Nearly two-thirds of US hospitals running Epic's electronic records platform had deployed an ambient AI documentation tool by mid-2025, according to a study in the American Journal of Managed Care. Physician use of voice-based documentation rose from 20% to 29% in under a year.
That pace is unusual for health IT, which typically moves through long procurement cycles and reluctant adoption. Ambient documentation spread partly by word of mouth among physicians.
The governance question is not whether these tools are unsafe. It is whether the frameworks determining accountability have developed at the same speed, and by several accounts they have not.
The Adoption Picture and Who Is Left Out
The numbers require care, because the denominator is narrower than headlines suggest.
Emory researchers examined a national sample of 6,561 US hospitals and identified 2,784, or 42.4%, using Epic as their primary inpatient system. Of those, 1,744, or 62.6%, had implemented ambient AI tools that capture patient-clinician conversations and generate draft notes for review. That is roughly two-thirds of Epic hospitals rather than two-thirds of all US hospitals. Three products, DAX Copilot, Abridge and ThinkAndor, accounted for use at more than 80% of adopting hospitals.
Adoption was uneven along lines that matter. It was higher among large hospitals, nonprofit facilities, metropolitan hospitals and those with stronger operating margins and higher staffing-adjusted workloads. Adjusted probabilities were 70.2% for nonprofit hospitals against 28.8% for for-profit ones, and uptake was lower in the Midwest than the South.
The authors note that uneven diffusion could widen disparities without targeted support for resource-constrained hospitals, which is a different concern from the safety questions that dominate discussion of clinical AI.
The individual-physician figure comes from a separate source: a 2026 Doximity survey of more than 3,100 physicians found voice-based documentation use jumped from 20% to 29% in under a year.
The Categories That Carry Different Risk
Clinical AI is not one thing, and the distinctions determine what oversight applies.
Documentation tools listen to a clinical encounter and generate a note. They perform clerical work, and as of 2026, ambient AI scribes are classified as administrative tools rather than medical devices, placing them outside FDA oversight.
Clinical decision support suggests diagnoses, flags findings or recommends treatment. The FDA's January 2026 revised guidance drew a specific line: such software escapes medical device regulation only if the clinician can independently verify the underlying logic. Software producing a recommendation a clinician cannot check falls under device regulation.
Risk prediction models estimate the probability of deterioration, readmission or sepsis, and often trigger clinical workflows automatically.
Patient monitoring systems analyze continuous data streams from bedside equipment and generate alerts when patterns suggest deterioration.
Those four categories differ in how directly they touch a clinical decision, and oversight has developed unevenly across them. A tool that drafts a note and a tool that recommends a medication are governed by different rules despite often being sold by the same vendor and installed in the same workflow.
The category boundary is where the current concern sits. Newer ambient systems do more than transcribe: they can identify care gaps, pre-populate orders for a physician to sign, and draft prior authorization requests. Documentation is clerical work. Order generation is closer to medicine, and one analysis argues health systems are treating that shift as a feature upgrade rather than a category change.
The Accountability Question
For documentation tools, the answer is settled and narrow: the clinician who signs the note is responsible for its contents. No ambient AI vendor accepts clinical liability, and physician review before signature is both a legal and professional requirement.
That answer is clear and it may be insufficient in practice. If a tool generates accurate notes nearly all the time, review becomes perfunctory, which is precisely when errors pass through.
For decision support, the picture is less settled. Existing malpractice frameworks were built around human judgment, and allocating responsibility between a clinician who followed a recommendation, a hospital that deployed the tool, and a vendor that built it has not been resolved by case law.
Analysts have noted this becoming a factor in physician employment negotiations, with indemnification for AI-related errors emerging alongside compensation and call schedules as a contract issue. That is a signal worth reading: when a question moves into employment contracts, it usually means the parties involved do not expect regulators to answer it soon.
The Safeguards That Currently Exist
Several layers apply, and they cover the categories unevenly.
FDA regulation covers software meeting the medical device definition, through a total product lifecycle framework extending from premarket review through post-market surveillance. Ambient documentation sits outside it.
Professional standards apply regardless of tool. A clinician remains responsible for the care delivered, and signing a note attests to its accuracy.
HIPAA governs protected health information, and a business associate agreement is required before any tool handling patient data goes live. That requirement is a recurring gap in practice, since staff sometimes use consumer AI tools not covered by any agreement, entering patient details into software that may store or process them outside the protections a hospital contract would impose.
Hospital governance is where most oversight actually occurs, through internal committees, vendor assessment, bias evaluation and local validation. That means quality varies by institution, and a tool validated carefully at one hospital may be deployed without validation at another.
State legislation has begun addressing clinical AI, producing a developing patchwork rather than a uniform standard, so what governs a tool can depend on which state a hospital sits in. MedicalDaily has reported on enforcement action in virtual behavioral health, an adjacent area where oversight followed practice.
The Practical Position for Patients
Patients are already encountering these tools, most often as ambient documentation during a routine visit, sometimes without being told.
Several things are reasonable to ask. Whether AI is being used to document the encounter, since consent requirements vary by state and encounter type. Whether the clinician reviews the note before signing, which they are obligated to do. And how to request a correction if a visit summary contains an error, since patients have a right under HIPAA to request amendment of their records.
Reviewing after-visit summaries and patient portal notes is worthwhile regardless of whether AI was involved. Errors in medication lists, allergies and diagnoses propagate through subsequent care, and a mistake caught by a patient at home is far easier to correct than one discovered months later by another clinician relying on it.
None of this suggests declining care at a facility using these tools. Documentation assistance is intended to return clinician attention to the patient, and evaluations have reported reductions in self-reported burnout and modest decreases in time spent in the electronic record, though clear financial returns have been harder to demonstrate. MedicalDaily has reported on quality gaps in AI note-taking tools used in mental health care specifically.
Key Questions Answered
How widely deployed is clinical AI? Of 6,561 US hospitals sampled, 2,784 used Epic, and 1,744 of those, or 62.6%, had deployed ambient AI documentation by mid-2025. That is roughly two-thirds of Epic hospitals, not of all US hospitals.
Is adoption even across hospitals? No. It was higher among large, nonprofit, metropolitan hospitals with stronger operating margins. Adjusted probabilities were 70.2% for nonprofit hospitals against 28.8% for for-profit ones, with lower uptake in the Midwest than the South.
Are these tools regulated by the FDA? Ambient documentation tools are classified as administrative rather than medical devices and sit outside FDA oversight. Clinical decision support falls under device regulation unless a clinician can independently verify the underlying logic.
Who is responsible if an AI-generated note contains an error? The clinician who signs it. No ambient AI vendor accepts clinical liability, and physician review before signature is legally and professionally required.
Why do the categories matter? Risks differ. A tool that transcribes an encounter poses different problems from one that suggests a diagnosis or pre-populates orders. Newer ambient systems are moving toward order generation, which analysts describe as a category change being treated as a feature upgrade.
What oversight currently exists? FDA regulation for qualifying devices, professional standards holding clinicians responsible, HIPAA requirements including business associate agreements, hospital governance committees, and a developing patchwork of state legislation.
What should patients ask? Whether AI is being used to document the visit, since consent rules vary by state, and how to request a correction if a summary contains an error. Patients have a right under HIPAA to request amendment of their records.
