AI Scribes: Modest Time Savings, Remarkable Burnout Reduction

What the gap tells us about building in health-tech

Ask anyone who has worked in a clinical setting why AI scribes took off, and you will usually hear the same answers.

First, clinicians were spending hours after work finishing notes in Electronic Health Record systems. It is a problem so common it earned its own name: pyjama time, the documentation that follows you home and gets finished on the couch after the kids are in bed.

Second, documentation was changing the experience of the consultation itself. Clinicians were trying to stay present with patients while simultaneously listening, typing, structuring notes, checking the screen, and holding in mind what would need to be finished later.

The pitch was straightforward: fix the notes, save time, reduce stress.

In important ways, the technology appears to work. But the data tells a more interesting story about how.

What the data actually shows

A multisite study tracking 8,581 clinicians across five major US academic health systems found that AI scribes saved around 16 minutes of documentation time per 8 scheduled patient hours (Rotenstein et al., 2026). The authors called this a "modest" finding. Modest seems right when set against the broader documentation burden: Sinsky et al. found that physicians spent around half their office day on EHR and desk work, and only about a quarter in direct clinical face time with patients. The Rotenstein study also found no significant average reduction in after-hours EHR time. In other words, the objective time savings were real, but limited. Pyjama time did not disappear.

But a separate study tells another side of the story. Olson et al. followed 263 clinicians using ambient scribes for 30 days and found that burnout dropped from 51.9% to 38.8%, a 13.1 percentage-point reduction. Clinicians also reported lower note-related cognitive task load and better focused attention on patients. (That figure deserves a caveat: the study followed self-selected adopters over just 30 days, with no control group and self-reported measures, so some of the drop may reflect early enthusiasm rather than a durable effect.)

Nevertheless, the contrast is the interesting part.

The strongest effects were not necessarily where the efficiency case would have predicted them. Objective time savings were modest. After-hours EHR time did not significantly change in the larger multisite study. Yet clinician-reported burnout and cognitive load improved substantially in separate research. That does not mean time savings were irrelevant. It means time savings may have been a lossy proxy for something more important: the cognitive experience of the clinical encounter.

The mechanism the metrics underweighted

The obvious response to documentation burden was to make note-taking faster. But the emerging evidence suggests the burden was never only about time. It was also about what documentation does to the clinician during the patient encounter, and what it leaves running in the background afterwards.

A clinician taking notes during a consultation is doing more than recording information. They are listening, interpreting, structuring, composing, checking the screen, maintaining rapport, preserving clinical details, and keeping a running list of what will need to be finished later.

That is a divided-attention task.

In cognitive load terms, much of this is extraneous load: burden created not by the clinical substance of the work, but by the way the work has to be performed. Documenting while listening forces the clinician to split attention between the patient and the note. The screen becomes a competing participant in the encounter.

There is also a second layer, which is less visible but arguably just as important. It's the feeling that their evening has already been claimed before they have left the building.

This is what we might call an open loop: an unfinished, must-be-completed task that the mind continues to hold in the background. The concept is grounded in task-incompletion research, including the Zeigarnik effect, which describes the tendency for unfinished tasks to remain cognitively active until they are resolved.

Before the scribe, the clinician is not just doing two things at once. They are also carrying a growing background process: "I will need to remember, reconstruct, and finish all of this later."

The scribe changes that experience. It does not simply make the note faster. It changes when and where the cognitive burden is carried. The note begins to close during the encounter itself.

That is a plausible explanation for why the felt benefit can be larger than the measured time saving. The scribe may relieve two burdens at once: divided attention in the room and the open loop that would otherwise accrue for later.

This interpretation is consistent with the cognitive-load evidence. Hudson et al. reported a 46.6% lower composite NASA-TLX score with an ambient scribe. Olson et al. separately found a significant reduction in note-related cognitive task load and an increase in focused attention on patients.

The important point is not that loop closure has been definitively proven as the causal mechanism. It has not. The point is that the pattern of evidence points toward a behavioural mechanism that standard efficiency metrics only partially capture.

Time saved is easy to count. Loop closure is harder to measure. That is exactly why it matters.

Why adoption was more precarious than it looked

The adoption data also tells a useful story.

Across the five health systems in the multisite study, only around one in five clinicians adopted AI scribes. Of those adopters, only about a third used the tool in at least half of their patient visits.

That matters because the efficiency case alone, roughly 16 minutes saved per shift, does not look strong enough to explain category-level enthusiasm on its own.

What appears to have mattered for sustained users was the felt experience: less mental juggling, more patient focus, less note-related strain, and a stronger sense that the encounter could be completed in the room rather than reconstructed later.

This is where the lesson extends beyond AI scribes.

Many health-tech products are evaluated through the metrics that justified the purchase: time saved, clicks reduced, throughput increased, forms completed, billing captured, etc. Those metrics matter. But they can still miss the behavioural mechanism that determines whether the product becomes part of real work.

A product can hit its documented metrics and still miss the human need underneath them. That is the gap product teams need to find early.

The rebound risk: when efficiency gets reabsorbed

As AI scribes become part of clinical infrastructure — nearly two-thirds of US hospitals using Epic systems have already adopted ambient AI tools (Yang & Graetz, 2026) — the question is no longer simply whether they reduce documentation time. It is what health systems do with the capacity they create.

This is where the rebound effect becomes relevant.

When technology creates efficiency, organisations often reinvest that efficiency into more activity rather than better experience. In healthcare, that could mean using AI scribes to increase visit volume rather than reduce cognitive strain, improve patient connection, or strengthen care quality.

That would be a serious misread of the value.

If the real benefit of AI scribes is partly cognitive relief, then converting every recovered minute into additional throughput risks eroding the very mechanism that made the tool valuable. The system becomes more efficient on paper while the clinician remains cognitively overloaded in practice.

The technology may work, but the incentive structure can still waste the benefit.

This is why efficiency metrics need to be interpreted within a wider health-system frame. The Quintuple Aim asks us to consider patient experience, population health, clinician well-being, health equity, and cost. Against that standard, the key question is not simply whether ambient AI saves time. It is whether the saved effort is reinvested in ways that improve care, preserve clinician capacity, and reduce inequity.

If measurement stops at minutes, clicks, and note completion, it may capture the purchase case while missing the value case.

Note bloat and the return of the open loop

There is also an emerging risk around note bloat.

AI scribes can generate structured notes quickly, but early evidence suggests they may sometimes lengthen documentation rather than tighten it. Owens et al. found that heavy users of an ambient tool saw note length grow by 542 characters even as documentation time fell. Stults et al. also found significant increases in both total note length and progress-note length.

The effect is not universal. Some studies report shorter notes among frequent users. And longer notes are not automatically worse. Extra length may reflect useful clinical detail rather than redundancy.

But from a behavioural perspective, the question is sharper than "are the notes longer?" The question is: does the output actually close the loop?

If the AI produces a long, redundant, or clinically noisy note that the clinician has to heavily review and edit later, the open loop has not disappeared. It has moved downstream.

That creates a testable behavioural prediction: the burnout benefit should decay as back-end editing burden increases.

This is the kind of mechanism health-tech teams should be measuring deliberately. Not just whether the tool produces output, but whether the human user experiences the task as resolved.

The deskilling question

There is another issue that deserves careful handling: what happens to clinical thinking when clinicians stop writing their own notes?

This is not an argument against AI scribes. It is a question about automation and human expertise.

Human-factors research has long described two risks: automation complacency, where reliance reduces vigilance, and automation bias, where people defer to system outputs even when they are wrong.

In medicine, those risks are not theoretical. Dratsch et al. showed that incorrect AI suggestions degraded radiologist performance across experience levels. Budzyń et al. found that after endoscopists were routinely exposed to AI assistance, their unaided adenoma detection rate fell from 28.4% to 22.4% when AI was removed.

Those are not AI scribe studies, and they should not be presented as evidence that AI scribes deskill clinicians. But they do establish a broader principle: when expertise is mediated by automation, teams should periodically measure what happens to human performance without the tool.

Documentation is often dull, repetitive, and burdensome. But the act of composing a note may also do useful cognitive work: reinforcing memory, sharpening interpretation, and consolidating the clinical encounter.

A recent MIT Media Lab preprint on AI-assisted essay writing offers an illustrative, non-clinical example. Participants who repeatedly offloaded writing to an AI assistant showed reduced neural engagement and weaker recall. The study is preliminary, and it is about students writing essays, not clinicians documenting care. It cannot carry the clinical claim on its own.

But read alongside the broader deskilling literature, it sharpens the question: if AI scribes remove the friction of note construction, what useful cognitive work might disappear with the burden?

The answer may be "very little." It may be "only in some specialties." It may be "only when review quality is poor." We do not know yet.

That is the point.

The value and risk of ambient AI may live in the same undermeasured layer. The value shows up as cognitive relief. The risk, if it exists, may accrue silently as cognitive offloading.

The only way to see both is to measure the human system, not just the tool.

The patient in the room

Most AI scribe research still views the patient through the clinician's eyes. That is understandable, given the original problem was clinician documentation burden. But it leaves a major gap.

The early patient evidence is reassuring in some respects. In one survey of 2,202 patients following outpatient visits, most found ambient AI scribes helpful and wanted their clinician to use the tool again.

But broader public awareness remains thin. In a separate survey of 12,153 Canadian adults, only 28.3% were aware of AI scribes. Many people are being asked to consent to something they do not yet fully understand.

Consent is not the same as comfort. And comfort is not the same as trust.

This matters because one of the most important variables in a clinical encounter is disclosure: what the patient is willing to say.

We already know from research on sensitive questions that people underreport stigmatized information when they feel exposed. Substance use, mental health symptoms, sexual health, domestic violence, medication adherence, financial stress, and other sensitive topics are not disclosed in a vacuum. They are shaped by context, perceived judgement, privacy, and trust.

An AI scribe could affect disclosure in either direction.

Some patients may hold back because a silent recording technology is present. Others may disclose more if the tool makes the clinician more attentive, less distracted, and more emotionally available. Some may not care at all.

The problem is not that we know AI scribes suppress disclosure. We don't. The problem is that ambient AI is often treated as if it were behaviourally neutral in the room, when that assumption is still under-evidenced.

You can measure documentation time. You can measure note length. You can measure clinician satisfaction.

It is much harder to measure the sentence a patient decided not to speak.

That is why patient trust cannot be treated as a consent workflow alone. It is an active condition of the encounter.

The behavioural layer is where value is preserved or lost

At Experience Foundry, this is the kind of problem we care about: how innovation is adopted, used, trusted, and sustained in the real world.

The question is deceptively simple: have you mapped the behaviour your product is actually changing?

Not just the feature. Not just the workflow. The behaviour underneath it: the cognitive patterns, friction points, moments of trust, open loops, avoided actions, and background tasks your product sits inside.

When teams skip that layer, they often miss where value is actually being created. They also miss where it will break under scale.

AI scribes were built around a real clinical pain point. They reduce documentation burden. They improve the experience of some clinicians. They may well become a durable part of clinical infrastructure.

But the most interesting lesson is not simply that AI scribes work.

It is that the strongest value may not sit exactly where the initial measurement frame expected it to sit. The efficiency story is real, but incomplete. The deeper story is cognitive relief, loop closure, patient presence, trust, and the careful management of automation risk.

That is a broader lesson for health-tech.

The teams that build durable products are usually the ones that find the behavioural mechanism early, measure it deliberately, and have the discipline to follow the data when it tells them something different from what they expected.

That is the work Experience Foundry helps teams do.

We help organisations look beneath the surface metrics of innovation: adoption, retention, engagement, workflow fit, trust, cognitive load, and value realisation. Not because efficiency does not matter, but because efficiency alone rarely explains whether a product becomes part of real human behaviour.

If you are measuring your product's success only through efficiency metrics, you may be missing the mechanism that determines whether value is actually created.

______________________________

References

Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterology & Hepatology. 2025;10(10):896–903.

Dratsch T, Chen X, Rezazade Mehrizi M, et al. Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology. 2023;307(4):e222176. doi:10.1148/radiol.222176

Hudson TJ, Albrecht M, Smith TR, et al. Impact of Ambient Artificial Intelligence Documentation on Cognitive Load. Mayo Clinic Proceedings: Digital Health. 2025;3(1):100193. doi:10.1016/j.mcpdig.2024.100193 (figures per published correction)

Kosmyna N, Hauptmann E, Yuan YT, et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. 2025.

Latkin CA, Edwards C, Davey-Rothwell MA, Tobin KE. The relationship between social desirability bias and self-reports of health, substance use, and social network factors among urban substance users in Baltimore, Maryland. Addictive Behaviors. 2017;73:133–136.

Nundy S, Cooper LA, Mate KS. The Quintuple Aim for Health Care Improvement: A New Imperative to Advance Health Equity. JAMA. 2022;327(6):521–522.

Olson KD, Meeker D, Troup M, et al. Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout. JAMA Network Open. 2025;8(10):e2534976.

Owens LM, Wilda JJ, Hahn PY, Koehler T, Fletcher JJ. The association between use of ambient voice technology documentation during primary care patient encounters, documentation burden, and provider burnout. Family Practice. 2024;41(2):86–91.

Parasuraman R, Manzey DH. Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors. 2010;52(3):381–410.

Chandrasekaran R, Moustakas E. Patient attitudes toward ambient artificial intelligence scribes in clinical care: insights from a cross-sectional study. Journal of the American Medical Informatics Association. 2026;33(2):263–272. doi:10.1093/jamia/ocaf218

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

Shah SJ, Naumann Murtagh K, Char D, et al. Patient perspectives on clinicians' use of ambient AI scribes. JAMIA Open. 2026;9(3):ooag104.

Sinsky C, Colligan L, Li L, et al. Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties. Annals of Internal Medicine. 2016;165(11):753–760.

Stults CD, Deng S, Martinez MC, et al. Evaluation of an Ambient Artificial Intelligence Documentation Platform for Clinicians. JAMA Network Open. 2025;8(5):e258614.

Tourangeau R, Yan T. Sensitive Questions in Surveys. Psychological Bulletin. 2007;133(5):859–883.

Yang F, Graetz I. Ambient AI Tool Adoption in US Hospitals and Associated Factors. American Journal of Managed Care. 2026;32(1):e25–e30. doi:10.37765/ajmc.2026.89876

Zeigarnik B. Über das Behalten von erledigten und unerledigten Handlungen. Psychologische Forschung. 1927;9:1–85.

‍