What problem does it solve?
Clinicians spend a large share of every working day on documentation. During the visit they type while the patient talks, and after clinic they finish notes and letters in the evening, the time clinicians call "pajama time". Patients notice the screen between them and their doctor, and health systems that deploy scribes name the documentation load as a driver of clinician burnout.
Human scribes and dictation help, but they are expensive or still take clinician time. Generative AI changed the economics: speech recognition that copes with a real consultation, followed by a language model that turns the conversation into a structured note in the clinician's preferred format. The risk moved with it. A fluent note that contains something nobody said, or leaves out a symptom, ends up in the medical record unless the clinician catches it.
How does it work?
- Ask for consent. The clinician tells the patient that an AI scribe will listen and records the consent; the patient can decline or stop it at any time, even mid visit.
- Capture the conversation. A phone, tablet or workstation app records the visit (in person, phone or video) and streams it to speech recognition with speaker separation.
- Draft the note. A language model turns the transcript into a note in the clinician's template and specialty format (history, examination, assessment and plan), and optionally a letter, patient instructions or suggested codes.
- Review and sign. The draft appears in the health record or next to it; the clinician checks it against what happened, edits it and signs it. Nothing enters the record unreviewed.
- Learn from edits. Edit rates, clinician feedback and quality samples show where the drafts are weak, per specialty and per template.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Mainstream
- Channels
- Internal tools, Mobile app
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 17,000 to 2.5 million | 2 | 2 organization |
| Handling time reduction | Too few to pool | 8.2% | 1 | 1 organization |
| Productivity gain | Too few to pool | 13.4% | 1 | 1 organization |
| Users served | Not pooled | 7260 | 1 | 1 organization |
Value drivers: Employee productivity, Customer experience, Speed and cycle time, Lower cost to serve.
Indicative value
A health system with 1,000 clinicians using an ambient scribe
USD 666,667 to USD 3.8 million
Clinician documentation time released per year
How this is calculated
Formula: clinicians * encountersPerClinician * minutesSaved / 60 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Clinicians using the scribe clinicians, clinicians | 1,000 | 1,000 | The reference health system. |
| Documented encounters per clinician per year encountersPerClinician, encounters per clinician per year | 2,000 | 3,000 | Editorial assumption for outpatient and primary care clinicians. Replace with your own visit volumes. |
| Documentation minutes saved per encounter minutesSaved, minutes per encounter | 0.2 | 0.5 | Derived from The Permanente Medical Group analysis on this page, with two assumptions of our own: a working day of 8 hours, and that the saving of 1,794 working days "in one year" can be set against the more than 2.5 million encounters counted over the 63 week evaluation. That gives about 0.34 minutes per encounter on average (about 0.42 if the encounters are scaled to 52 weeks). The top third of users accounted for 89% of activations, and high users saved two and a half times more per note than infrequent users, so the average is close to what frequent users saved (roughly 0.35 to 0.45 minutes); occasional users saved less. The range stays around that evidence. Replace with your own time data. |
| Fully loaded cost per clinician hour costPerHour, USD per hour | 100 | 150 | Editorial assumption. Replace with your own fully loaded clinician cost. |
What it leaves out: Time released is not cash saved unless it becomes extra appointments or less overtime. The figure leaves out licence and integration costs, clinician review time for drafts, the effect on burnout and retention, and any change in coding completeness.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
US Department of Veterans Affairs, Veterans Health Administration
United States · Healthcare · 2026
The 2025 federal AI use case inventory lists Abridge and Knowtex ambient scribe pilots at VA, both flagged as high impact. As of June 2026 VHA has deployed Ambient Scribe to all Patient Aligned Care Team primary care providers, including physicians, physician assistants and advanced practice nurses, after a phased rollout that began at 10 VA medical centers. With the Veteran's verbal consent the tool drafts the progress note from the conversation; the provider reviews and edits it before signing it into the record, and Veterans can opt out at any time, even mid visit. VA plans to extend it to selected outpatient specialty care. No outcome figures are published by VA.
No outcome disclosed.
Great Ormond Street Hospital for Children NHS Foundation Trust
United Kingdom · Healthcare · 2025
An NHS England sponsored study led by the GOSH DRIVE innovation unit tested the TORTUS ambient scribe at nine London sites, including hospitals, GP practices, mental health services and ambulance teams, over more than 17,000 patient encounters. The tool transcribes the consultation and drafts a clinic note and letter that the clinician checks and edits before saving. Direct patient interaction time rose and appointments got shorter; in A&E at St George's University Hospital, clinicians saw more patients per shift. A rollout across GOSH outpatient settings was planned to follow.
- Interactions handled: at least 17,000, evaluation across nine London NHS sites
"Over 17,000 patient encounters were evaluated across a diverse range of sites including hospitals, GP practices, mental health services and ambulance teams."
Claimed by: organization - Handling time reduction: 8.2%, overall appointment length, trial sites
"Results showed a 23.5% increase in direct patient interaction time during appointments, alongside an 8.2% reduction in overall appointment length when AI-scribes were used."
Claimed by: organization - Productivity gain: 13.4%, A&E at St George's University Hospital, patients seen per shift
"A&E saw particularly strong results, with a 13.4% increase in patients seen per shift."
Claimed by: organization
Kaiser Permanente
United States · Healthcare · 2024
Kaiser Permanente made an ambient documentation tool from Abridge available to doctors and other clinicians at its 40 hospitals and more than 600 medical offices in August 2024, after a year of testing. With the patient's consent, the tool listens to the visit and drafts the clinical note, which the clinician reviews before it enters the record. An analysis by The Permanente Medical Group in Northern California, published in NEJM Catalyst, found that the scribes saved the equivalent of 1,794 working days in one year, and that time savings were concentrated among the most frequent users. The tool does not make decisions or recommendations about care.
- Users served: 7260, The Permanente Medical Group, 63 week evaluation period
"AI scribes were used by 7,260 Permanente physicians in more than 2.5 million patient encounters during the evaluation period."
Claimed by: organization - Interactions handled: at least 2.5 million, The Permanente Medical Group, 63 week evaluation period, patient encounters
"AI scribes were used by 7,260 Permanente physicians in more than 2.5 million patient encounters during the evaluation period."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Note templates and documentation standards per specialty
- A consent process and patient information text
- A clinical safety case and a data protection impact assessment
- A sample of real consultations (with consent) to test draft quality per specialty
Systems to integrate
- Electronic health record for patient context and filing the signed note
- Clinician devices (mobile app, desktop, dictation hardware)
- Identity and single sign on for clinicians
- Audit logging and records retention for audio and transcripts
Complexity: Medium
Mature products exist, so the work is in integration with the health record, consent and information governance, clinical safety assessment, device and network setup in clinics, and training clinicians to review drafts properly. Specialty templates and languages take tuning.
- 1
Start with willing clinicians in a few specialties
At The Permanente Medical Group in Northern California, mental health, emergency medicine and primary care doctors were the most likely to use the scribe. Pick specialties with long conversations and heavy notes, and clinicians who want the tool.
- 2
Settle consent, retention and safety before go live
Decide how consent is asked and recorded, whether audio is kept or deleted after the note is signed, and complete the clinical safety assessment and data protection impact assessment.
- 3
Tune templates per specialty
Build note formats with clinicians in each specialty and let individual clinicians adjust style, so the draft needs editing rather than rewriting.
- 4
Train clinicians to review, not to trust
Teach what the tool gets wrong (medication names, negations, who said what) and make clear that the signature means the clinician has checked the note.
- 5
Measure time and quality, then widen
Track documentation time, after hours time, edit rates and patient feedback against a baseline, and expand to more specialties once quality holds.
Guardrails
- No note enters the record without clinician review and signature
- Patient consent recorded for every encounter, with an easy way to decline or stop
- The scribe documents only; it does not suggest diagnoses or place orders without clinician action
- Audio and transcripts retained only as long as policy allows, with access logging
- Health data processed in approved regions under HIPAA or GDPR special category rules
KPIs to instrument
- Documentation time per encounter and after hours time in the record, before and after
- Share of encounters where the scribe is used, per clinician and specialty
- Edit distance between draft and signed note
- Omissions and invented content found in quality samples
- Patient consent and decline rates, and patient feedback
Human in the loop
The clinician reviews, edits and signs every note and remains accountable for the record. A clinical safety officer owns the risk log, and a quality team samples signed notes against transcripts to find omissions and invented content.
Common failure modes
- Invented or misattributed content
- The draft contains a symptom, medication or statement nobody said, or attributes the patient's words to the clinician. Sample notes against transcripts and teach clinicians what to check.
- Automation complacency
- Clinicians sign drafts after a glance because they are usually right. Measure review time and edit rates and make quality feedback visible.
- Low use after launch
- Most of the time savings go to frequent users; occasional users gain little. Support adoption with training and specialty templates rather than counting licences.
- Consent that is not real
- Patients are not told clearly or feel they cannot refuse. Script the consent, make declining easy and track decline rates.
What are the risks and rules?
EU AI Act
Depends on design
A scribe that only transcribes and summarises for a clinician to review is not listed in Annex III and is usually minimal risk, although the provider of a system that generates text can still owe the Article 50(2) duty to mark output as AI generated, unless an exception such as an assistive function for standard editing applies. If the product qualifies as medical device software under the EU Medical Device Regulation and needs a notified body assessment, for example because it suggests diagnoses or treatment, it becomes high risk under Article 6(1) and Annex I. Health data in audio and notes falls under GDPR Article 9 in every case.
Guidance
- MHRA clarifies regulatory status of ambient voice technologies used in the NHS (Medicines and Healthcare products Regulatory Agency, Europe). Confirms that products used solely for transcription, summarising consultations, drafting letters or suggesting codes for clinician review are not regulated as medical devices in Great Britain, while products that support diagnosis or treatment, or act without clinician review, are.
- Article 6, classification rules for high risk AI systems (European Union, Europe). An AI system that is, or is a safety component of, a product covered by EU harmonisation legislation such as the Medical Device Regulation and needs third party conformity assessment is high risk.
- Regulation (EU) 2017/745 on medical devices (European Union, Europe). Decides whether a scribe with clinical functions is medical device software, and its risk class.
Controls to put in place
- Clinical safety case and hazard log for the scribe, owned by a named clinical safety officer
- Data protection impact assessment covering audio, transcripts and vendor processing
- Consent procedure and patient information in plain language
- Audit trail of drafts, edits and signatures per note
- Periodic quality sampling of signed notes against source audio or transcripts
When it went wrong elsewhere
- OpenAI's transcription tool hallucinates more than any other, experts say, but hospitals keep using it. An Associated Press investigation reported that the Whisper speech model can invent text, and that a Whisper based medical transcription tool from Nabla, used by over 30,000 clinicians and 40 health systems, deletes the original audio, so transcripts cannot be checked against the recording. Nabla said clinicians must edit and approve notes.
Frequently asked questions
- How much time do ambient AI scribes save?
- The Permanente Medical Group in Northern California reports that AI scribes saved its physicians the equivalent of 1,794 working days in one year; over a 63 week evaluation they were used in more than 2.5 million encounters, and high users saved two and a half times more time per note than infrequent users. In an NHS England sponsored study led by Great Ormond Street Hospital, appointments were 8.2% shorter and A&E clinicians saw 13.4% more patients per shift.
- Is an AI scribe a medical device?
- It depends on what it does. The UK MHRA confirmed in July 2026 that tools used solely to transcribe, summarise, draft letters or suggest codes for clinician review are not medical devices, while tools that support diagnosis or treatment, or act without review, are. In the EU, a scribe that qualifies as medical device software under the Medical Device Regulation and needs a notified body is also high risk under the AI Act.
- What are the main safety risks?
- Invented content, omissions and misattributed statements in a note that looks complete. Keep clinician review and signature mandatory, sample signed notes against transcripts, and be careful with tools that delete the audio before anyone can check the transcript.
How to cite this page
Blits.ai AI Use Case Library, "AI ambient scribe for clinical documentation", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/ambient-clinical-documentation. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published