Klinmate

Do AI medical scribes really save doctors time? What the first randomised trials found.

Documentation is the biggest drain on a doctor's day. In 2025 the first randomised trials of ambient AI scribes were published. The results are more nuanced than the marketing, and more useful.

of EHR and desk work per hour with patients
~2 h
physicians in the first scribe RCT
238
better burnout scores with AI scribes
~7%

Klinmate team · · 8 min read · 7 references

Key takeaways

  • In ambulatory practice, doctors spend nearly 2 hours on EHR and desk work for every hour of direct patient time.
  • In a 238-physician randomised trial, one AI scribe cut note-writing time by about 9.5% relative to usual care; the other showed no significant time saving.
  • Both trials found meaningful improvements in burnout and work exhaustion, even where time savings were small.
  • AI notes occasionally contain clinically significant errors, so every note still needs a doctor's review before signing.

The paperwork problem, in numbers

A landmark time-and-motion study in Annals of Internal Medicine observed physicians in four specialties and found that they spent 49.2% of their office day on electronic health record (EHR) and desk work, but only 27% in direct face time with patients. For every hour with patients, they spent nearly two more hours on screens and paperwork[1].

EHR event logs tell the same story. Family physicians in a Wisconsin health system spent about 5.9 hours of an 11.4-hour workday in the EHR, including about 1.4 hours after clinic. Documentation, ordering, billing and other clerical work made up 44% of that time, and the inbox another 24%[2]. This "pajama time" is consistently linked to burnout.

What is an ambient AI scribe?

An ambient AI scribe listens to the consultation (with the patient's permission), converts the conversation to text, and uses a large language model to draft a structured clinical note. The doctor reviews, edits and signs it. The idea is simple: let the doctor look at the patient, not the keyboard.

The underlying technology has matured quickly. In a 2024 study in Nature Medicine, clinicians reading summaries of clinical text rated the output of the best adapted language models as equivalent to or better than summaries by medical experts in most cases[3]. Engineering research continues to target the main weakness, clinical accuracy. For example, a 2025 paper in the IEEE Journal of Biomedical and Health Informatics proposes new methods to make multi-turn medical dialogue systems more clinically accurate[4].

What the randomised trials found

UCLA Health, NEJM AI 2025. 238 outpatient physicians from 14 specialties were randomised to one of two commercial ambient scribes (Microsoft DAX Copilot or Nabla) or to usual care, across about 72,000 patient encounters over two months[5].

  • Time: Nabla users' average time per note fell from about 4 min 30 s to 3 min 49 s, a 9.5% greater reduction than the control group. The DAX group's reduction was smaller and not statistically significant[5].
  • Well-being: Both scribe groups showed modest improvements of around 7% in burnout scores, with lower cognitive workload and work exhaustion[5].
  • Safety: Physicians reported that notes occasionally contained clinically significant inaccuracies, mainly omissions and pronoun errors, and one mild patient-safety event was reported[5].

UW Health, NEJM AI 2025. In a 24-week stepped-wedge randomised trial with 66 clinicians, ambient AI significantly reduced work exhaustion and interpersonal disengagement by 0.44 points on a 5-point scale, and reduced task load[6].

Kaiser Permanente, NEJM Catalyst 2024. Before the trials, The Permanente Medical Group made an ambient scribe available to 10,000 physicians and staff. Within 10 weeks, 3,442 physicians had used it in up to 303,266 encounters. Early evaluation found high-quality draft notes and positive patient feedback, alongside a small number of errors[7].

What this means for your practice

The honest summary: AI scribes are not a time machine, but they do make the work feel lighter. Measured time savings per note are modest and depend on the product. The more consistent benefit is lower cognitive load and exhaustion, which matters in a profession where burnout drives people out of clinical work. The trials also suggest the benefit grows with how consistently doctors use the tool.

Three cautions apply:

  • Review every note. Errors of omission are easy to miss when a note looks fluent. The doctor remains responsible for what is signed.
  • Ask for consent. Patients should know when a conversation is being processed by AI and be able to say no.
  • Know where the data goes. Ask any vendor where recordings and transcripts are processed and stored, how long they are kept, and whether they are used to train models.

Questions to ask before choosing an AI documentation tool

  • Has it been evaluated in a published study, ideally a randomised one, and in a setting like yours?
  • Does it support your language(s), including mixed Arabic and English consultations if that's how your clinic talks?
  • How does it show uncertainty, and how easy is it to correct a draft?
  • Is audio deleted after transcription? Where is data hosted, and is it used for training?
  • How does the note fit into your records, and does it save time end-to-end, not just in typing?

Where Klinmate fits

Klin is the AI built into Klinmate for doctors, specialists and their teams. It is being designed around the lessons above: AI drafts and assists, and the clinician reviews and decides. Klinmate is launching soon. Request a demo or see what's included.

This guide is general information for clinic owners and healthcare professionals. It is not medical, legal or regulatory advice, and it is not a substitute for your own professional judgement.

References

  1. C. Sinsky, L. Colligan, L. Li, M. Prgomet, et al., "Allocation of physician time in ambulatory practice: a time and motion study in 4 specialties," Ann. Intern. Med., vol. 165, no. 11, pp. 753–760, 2016. doi.org/10.7326/M16-0961
  2. B. G. Arndt, J. W. Beasley, M. D. Watkinson, J. L. Temte, et al., "Tethered to the EHR: primary care physician workload assessment using EHR event log data and time-motion observations," Ann. Fam. Med., vol. 15, no. 5, pp. 419–426, 2017. doi.org/10.1370/afm.2121
  3. D. Van Veen, C. Van Uden, L. Blankemeier, J.-B. Delbrouck, et al., "Adapted large language models can outperform medical experts in clinical text summarization," Nat. Med., vol. 30, no. 4, pp. 1134–1142, 2024. doi.org/10.1038/s41591-024-02855-5
  4. Z. Liu, Y. Quan, X. Lyu, M. J. F. Alenazi, et al., "Enhancing clinical accuracy of medical chatbots with large language models," IEEE J. Biomed. Health Inform., vol. 29, no. 9, pp. 6395–6405, 2025. doi.org/10.1109/JBHI.2024.3470323
  5. P. J. Lukac, W. Turner, S. Vangala, A. T. Chin, et al., "Ambient AI scribes in clinical practice: a randomized trial," NEJM AI, vol. 2, no. 12, 2025. doi.org/10.1056/AIoa2501000
  6. M. Afshar, M. R. Baumann, F. Resnik, J. Hintzke, et al., "A pragmatic randomized controlled trial of ambient artificial intelligence to improve health practitioner well-being," NEJM AI, vol. 2, no. 12, 2025. doi.org/10.1056/AIoa2500945
  7. A. A. Tierney, G. Gayre, B. Hoberman, B. Mattern, et al., "Ambient artificial intelligence scribes to alleviate the burden of clinical documentation," NEJM Catalyst, vol. 5, no. 3, 2024. doi.org/10.1056/CAT.23.0404

How we write guides: every figure is taken from the cited source, and we say where evidence is weaker (preprints, before-and-after or single-site studies). Spotted an error? Email [email protected] and we'll correct it.

Run a calmer clinic with Klinmate.

Appointments, patient records, payments and Klin, the AI front desk on WhatsApp, in one place. Launching soon.

Request a demo Read more guides