How much time does a doctor save when AI writes the medical note?

In the largest study so far (JAMA, 2026, 8,581 clinicians in the US), adoption was associated with 16 fewer minutes of documentation per 8 hours of scheduled visits. Heavy users saved 27.3 minutes. After-hours work did not fall significantly. The studies cover consultations in English; there are no published data for Romanian.

Am I allowed to record the consultation so the AI can transcribe it?

Romania’s Patients’ Rights Law (Law 46/2003) requires the patient’s consent for photographing and filming. Audio recording is not named explicitly, but the safest practice is the same standard: the patient is informed beforehand, gives consent, and the consent is recorded in the file. In the UCLA randomised trial, fewer than 10% of patients declined. Beyond consent, the clinic needs data processing agreements with its vendors and, depending on its size, an impact assessment.

Is software that proposes a diagnosis and an ICD-10 code a medical device?

It depends on the intended purpose declared by the manufacturer and on what the software actually does. According to the MDCG 2019-11 guidance, software that only stores records or replaces the paper file is not a medical device. If, however, it is a medical device and provides information for diagnostic or therapeutic decisions, Rule 11 of the MDR places it at least in class IIa. The UK MHRA guidance of July 2026 draws the line at suggesting diagnoses and at acting without the doctor’s verification. Ask the vendor whether the proposed code comes from what the doctor said or is inferred by the AI.

AI in private clinics: front desk and the consultation note

Q: What is an AI scribe and how does it write the consultation note?

An AI scribe is software that records the conversation between doctor and patient, transcribes it and proposes a structured note: reason for the visit, history, clinical examination, conclusions, recommendations. In the literature it is called ambient clinical documentation. The result is a draft, not a final note. The doctor reads it, corrects it and validates it.

The promise around AI in private clinics is a simple one: the consultation note writes itself. The largest study published so far on so-called “AI scribes” — software that listens to the consultation and drafts the medical note — followed 8,581 outpatient clinicians across five US academic health systems and measured something else (JAMA, April 2026). On average, clinicians who adopted one spent 16 minutes less on documentation for every 8 hours of scheduled visits. Evening work, the kind done after hours, did not change in a statistically significant way.

Sixteen minutes is not nothing. Nor is it the revolution from the sales decks. It is a real gain, measured against a comparison group, and it shows fairly precisely where AI helps in a clinic and where it does not. It takes over the drafting and the front-desk phone. It does not take over validating the note, the diagnosis, triage or emergencies. That boundary is what this article is about, written for owners, managers and doctors in private clinics and practices in Romania. The concrete example at the end is one of our own products, and we present it as such.

The phone and the note: where a private clinic loses time

In a private clinic, two administrative tasks compete with medical care. At the front desk, the phone rings for bookings, rescheduling, cancellations, questions about prices, test preparation or parking — often exactly when the receptionist is dealing with a patient at the desk, or after the clinic has closed. In the consulting room, the doctor listens, examines and then has to turn the conversation into a note: reason for the visit, history, clinical examination, conclusions, recommendations. Either it gets written during the visit, eyes on the screen, or filled in later, from memory.

The cost does not appear on any invoice. It shows up in the patient who cannot get through, in the note completed at the end of the day, when the details the patient mentioned have faded, and in the doctor’s fatigue. A 2023 study by the Romanian College of Physicians, cited by the public broadcaster TVR in November 2025, found that 36% of doctors are at high risk of burnout. The TVR report does not describe the methodology.

Improvised solutions fail in predictable ways:

A generic voice assistant, set up as if for a shop, also answers medical questions. A question about chest pain is not a booking.
Test results get read out over the phone to someone who only gave a name.
The consultation is dictated or pasted into a chatbot on the doctor’s personal account — health data sent to a provider the clinic has no contract with. We have written separately about shadow AI.
The AI-generated note is signed after a quick glance, even though, as the studies below show, the typical problem is not what appears wrong but what is missing.

What the studies show about AI-drafted consultation notes

An AI scribe (in the literature, “ambient clinical documentation”) records the conversation between doctor and patient, transcribes it and proposes a structured note, which the doctor corrects and signs. Read together, the studies published in 2025–2026 say something more nuanced than the brochures.

The JAMA study (Rotenstein et al., 2026) is the largest. It compared 1,809 clinicians who adopted an AI scribe with 6,772 who did not. Per 8 hours of scheduled visits, adoption was associated with 16.0 fewer minutes of documentation and 13.4 fewer minutes of total time in the electronic health record, plus 0.49 additional visits per week. Heavy users, who used the AI for at least half of their notes, saved 27.3 minutes of documentation. The study is not randomised and the centres are academic, so the result does not translate directly to a private practice.

The only large randomised trial comes from UCLA (NEJM AI, 2025): 238 doctors across 14 specialties, roughly 72,000 visits, split between two products and a control group. With Nabla, time per note fell from 4:30 to 3:49 minutes, against a drop from 4:22 to 4:04 in the control group — a 9.5% larger reduction, statistically significant. With DAX, the reduction was not significant. Burnout improved by about 7% relative to control, an effect the authors themselves call “modest”. Fewer than 10% of patients declined the use of the AI scribe.

At Kaiser Permanente, 7,260 doctors used AI scribes in more than 2.5 million visits over 63 weeks, with reported savings equivalent to 1,794 working days in one year. 84% of doctors reported a positive effect on their interaction with patients, and 47% of patients noticed that the doctor looked at the computer less. The data, however, come from the organisation that deployed the system and have no control group. Kaiser also reports less note-writing outside working hours. The JAMA study, which does have a comparison group, finds no significant change after hours (−3.1 minutes), and its authors suggest the time saved probably goes into other tasks.

On burnout, the pattern repeats. Studies without a control group find large drops: among 263 clinicians in six health systems, the share with burnout fell from 51.9% to 38.8% after 30 days (Olson et al., JAMA Network Open, 2025), and at Mass General Brigham prevalence fell by 21.2 percentage points at 84 days, in a joint study with Emory where response rates were 11–30%. The randomised trial finds the 7% mentioned above. The gap is explained by voluntary participation, low response rates and the novelty effect. The direction is consistent; the size of the effect is uncertain. An independent assessment by the Peterson Health Technology Institute from March 2025, cited by Fierce Healthcare, found evidence of reduced burnout but no proven financial return.

One detail that matters directly for a clinic in Romania: all of these studies are from the US, on consultations in English. There are no published data on how an AI scribe performs on consultations in Romanian.

Omission: the error that survives a re-read

The errors change how the tool has to be used. A real-world pilot (Taylor et al., JMIR Medical Informatics, 2026) followed 31 doctors and 7,545 notes, of which 356 were assessed in detail. Omissions appeared in 18% of them, hallucinations — information that had not been said — in 11.5%, and incorrect inclusions in 9.3%. 94.7% of notes had no significant errors, but 5.3%, that is 19 notes, had errors the doctors judged to carry serious or imminent risk. At the median, doctors changed 9% of the words.

A simulation published in Mayo Clinic Proceedings: Digital Health (Anderson et al., 2025) ran 14 simulated visits through five platforms. On average, 26.3% of key clinical elements were omitted or rendered incorrectly, and 76.3% of the errors were omissions. Each case produced, on average, 3.0 errors with the potential for moderate or severe harm. The figures are not directly comparable: one counts notes, the other clinical elements, on real versus simulated cases. But they show the same thing: the dominant error is omission.

That is hard to catch. A mistake you see when re-reading, because something appears that was not said. An omission you do not see, because the note looks complete: the allergy mentioned in passing, or the medication the patient said they had stopped, simply is not there.

A wrong note shows up on a re-read. An incomplete note looks exactly like a good one.

Hence the practical rule: the AI writes the draft, and the doctor validates the note, makes the diagnosis and is accountable for it. What also matters is which features help the doctor find the gaps, not just how fluent the text sounds: quotes from the conversation, to check where a statement comes from, and an explicit list of what was not discussed. There is one more boundary, a regulatory one. As long as the product only documents, it is a writing tool. Once it starts suggesting diagnoses, it may move towards medical-device territory, and the classification depends on the intended purpose declared by the manufacturer.

AI in private clinics: what to check before you switch it on

At the front desk

The assistant introduces itself as a virtual assistant in its first sentence. We have written about the AI Act transparency obligation and about recording calls in our article on a voice agent for a small business phone line.
It books against the real schedule — doctor, room and service duration — not a generic calendar. And for family members too, because the caller is not always the patient.
It answers only from the clinic’s own documents: services, prices, doctors, test preparation. When it does not know, it says so, and the question goes to reception.
It gives no medical advice, makes no diagnosis, does no triage. In an emergency it tells the patient to call 112. Test this before launch with questions phrased the way a frightened patient would ask them, not an engineer.
For test results it communicates only the status, never the values, and only after verifying identity. The doctor flags the result first, so the assistant does not interpret anything.
Know in advance who takes the call when a patient is frustrated or asks for a person, both during reception hours and outside them. For more examples by industry, including medical practices, see our article on AI voice agents in customer support.

In the consultation

1Define the roles. Health data are a special category (Article 9 GDPR), and the legal basis for medical care is Article 9(2)(h). The clinic is the controller. The AI platform, the voice provider and the hosting are processors and sign a data processing agreement with the clinic (DPA, Article 28).
2Decide whether to carry out a data protection impact assessment (DPIA, Article 35). Decision no. 174/2018 of ANSPDCP, the Romanian data protection authority, makes it mandatory for large-scale processing of health data and for large-scale processing using new technologies. The same text adopts the nuance in Recital 91 GDPR: the data of an individual doctor’s patients are, in principle, not “large scale”. For a single-doctor practice, a DPIA is not automatically mandatory. For a clinic with several doctors introducing transcription of consultations, it is prudent to carry one out.
3Ask for the patient’s consent and record it. Romania’s Patients’ Rights Law (Law 46/2003) requires consent for photographing and filming in a healthcare facility. Audio recording is not named explicitly, but the safest practice is the same standard: the patient is informed beforehand, gives consent, and the consent is recorded in the file. Consent is an additional safeguard; it does not replace the legal basis under the GDPR.
4Ask the vendor where the diagnosis code comes from. The MDCG 2019-11 guidance (revised in June 2025, not legally binding) says that not all software used in healthcare is a medical device. If software qualifies as a medical device and provides information used to take diagnostic or therapeutic decisions, Rule 11 of the MDR places it at least in class IIa. The decisive criterion is the intended purpose stated by the manufacturer. The clearest reference point, valid only in the United Kingdom, is the MHRA guidance of 29 July 2026. A scribe that transcribes, summarises or suggests codes matching terms explicitly spoken, for the doctor to verify, is not a medical device. One that suggests diagnoses or acts without verification is. The concrete question: “Does the proposed code come from what the doctor said, or is it inferred by the AI?”
5Check the AI Act timeline. An AI system that is a medical device and requires a notified body (classes IIa, IIb, III) is high-risk. Plain class I is not (MDCG 2025-6). The Digital Omnibus, in force since 27 July 2026, has postponed the high-risk obligations for medical devices to 2 August 2028. The transparency obligation in Article 50 already applies, since 2 August 2026.
6Run a local pilot, with verification. One or two doctors, a few weeks, notes read in full with attention to the list of what was not discussed, and your own time measured before and after. The studies are in English, so this is the only way to learn how it performs in Romanian.

A concrete example: Magic AI Clinici

Magic AI Clinici is our product, built on the same logic. At the front desk, the voice assistant, Ana (the clinic chooses the name), always introduces itself as a virtual assistant. It answers 24/7 on the clinic’s number and in the chat and voice widget on the website, handles several calls at once and speaks Romanian, English and German. It makes, moves, cancels and confirms appointments against the real schedule, by doctor, room and service duration, including for a child or a parent. It recognises returning patients by their phone number, and new patients spell out their name. When a patient is frustrated, it transfers the call to reception during reception hours and takes a message outside them.

For results, the assistant communicates only the status, after verification with the date of birth and the last 4 digits of the CNP, the Romanian personal identification number. After 3 failed attempts, access is blocked for 24 hours. The doctor flags the result first (no issues, follow-up needed or urgent), and the patient receives the link by email. The assistant also calls patients: reminders 24 hours and 2 hours before, with a text message if the patient does not answer, a call to book a follow-up when a result is urgent, and rescheduling when a doctor becomes unavailable.

The digital consultation is a separate option. The doctor records from the browser only with the patient’s written consent, recorded in the file. The audio recording is deleted immediately after transcription. The AI proposes a structured note: reason for the visit, history, clinical examination, paraclinical investigations, conclusions, a diagnosis with a proposed ICD-10 code, and recommendations. It adds quotes from the conversation for verification and a list of what was not discussed — tools designed for the omission problem, which nonetheless do not replace reading the note in full. The doctor corrects and validates, the note is saved in the patient’s folder in Google Drive, and recommendations go out by email only if the doctor decides so. The product page does not claim medical-device status, and neither do we here. The question in point 4 is one a clinic has every right to ask us too.

The data sit in the clinic’s own Google Sheets and Google Drive. The CNP and the notes are encrypted, and every opening or change of a note goes into the audit log. The app has separate roles, doctors see only their own patients, and doctors and managers sign in with two-factor authentication. At the patient’s request, the data are exported as a ZIP archive or deleted. Calls do, however, pass through the voice provider: recordings and transcripts are deleted there after 30 days, and transcripts are kept encrypted for 90 days, after which only the summary remains. That is why the data processing agreement matters here too.

The limits, stated openly: Magic AI does not yet integrate with other clinic software, so the app becomes the clinic’s booking system. Nor is it integrated with CNAS/SIUI, the reporting system of Romania’s national health insurance. For a clinic contracted with the health insurance fund, or with software it does not want to replace, that can decide everything. On the product page there is a fictitious clinic, “Clinica Magic AI”, on the test number +40 373 818 611 (a Romanian number), where you can make, move and cancel appointments, plus a narrated presentation and the cost calculator. If you need something other than a standard product, we also build custom AI agents and applications.

Conclusion

AI does not write the note instead of the doctor. It writes the draft, which saves real minutes but not hours, and its mistakes are mostly what it leaves out, so its value depends on how easily it lets the doctor find the gaps before signing. At the front desk the rule is the same: the assistant books, confirms and informs, while triage, result values and emergencies stay with people. If you would like to look together at what AI could take over in your clinic and what should stay with your team, we suggest a 30-minute conversation, with no obligation.

Sources

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, doi:10.1001/jama.2026.2253 (jamanetwork.com)
Lukac PJ, Turner W, Vangala S et al. — “Ambient AI Scribes in Clinical Practice: A Randomized Trial”, NEJM AI, 2025, DOI 10.1056/AIoa2501000; UCLA Health press release, 26 November 2025 (uclahealth.org)
Kaiser Permanente / The Permanente Medical Group — NEJM Catalyst, DOI 10.1056/CAT.25.0040; summary on permanente.org, 7 April 2025
Olson et al. — “Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout”, JAMA Network Open, 2025;8(10):e2534976 (newswise.com)
You JG et al. — Mass General Brigham and Emory study, JAMA Network Open, August 2025 (news.emory.edu, jamanetwork.com)
Peterson Health Technology Institute — assessment of AI scribes, March 2025, cited by Fierce Healthcare and Axios (fiercehealthcare.com, axios.com)
Taylor et al. — “Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study”, JMIR Medical Informatics, 2026;e86474 (medinform.jmir.org)
Anderson TN, Mohan V, Dorr DA et al. — “Evaluating the Quality and Safety of Ambient Digital Scribe Platforms Using Simulated Ambulatory Encounters”, Mayo Clinic Proceedings: Digital Health, 2025;3(4):100292 (pmc.ncbi.nlm.nih.gov)
Romanian College of Physicians (Colegiul Medicilor din România) — burnout study, 2023, cited by TVR Info, 6 November 2025 (tvrinfo.ro)
Regulation (EU) 2016/679 (GDPR) — Articles 9, 28, 35 and Recital 91 (eur-lex.europa.eu); ANSPDCP (Romanian data protection authority) — Decision no. 174/2018, Official Gazette no. 919 of 31 October 2018
Romanian Law no. 46/2003 on patients’ rights, Article 20, consolidated version published by the Ministry of Health (ms.ro)
MDCG 2019-11 rev.1 — guidance on qualification and classification of software as a medical device, June 2025 (health.ec.europa.eu); Regulation (EU) 2017/745 (MDR), Annex VIII, Rule 11
MHRA — “Guidance on Ambient Voice Technology-Enabled Products: Device qualification and classification”, v1.0, 29 July 2026 (gov.uk)
MDCG 2025-6 — FAQ on the interplay between the MDR/IVDR and the AI Act, 19 June 2025, via Hogan Lovells (hlc.com)
Analyses of the Digital Omnibus on AI and the AI Act timeline (gibsondunn.com, houthoff.com); Regulation (EU) 2024/1689 (AI Act), Article 50