Few clinical technologies have been adopted as quickly as ambient AI scribes: software that listens to a consultation and drafts the clinical note. The appeal is obvious. Documentation is a major driver of clinician burnout, and a tool that gives time back is an easy sell.
The evidence base has caught up enough to say something more useful than "promising". As someone who designs healthcare data and AI platforms, here is my reading and how I would deploy them.
What the trials found
The clearest evidence so far comes from a three-arm randomised trial at UCLA, published in NEJM AI. It enrolled 238 outpatient physicians across 14 specialties and randomly assigned them to one of two commercial scribes or usual care12. One product produced a statistically significant reduction in time spent per note, of around 41 seconds compared with 18 seconds in the control group; the other's reduction was not significant against control1. Both showed potential benefits on burnout and work-related stress, and the authors called for longer, multi-site studies1.
Other trials point in a similar direction. A multistate pragmatic trial reported reduced documentation time and lower work exhaustion3, and a University of Wisconsin trial published in NEJM AI examined both efficiency and practitioner well-being4.
The fair summary: real but modest time savings that vary by product, with the clearest benefits in how clinicians feel about their work. That is worth having, but it is not the transformation some business cases assume.
The error problem
Accuracy is the harder question. In a JMIR Medical Informatics pilot, physicians reviewed a sample of AI-generated clinic notes. Omissions were the most common error, found in 18% of reviewed notes, followed by hallucinations at 11.5%. Most errors were rated mild to moderate, but about 2.5% of notes contained errors rated as posing serious or imminent risk5.
Those numbers come from one pilot and will differ by product and setting. The pattern is what matters: errors are common enough that clinician review is not optional, and some are serious. A note is a legal and clinical record. An omitted allergy or an invented examination finding can travel through the record for years.
The controls I would put in place
| Area | Control |
|---|---|
| Patient trust | Clear notice to patients that the consultation is recorded for documentation, with an easy way to decline |
| Clinical safety | Mandatory clinician review and sign-off; the draft is never filed automatically |
| Quality monitoring | Track edit rates, error reports and specialty differences; sample notes for structured review |
| Data protection | Defined retention for audio and transcripts, encryption, and contracts that limit vendor use of recordings |
| Integration | Write drafts into the EHR through standard interfaces, with provenance marking the note as AI-assisted |
| Evaluation | Local pilots comparing products on your specialties before scaling, not just vendor-reported figures |
Audio and transcripts are new stores of highly sensitive data. They belong in the asset inventory and data-flow maps I describe in the proposed HIPAA Security Rule update, and they should be covered by the same encryption and access controls as the record itself.
Measure it like any clinical intervention
- Baseline first. Documentation time from EHR audit logs, after-hours work and a validated burnout measure, before rollout.
- Compare products on your own workload. The UCLA trial showed different products performing differently1. Your specialties and patient mix will matter.
- Build an evaluation set of de-identified or synthetic encounters with reference notes, so updates to the vendor's model can be re-tested. The approach in evaluation sets that predict production applies directly.
- Watch for drift. Vendors update models frequently. Track edit rates and error reports per release.
- Report honestly. If time savings are small but well-being improves, say so. That is still a valid reason to invest.
The bigger opportunity
A scribe produces structured clinical content at the point of care. Mapped carefully into FHIR and on to analytical models such as OMOP, that content could improve data quality for research and population health. That only works if the notes are accurate, which brings us back to review and monitoring.
How I can help
I help health systems and health tech companies evaluate, integrate and govern clinical AI, from data flows and security to measurement. See healthcare data architecture, or analyse your challenge and bring the result to a call.
- 1
- 2
- 3
- 4Studies find AI technology for clinical documentation aids efficiency and reduces burnoutUniversity of Wisconsin · 12 Dec 2025
- 5Physician review of errors in AI-generated clinic notes (e86474)JMIR Medical Informatics · 2026