states requiring all parties to consent rather than one. Ask your attorney which rule applies, then build the plainest possible practice: a line in the check-in paperwork, signage, and a spoken sentence from the clinician offering an easy opt-out. Do not let the vendor’s generic consent language be your only control.
What happens when it is down, or the internet is? Ask specifically whether the encounter is lost or recoverable, and what you are meant to do mid-clinic. You need a fallback that is not “type it from memory at 7pm.”
Can you leave with your work? Notice period, and whether signed notes live in your EHR rather than only in the vendor’s portal. Output that exists only inside their system is a dependency, not a subscription.

On regulation, verify rather than assume. Section 520(o)(1) of the FD&C Act, added by section 3060 of the 21st Century Cures Act, excludes several software categories from the device definition, among them software for administrative support of a health care facility, software serving as electronic patient records to the extent it transfers, stores, converts formats or displays the equivalent of a paper chart, and software for transferring, storing, converting or displaying device data “unless such function is intended to interpret or analyze” it. FDA’s line sits at interpretation or analysis intended to inform diagnosis or treatment, which is where the clinical decision support criteria in 520(o)(1)(E) apply. A tool that drafts a note from a conversation generally sits on the excluded side; a feature that suggests diagnoses, flags conditions or proposes codes edges toward the line. Get the vendor’s written regulatory position on each feature you switch on, and be clear what it means: no clearance also means no regulator has assessed whether the notes are accurate. That assessment is yours.

What to price it against, and how to run a trial worth trusting

The scribe is not competing with doing nothing. It competes with four cheaper or better-aimed options, and a tight-margin practice should price all of them:

A human virtual scribe. Costs considerably more per month and does considerably more: the note, plus the orders queue, the referral letter, the prior auth pack, the patient instructions. If your bottleneck is everything around the note, a person is the better buy and it is not close.
A shared in-person scribe across two clinicians. Awkward to schedule, unbeatable in procedural and high-touch clinics, and the scribe learns your practice.
Template and macro optimisation. Free, dull, often worth more than either. Pull your EHR usage analytics, find the note types eating your time, fix those five templates. Do this first whatever you buy, or you will pay a subscription to automate a bad template.
Restructuring the visit. Document in the room and close the note before the patient stands up. The highest-return unpaid change available to most primary care physicians, and the hardest, which is why people buy software instead.

If you still want the trial, the rep will offer two weeks. Take four, and run it so it tells you something.

One clinician, and not the most enthusiastic one. The early adopter will love it regardless; pick the sceptic who still files notes on time.
Pull four weeks of baseline first from your EHR’s clinician efficiency reporting: time in notes per appointment, after-hours minutes per day, same-day note closure. Without a baseline you are measuring your mood.
Change nothing else, the schedule template least of all. One variable.
Run four weeks, not two, and discard week one, the trust-building period, which will look terrible.
Measure two things separately. The EHR metrics, and a one-line diary entry each night: what time did I actually finish? When they disagree, believe the EHR for the business decision and the diary for the human one. Keep the edit tally alongside.

Adoption is normal now: the AMA’s Center for Digital Health and AI surveyed 1,692 US physicians in January and February 2026 and found 81% using AI professionally, 30% of them for progress notes, care plans or discharge instructions. Self-reported, voluntary survey. It tells you everyone is doing this. It does not tell you it pays for itself in your practice.

So who should buy. A physician regularly finishing notes after 8pm, in a narrative specialty, with stable staffing, no larger operational fire, templates already fixed, who can honestly say the reclaimed hour has a use. For that person, $22 to $63 an hour of evening is one of the better deals in practice management, and the burnout findings, soft as they are, are probably real for them too. Who should not: the practice whose margin problem is denials, no-shows, unanswered phones or an AR pile. None of those are documentation problems.

If the diagnosis comes back as the second kind, that is the work AB7 Solutions does: medical scribing with real people where a tool cannot carry the load, plus billing and RCM support, eligibility and prior authorisation, and remote healthcare admin staffing under a proper BAA. Worth a twenty-minute call before you sign any subscription, if only to rule it out. Call +1 321 341 7733, write to ab@ab7solutions.com or director@ab7solutions.com, or look at www.ab7solutions.com.

Sources: Duggan MJ et al., Clinician Experiences With Ambient Scribe Technology to Assist With Documentation Burden and Efficiency, JAMA Network Open, 19 February 2025; Shah SJ et al., Ambient artificial intelligence scribes: physician burnout and perspectives on usability and documentation burden, JAMIA 32(2):375–380; Ma SP et al., Ambient artificial intelligence scribes: utilization and impact on documentation time, JAMIA 32(2):381–385; Guo Y et al., Evaluating ambient artificial intelligence documentation, JAMIA 33(2):273–282, and Clinicians’ rationale for editing ambient AI–drafted clinical notes, JAMIA 33(7):1345–1353; Biro JM et al., The value of simulation testing for the evaluation of ambient digital scribes, JAMIA 32(5):928–931; CMS, PFS Payment for Office/Outpatient E/M Visits fact sheet (11 January 2021) and Evaluation and Management Visits; FDA, Changes to Existing Medical Software Policies Resulting from Section 3060 of the 21st Century Cures Act and Clinical Decision Support Software; HHS Office for Civil Rights, Sample Business Associate Agreement Provisions; AMA Center for Digital Health and AI, physician AI sentiment report (n=1,692, fielded 15 January to 2 February 2026, self-reported); Freed published pricing (vendor-listed, accessed September 2026).