3 SEP 2026 — The United States Food and Drug Administration's TEMPO pilot lets generative-AI devices reach Medicare patients without marketing authorisation, and four have now been accepted. Most coverage frames it as a route to market before clearance. Clearance is not the only thing being set aside. The agency's enforcement discretion also covers the investigational device exemption, the informed consent regulations and institutional review board oversight.
What TEMPO is
The name stands for Technology-Enabled Meaningful Patient Outcomes. The pilot was published in the Federal Register in December 2025, opened for statements of interest on 2 January 2026, and named its first participant in July. Roughly forty manufacturers are expected in total, about ten in each of four clinical areas.
Those areas are early cardio-kidney-metabolic conditions, established cardio-kidney-metabolic disease, musculoskeletal chronic pain, and behavioural health covering depression and anxiety. Participants must be United States manufacturers with a working prototype, a device intended for clinician-supervised outpatient use, and no serious safety risk.
Among the devices now accepted are Cadence, which makes software for hypertension management, and Limbic, which will deliver cognitive behavioural therapy to Medicare beneficiaries with clinically significant depression or anxiety through scheduled telephone conversations with an AI voice agent, with licensed clinicians supervising each patient's care.
Neither a cleared product nor a clinical trial
American medical devices typically reach patients through two pathways. They are cleared or approved for marketing, or they are studied as part of a clinical investigation. The research route requires an investigational device exemption, review board approval and research informed consent.
TEMPO creates a third route. The agency says it may decline to enforce premarket notification, the investigational device exemption requirements, the informed consent regulations at 21 CFR Part 50 and the review board requirements at 21 CFR Part 56.
Read those last two carefully. It would be wrong to say patients receive these devices without consent. Those provisions govern human subjects research. Setting them aside means the deployment is not being run as a clinical investigation, so ordinary clinical consent between a patient and their clinician still applies. What drops away is the oversight apparatus that exists because research subjects bear risk in the service of generating knowledge, and generating knowledge is this pilot's explicit goal.
The containment is a payment model
The boundary is not a regulation. Enforcement discretion applies only while a device is offered to or by participants in the Centers for Medicare and Medicaid Services' ACCESS model, an experiment in paying for technology that helps beneficiaries manage chronic conditions, and the device may not be marketed for the same intended use outside it.
That is an odd arrangement. The population exposed to an unauthorised device is defined by who is enrolled in a Medicare payment experiment, so the boundary of the regulatory exception is drawn by a reimbursement programme rather than by the device's risk profile.
The payment model is also a hard boundary. A manufacturer cannot use TEMPO as a general-purpose bypass, because leaving the payment model means leaving the discretion, and the agency expects each participant to file for marketing authorisation eventually using the data the pilot generates.
What the pilot asks in return
Participants must supply safety and effectiveness data supporting a reasonable expectation of benefit before entry, quality management system documentation held at a registered United States facility, and a risk mitigation plan covering real-world performance collection, monitoring, analysis and reporting.
They must set performance goals with a statistical analysis plan, and report at proposed six-month intervals on adverse events, newly identified risks and progress against timelines.
Watch the six-month reporting interval. For a hypertension tool, two reports a year is a reasonable cadence for spotting a signal. For a voice agent delivering therapy to people with clinically significant depression, the most serious events are precisely the ones nobody can wait six months to hear about, and the safeguard there is the supervising clinician rather than the reporting schedule.
Why generative AI made this necessary
The existing pathways assume a device that is fixed at the moment it is reviewed. A model that generates language does not hold still: its behaviour depends on the conversation it is in, and the same input can produce different outputs on different days.
The agency has a mechanism for change, the predetermined change control plan, which lets a manufacturer specify in advance what modifications it may make without a new submission. That works for a retrained classifier with a defined performance envelope. It fits poorly around a system whose output space is open.
TEMPO is the agency admitting it does not yet know how to evaluate these products from a paper submission, and would rather learn from a deployment it can watch than from one it cannot. That stance is defensible, but it is also a decision to gather the knowledge from patients in the field rather than from subjects in a trial.
What to watch
The pilot's design already contains the test. Every participant is expected to convert to a marketing submission using the data collected, so the question is whether any of them does, and whether the agency clears it.
If devices enter, generate real-world data and emerge with authorisation, TEMPO will have worked as a temporary bridge. If they enter and stay, renewed indefinitely inside a payment model that has become their permanent home, the pilot will have become the pathway rather than the route to one. The four accepted so far are the first cohort to watch.
View author profile → · Editorial policy
