What Happened
For most of the past decade, the U.S. Food and Drug Administration has held a consistent line on clinical software: if a program makes or informs a decision about a patient's care, it is a medical device, and medical devices require authorization before they reach anyone. This summer the agency began carving out an exception. Under a pilot called TEMPO — Technology-Enabled Meaningful Patient Outcomes for Digital Health Devices — four companies have now been cleared to put generative-AI-driven tools in front of Medicare patients without first obtaining marketing authorization.
The legal mechanism is enforcement discretion, which is less dramatic than it sounds and more consequential than it looks. The FDA is not declaring these products safe or effective. It is stating that, for a narrowly defined set of uses, it does not currently intend to enforce premarket authorization and investigational-device requirements. The agency's public participant list says as much in plain language: the effectiveness of the selected devices, for the intended uses under which they are participating, has not yet been evaluated.
The four participants are an unusually clear cross-section of where applied AI in medicine currently sits. SonderMind is offering an adjunctive smartphone application for adults aged 22 and over being treated for moderate depression or anxiety. Limbic, a London-based company, is offering Unpacked, in which an AI voice agent delivers structured cognitive behavioral therapy over scheduled telephone calls under clinician supervision. Cadence Solutions is offering HypertensionOS, which supports protocol-bound initiation and titration of blood-pressure medication for adults with Stage 2 hypertension, prescribed by a licensed clinician. Dexcom, by far the largest and most established of the four, is offering a glucose health program that combines continuous monitoring with AI-generated guidance and screening support for prediabetes and type 2 diabetes.
In exchange for early access to patients, every participating manufacturer takes on an obligation: collect, monitor and report real-world data on how the device performs for the intended use it entered the pilot with. The expectation is that companies will eventually file for conventional marketing authorization and use that accumulated evidence to support the application. TEMPO, in other words, inverts the usual sequence. Evidence is gathered in the field first, and the regulatory judgment comes afterward.
Why It Matters
What makes TEMPO more than a paperwork curiosity is what it is tethered to. Every participant must offer its device through a specific Medicare program: the Centers for Medicare & Medicaid Services Innovation Center's ACCESS model — Advancing Chronic Care with Effective, Scalable Solutions — a ten-year, outcomes-based payment experiment aimed at chronic disease. The FDA's enforcement discretion applies only inside that boundary. Step outside the ACCESS context and the ordinary rules snap back into place.
That design is doing something quietly clever. Two agencies under the same department have converging problems, and TEMPO lets each solve the other's. The FDA has no settled framework for reviewing a device whose outputs are generated fresh each time rather than fixed at the point of clearance; its existing playbook assumes a system you can fully characterize in advance. CMS, meanwhile, has built a payment model whose rates are deliberately lean, which means it needs a supply of technologies that can deliver care at a cost the model can actually sustain. Ross Harper, Limbic's co-founder and chief executive, put the economics bluntly to Fierce Healthcare, describing the use of new technology to drive outcomes at lower cost as effectively built into the ACCESS payment rate.
The scale involved is what gives the arrangement weight. Roughly 34 million Medicare beneficiaries sit within reach of the model, and CMS has already brought in around 150 digital health companies and provider organizations as ACCESS participants. No AI health product has previously had a route to a patient population of that size without clearing the premarket bar first.
There is a temptation to read this as the FDA lowering its standards, and that reading is too easy. The standards it is setting aside were never doing the work people assume. More than a thousand AI- and machine-learning-enabled devices have been cleared through conventional pathways, yet a recent analysis of 1,357 of them found only 34 linked to a registered clinical trial, and just three evaluated against patient-centered outcomes. Clearance, in practice, has largely certified that a model performs on a dataset, not that a patient ends up better off. TEMPO at least makes outcome reporting a condition of participation. Whether that trade is a good one depends entirely on how seriously the reporting is taken.
Reaction
Response to the pilot has split along a fairly predictable seam. Companies see an opening. Clinicians and health-policy researchers see a precedent, and are less certain they like the shape of it.
Limbic's case for itself is the most fully articulated, partly because its product is the one that will draw the most scrutiny. Harper's framing is that the rules simply predate the technology — existing frameworks, as he told Fierce Healthcare, "were not written for generative AI in healthcare." His argument for safety rests on structure rather than on the model: a clinician stays formally responsible for every patient, reviews session content and progress, receives real-time safety alerts, and can intervene or take over at any point. Limbic says it submitted more than a hundred pages of documentation to get in, including nine peer-reviewed clinical studies. The company has supported roughly 850,000 patients across a dozen U.S. states and the United Kingdom, where it was the first mental health AI product certified as a Class IIa medical device.
The accessibility argument is genuinely strong and easy to overlook. Unpacked runs over an ordinary phone call, which means it works on a landline or a flip phone — no app, no smartphone, no reliable broadband required. For a Medicare population in which older adults are consistently under-represented among people receiving talk therapy, that is not a small detail. Limbic is also waiving copays for ACCESS participants.
The skeptical reading focuses on the exclusions. Unpacked's contraindication list is long and specific: suicidal or homicidal ideation, psychotic features, moderate to severe dementia, eating disorders of any severity, primary substance use disorder, acute physical instability, and more. Those exclusions are the safety architecture, and they are only as reliable as the screening that enforces them. Several U.S. states have been moving in the opposite direction, introducing measures to restrict autonomous AI therapy, and a federal program that permits AI-delivered treatment — even with a clinician formally in the loop — cuts against that current. The tension is unresolved, and TEMPO does not pretend to resolve it.
What's Next
The thing worth watching is not the announcement but the data. Each participant is now generating exactly the kind of evidence the field has been short of, and the four products will not be equally easy to judge.
Cadence's hypertension tool is the cleanest test in the group. Blood pressure is measured, not reported; medication titration follows protocols that can be audited step by step; and the endpoint is unambiguous. If HypertensionOS moves numbers in a Medicare population over a year, that is a readable result. The behavioral health products are harder. Depression and anxiety outcomes rest on self-report instruments, patients in ACCESS are typically receiving other care at the same time, and attributing improvement to any single component of a care package is a well-known statistical headache.
The participant roster is also not fixed. The FDA has said it intends to update the list as further manufacturers are selected, and the composition of the next cohort will say a lot about the agency's appetite. So far every selected device sits in behavioral health or cardio-kidney-metabolic care — areas with established measurement conventions and relatively forgiving time horizons. Whether TEMPO extends into diagnostic or acute-care territory is the question that matters for the pilot's eventual scope.
The real milestone, though, will be conversion. TEMPO is only a meaningful regulatory innovation if at least one participant takes real-world evidence gathered inside the pilot and turns it into a genuine marketing authorization. Until that happens, the program is an interesting experiment in how to regulate, not a precedent anyone can build on.
Closing Thoughts
What makes TEMPO interesting is less what it permits than what it concedes. The premarket framework the FDA has spent decades refining assumes a device that can be fully described before it is used — fixed inputs, fixed logic, fixed outputs, characterized in a submission and frozen at clearance. A system that composes a fresh response to each patient does not fit that description, and no amount of procedural adjustment makes it fit.
Faced with that, the agency had a narrow set of options. It could stretch the existing rules until they covered something they were not built for. It could hold the line and watch generative AI health tools proliferate in the unregulated wellness market, outside its view entirely. Or it could let a small, named, closely watched group operate in the open and learn from what happens. TEMPO is the third choice, and as choices go it is defensible — arguably more honest than the first two.
The risk is structural rather than technological. Enforcement discretion is a decision not to act, and decisions not to act are easier to widen than to withdraw. Four companies, one payment model and a mandatory reporting obligation is a contained experiment. The same posture applied to forty companies, or applied without the reporting requirement attached, would be something else entirely. The guardrail here is not the clinician in the loop or the contraindication list. It is the size of the cohort and the seriousness of the data it produces — and both of those are decisions the FDA will have to keep making, repeatedly, in public.
한글 요약
미국 식품의약국(FDA)이 TEMPO(Technology-Enabled Meaningful Patient Outcomes) 시범 프로그램을 통해 생성형 AI 기반 디지털 헬스 기기 4종을 사전 허가 없이 메디케어 환자에게 제공할 수 있도록 허용했습니다. 선정된 곳은 SonderMind(우울·불안 보조 앱), Limbic(AI 음성 에이전트가 전화로 인지행동치료를 제공하는 Unpacked), Cadence Solutions(2기 고혈압 약물 조절 지원 HypertensionOS), Dexcom(연속혈당측정 기반 대사 관리 프로그램)입니다. FDA는 이 기기들의 유효성을 아직 평가하지 않았다고 명시했으며, 시판 전 허가 요건에 대해 '집행 재량'을 행사하는 방식을 택했습니다.
핵심은 이 허용이 CMS 혁신센터의 ACCESS 모델이라는 특정 메디케어 만성질환 지불 모델 안에서만 유효하다는 점입니다. 약 3,400만 명의 메디케어 수혜자가 대상 범위에 들어가며, 참여 기업은 실제 사용 데이터를 수집·보고할 의무를 집니다. 기존에 허가된 AI 의료기기 1,357개 중 등록된 임상시험과 연결된 것은 34개, 환자 중심 결과로 평가된 것은 3개에 불과하다는 분석을 감안하면, TEMPO는 기준을 낮췄다기보다 평가의 순서를 뒤집은 시도에 가깝습니다.
다만 논란도 분명합니다. Limbic의 Unpacked는 자살 사고, 정신병적 증상, 중등도 이상 치매, 섭식장애 등 광범위한 금기 목록으로 안전성을 확보하는 구조인데, 이는 선별 절차가 얼마나 정확한지에 전적으로 의존합니다. 미국 여러 주가 자율적 AI 심리치료를 제한하는 방향으로 움직이고 있어 연방 차원의 이번 조치와 방향이 엇갈리기도 합니다. 관건은 참여 기업 중 한 곳이라도 이 실사용 데이터를 실제 시판 허가로 전환할 수 있느냐이며, 그 전까지 TEMPO는 선례가 아니라 규제 실험에 머물게 됩니다.
참고: FDA — Participants Selected for TEMPO for Digital Health Devices Pilot · Fierce Healthcare — Limbic selected to FDA pilot · STAT — FDA pilot offers generative AI medical devices a path to patients