By Dr. Adarsh K. Gupta, DO, MS — Author of The Science of Habits | Cochrane Associate Editor | 25+ Years Clinical Practice

An ambient AI scribe listens in the exam room and helps write the note. It can give some minutes back.

The important question is not “Did we buy a scribe?” It is:

What happens to those minutes?

They can go to:

  • The patient in the room (more eye contact, fewer glances at the screen)
  • The clinician’s evening (less charting after dinner)
  • The schedule (another visit squeezed in)

Those savings are real — but usually modest, and they differ by product. A dashboard is good at counting visits. Attention has to be chosen on purpose. After 25 years in practice, I keep watching tools arrive faster than the habits we need to use them well.

What ambient AI scribe research actually shows

Three findings are solid enough to plan around. They are not a sales slogan. They are a strategy brief.

1. You get some minutes back — and the brand matters

Big observational study (JAMA, April 2026) — Rotenstein, Holmgren, and colleagues [1]:

  • Followed clinicians across five centers (1,809 adopters out of 8,581 clinicians)
  • Adopters spent about 13 fewer minutes in the chart and about 16 fewer minutes documenting, per 8 hours of scheduled care
  • Weekly visits rose by about half a visit
  • After-hours chart time did not clearly fall
  • Primary care saw larger savings (~25 fewer minutes of chart time per 8 hours)

On their baseline, chart work already eats about 2.3 hours of every 8 hours of patient care. Sixteen minutes back is helpful — and people who choose the tool early are not the same as people who wait.

Randomized trial (NEJM AI, November 2025) — Lukac, Chin, and colleagues [3] (NCT06792890):

  • 238 outpatient physicians, 14 specialties
  • Compared Microsoft’s DAX Copilot, Nabla, and usual care
  • Nabla cut time-in-note by about 9.5% vs control
  • DAX’s ~1.7% change looked like no clear difference
  • Clinicians used the tool on roughly one-third of visits (either product)
  • Clinically important errors were described as occasional with both
  • The trial recorded one mild adverse event

Bottom line: An ambient AI scribe is not one product. Ask which tool, for whom, and with what proof.

2. Exhaustion can ease — while “meaning” stays flat

Stepped-wedge trial (NEJM AI, November 2025) — Afshar and colleagues [4] (NCT06517082):

  • 66 practitioners over 24 weeks
  • About 38% of notes used the ambient tool
  • Work exhaustion and “checking out” from people fell by 0.44 points on a 5-point scale
  • Sense of professional fulfillment barely moved (+0.14) and missed the study’s preset bar
  • Documentation fell by about 0.36 hours a day
  • The after-hours win disappeared once extreme days were removed

Bottom line for buyers: Less drain does not automatically mean more meaning. Staffing, autonomy, and team climate still decide whether a lighter note becomes a better day.

3. A hopeful first month is a signal — not a finish line

30-day study (JAMA Network Open, October 2025) — Olson and colleagues [5]:

  • Burnout (single question) fell from 51.9% to 38.8%
  • Feeling of undivided attention rose a bit
  • Self-reported after-hours documentation fell by about 0.9 hours

Hold the design next to that hope:

  • Before-and-after (not a strong trial design)
  • About one month
  • One platform
  • Roughly 60% of enrollees finished both surveys

Bottom line: A good first month is a reason to keep measuring, not to declare victory.

For a GRADE-style read of population, comparison, outcome, and certainty, use CertaintyGrade.com. The evidence is real and uneven. It is already enough to stop treating “the scribe” as one single intervention.

Where does a saved ambient AI scribe minute go?

In the large JAMA adoption study [1], a clear pattern showed up:

  • Documentation time fell
  • Visits rose a little
  • After-hours chart time did not clearly fall
That pattern is a choice, even when nobody writes it down.

Saved minutes can go to:

  1. The patient — more undivided attention (what Olson’s group heard [5])
  2. The evening — inbox, sleep, recovery (fragile in Afshar’s trial [4])
  3. Throughput — about half an extra visit a week means the organization already assigned the minute to capacity

Aaron Tierney, Kristine Lee, and Vincent Liu wrote beside that JAMA study [2]: scribes might help parts of the “quintuple aim” (population health, cost, experience, clinician well-being, equity) — while the evidence stays uneven across those aims.

A dashboard crowns whatever it can count. Unless leaders name a second outcome and measure it, the schedule wins by default.

Also notice whose discomfort the tool relieves:

  • More eye contact → helps the relationship
  • An added slot + a skimmed signature → helps the schedule and borrows against the record

Fluent draft errors are cousins of a problem I have covered before (AI hallucinations in medicine). Lukac’s trial found the ordinary version [3]: occasional meaningful inaccuracies, and one mild adverse event. The signature still owns the plan.

The habit underneath the ambient AI scribe switch

Three different things get confused:

| What happened | What it actually is | |---------------|---------------------| | Someone bought a product license | Procurement | | Someone turned the tool on | Utilization | | Someone signed a draft without reading | A safety risk |

In the UCLA trial, use was only about one-third of visits [3]. In the JAMA study, people who used the scribe on half or more of visits saved more documentation time [1]. The unread signature is the version most likely to cause harm. Several factors can affect scribe accuracy — for example, how loudly the patient speaks, accent, and distance from the microphone. That is why reviewing the draft before signing matters.

Two simple habits protect what the software returns:

  • Presence — eyes and a pause for the person in the room
  • Review — read the plan, the medication list, and anything the draft smoothed over before it becomes the chart

Both fail when:

  • The cue is weak (“I’ll review later”)
  • The reward is late (nobody thanks you for a careful note)
  • The template treats open time as waste

The Habit Loop, as a protocol, lives on HabitRx.app. A personal starting point is the Free Habit Score. Why the environment beats a lecture on willpower is the subject of The Science of Habits.

A schedule that spends every recovered minute will outrun any request to “be more present.”

Five questions before buying an ambient AI scribe

These match the Clinical AI Safety & Evaluation Checklist:

  1. What problem are we naming — and in whose words?
    Exhaustion scores? After-hours chart time? Feeling heard? Visit capacity?
    Afshar’s trial moved exhaustion and left fulfillment basically unchanged [4]. An unnamed aim gets replaced by whatever the dashboard already counts.

  2. Which product — and which evidence?
    Nabla and DAX did not perform alike [3]. Both were used on about 30% of visits.
    Ask for the outcome, the comparison group, and the specialty mix. A demo is the best afternoon. The contract has to survive an average Tuesday.

  3. Who reviews the draft when it is wrong?
    The note stays a draft until a clinician accepts it. Track inaccuracies as safety events.
    The FDA is still asking how generative clinical tools should be tested; physicians can weigh in through October 19 (the FDA’s generative AI questions). The review habit can start locally now.

  4. What will you do with the saved time?
    Write it where staff can see it. If the minutes become added visits, say that out loud when you announce the purchase.

  5. What would make you stop?
    Examples: skimmed signatures, almost nobody using the tool, a safety signal, or notes patients cannot follow. Stopping rules are part of safety.

Printable checklist: /clinical-ai#checklist. The Clinical AI Brief (footer on this site) is the ongoing version of the same conversation.

What this leaves the physician to protect

Patients will trust the clinician who can tell a saved minute from a spent one.

An ambient AI scribe is at its best when it returns:

  • The face
  • The pause
  • A note the clinician will own

It speeds the treadmill when every recovered minute is already assigned to the next slot.

The minutes are the receipt. Attention is the practice.

Frequently asked questions

Do ambient AI scribes reduce physician burnout?

Sometimes exhaustion eases. A deeper sense of fulfillment has not been clearly shown to rise.

  • Afshar: exhaustion and interpersonal disengagement fell 0.44 points; fulfillment missed the preset bar [4]
  • Olson: a single-item burnout measure fell from 51.9% to 38.8% in a short before-and-after design that cannot prove the scribe caused the change [5]

Do ambient AI scribes give clinicians their evenings back?

Publish your local number before you promise it.

  • Across five academic centers, Rotenstein found no clear drop in after-hours chart time, even as in-clinic documentation fell [1]
  • Afshar’s after-hours win depended on a few extreme days [4]

Are ambient AI scribe notes safe to sign without reading?

No. The clinician who signs still has to read.

  • Lukac’s trial: clinically significant inaccuracies were called occasional with both products [3]
  • One mild adverse event was recorded
  • A workflow that makes review optional has already weakened the safety case

Should a health system buy an ambient AI scribe to add visits?

Only if added visits are the goal you are willing to name.

  • The 2026 JAMA study linked scribe adoption with about half an extra visit per clinician per week, plus modest time savings [1]
  • Experience, equity, and a sustainable workforce are separate questions — and the accompanying editorial argued the evidence is still uneven there [2]

What should a clinician protect when the schedule absorbs ambient AI scribe time savings?

Protect two things:

  • Attention in the room
  • Reading the draft before it becomes the record

Use has hovered near a third of visits in trials [3], so the tool is off on many days it could have helped.

The Free Habit Score shows the cue and reward around those habits. The Clinical AI checklist asks the system version in advance: what is the recovered time for?

A note on scope

This is educational strategy for clinicians and health-system leaders. It is general information — not medical advice, a device evaluation, or a recommendation for or against any named product. Figures come from the cited studies. Local results will differ by specialty, vendor, how often people use the tool, and how saved time is spent. Decisions about a specific patient belong to the clinician who knows that patient.

Sources

  1. Rotenstein LS, Holmgren AJ, Thombley R, et al. Changes in clinician time expenditure and visit quantity with adoption of artificial intelligence–powered scribes: a multisite study. JAMA. 2026;335(16):1408-1417. doi:10.1001/jama.2026.2253
  2. Tierney AA, Lee K, Liu VX. Ambient AI scribes and the quintuple aim: what is counted—and what matters. JAMA. 2026;335(16):1393-1395. doi:10.1001/jama.2026.3529
  3. Lukac PJ, Chin AT, Turner W, et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI. 2025;2(12). doi:10.1056/AIoa2501000
  4. Afshar M, Ryan Baumann M, Resnik F, et al. A pragmatic randomized controlled trial of ambient artificial intelligence to improve health practitioner well-being. NEJM AI. Published online November 26, 2025. doi:10.1056/AIoa2500945
  5. Olson KD, Meeker D, Troup M, et al. Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Netw Open. 2025;8(10):e2534976. doi:10.1001/jamanetworkopen.2025.34976