Who watches the AI in the exam room? Not the FDA.

A primary-care visit sounds like this from the exam room. A clinician asks about sleep. The patient describes a side effect. A parent fills in a detail about a child's asthma. Behind the conversation, an ambient AI scribe is listening and writing a draft chart note that the clinician will later edit and sign.

The scribe is doing clinical work, capturing the encounter that drives billing, coding, and the medical record, but as of 2026 it is not regulated as a medical device. The Food and Drug Administration's 2016 clinical decision-support guidance classifies ambient AI documentation as administrative clinical software, placing it outside the agency's medical-device oversight — a boundary the agency has not reopened for ambient AI. The classification is the load-bearing policy fact behind a new national study of how fast this technology has spread, and it is the part wire coverage skips.

Researchers at Emory University's Rollins School of Public Health, writing in the American Journal of Managed Care, examined 6,561 US hospitals and identified 2,784 (42.4%) that run Epic as their primary inpatient record system. Of those, 1,744, or 62.6%, had implemented an ambient AI documentation tool by mid-2025 (AJMC study; Emory release).

Three products account for more than 80% of those adoptions: Microsoft's DAX Copilot, Abridge, and ThinkAndor. The concentration matters because the deployment pattern is no longer a curiosity at the margin. It is the dominant experience at large US hospitals, and the regulatory line that put ambient scribes outside device oversight was drawn around a different kind of software.

The FDA's clinical decision-support guidance, issued in 2016, drew the line at software whose output a physician can independently review. Static alerts about drug interactions or dosing thresholds fit that test. Ambient documentation fits it on paper, but the practical workflow, a clinician editing an AI's draft of a routine visit, signing within minutes, and moving to the next patient, is a different kind of software. The 2016 rule now governs a deployment pattern that did not exist at scale when it was written, and the agency has not reopened the boundary for ambient AI.

Adoption is also uneven. The Emory team found large, nonprofit, metropolitan hospitals with higher operating margins and heavier staffing-adjusted workloads were the most likely to have rolled out an ambient scribe. Adjusted predicted adoption probability reached 70.2% at nonprofit hospitals versus 28.8% at for-profit hospitals, and uptake was lower in the Midwest than in the South (AJMC study).

The AJMC authors warn that without targeted support for resource-constrained hospitals, the technology risks widening the existing gap between well-funded academic medical centers and the rural or community safety-net systems that already run thinner. A second independent signal points in the same direction. A 2026 Doximity survey of more than 3,100 physicians found voice-based documentation use rose from 20% to 29% in under a year. Doximity's figures are self-reported and directional rather than an audit, but they confirm the AJMC pattern from the physician side of the room (Medical Daily summary).

The regulatory category matters because it determines who is accountable when a draft note captures a symptom the patient mentioned, when a coding suggestion shapes the bill, or when a clinician edits the AI's summary under time pressure. Ambient scribes are not making diagnostic calls, and the FDA's safety question is the right question to ask of software that does. The governance question the 2016 guidance does not answer is what the chart note actually is, who owns the edits, and who is on the hook when the workflow moves faster than the documentation.

For the category to catch up, two things have to move. The FDA would need to update or replace its 2016 clinical decision-support guidance to address ambient documentation explicitly. A payment or accreditation lever, whether CMS conditions of participation, Joint Commission standards, or state scope-of-practice rules, would have to put a floor under documentation quality at the hospitals least able to absorb a vendor failure. The Emory authors' concrete near-term is targeted adoption grants, shared template libraries, and on-site implementation support for resource-constrained hospitals. The regulatory redraw is the harder lift, and the 2016 guidance that defines the boundary remains the binding document for the technology in the exam room.