EHR scribe vendors must log provenance
- Health IT guidance around ambient artificial intelligence scribes is shifting toward storing how a note was made, not just the final text. - The core ask is to preserve model names, versioning, prompts, timestamps, and clinician review steps alongside generated draft notes and edits. - The push tracks rapid scribe adoption and new federal transparency rules for health software. (healthit.gov)
Ambient artificial intelligence scribes are built to listen to a visit and draft a clinical note. The new governance push is to save the note’s origin story too: prompts, model version, timestamps, and who approved the draft. (jamanetwork.com) (build.fhir.org) In health records, that origin story is called provenance. The Health Level Seven Fast Healthcare Interoperability Resources standard uses a Provenance resource to record the agents, systems, and actions involved in producing a version of a record. (build.fhir.org) (fhir.hl7.org) That matters because most ambient scribes do not write a final chart entry on their own. They generate a draft from audio or transcript, and clinicians are expected to review, edit, and sign the note before it becomes part of the medical record. (ama-assn.org) (cmpa-acpm.ca) The policy pressure is rising as these tools spread. A 2025 JAMA Network Open commentary called ambient scribes the fastest-adopted generative artificial intelligence tool in health care, and a Peterson Health Technology Institute report said roughly 60 ambient scribe products were already being implemented in practice. (jamanetwork.com) (phti.org) Physician use of health artificial intelligence has also climbed fast. The American Medical Association said 81% of physicians in its 2026 survey reported using AI professionally, more than double the rate from 2023. (ama-assn.org) Federal regulators are already moving on transparency, even if their rules are not written specifically for ambient scribes. The Office of the National Coordinator’s HTI-1 final rule created new transparency requirements for certain artificial intelligence and predictive algorithms in certified health information technology. (healthit.gov) HL7’s draft Artificial Intelligence Transparency on FHIR work points in the same direction. Its examples call for model identification and versioning, generation timestamps, context for human oversight, confidence information, and human review status. (build.fhir.org) The practical implication for electronic health record vendors and scribe companies is narrower than “save everything forever.” Provenance records can capture the last meaningful system hop, the producing model, and the clinician’s review action without turning the patient chart into a raw log dump. That is an inference from current FHIR provenance guidance and the AI transparency implementation work. (build.fhir.org 1) (build.fhir.org 2) Health systems are asking for that level of traceability because the stakes are no longer theoretical. The Permanente Medical Group said its ambient scribe program was used in more than 2.5 million encounters in a year and saved an estimated 15,791 hours of documentation time. (ama-assn.org) As ambient notes become routine, the compliance question is shifting from “Was AI used?” to “Which model, with which prompt, on which draft, and who changed it before signoff?” Provenance is the record structure built to answer that. (fhir.hl7.org) (healthit.gov)
