The Food and Drug Administration has begun letting a small group of generative-AI health tools onto the market before they receive standard marketing authorization, betting that real-world data collected from Medicare patients can substitute, at least temporarily, for the years-long clearance process that normally gates new medical devices. The agency’s Technology-Enabled Meaningful Patient Outcomes (TEMPO) for Digital Health Devices Pilot, whose first four participants were announced July 22, 2026, marks one of the most significant regulatory experiments the FDA has run on artificial intelligence to date.
What TEMPO Actually Does
Under TEMPO, a manufacturer can ask the FDA to exercise ‘enforcement discretion’ over certain premarket requirements if its device is offered to participants in the Centers for Medicare and Medicaid Innovation’s ACCESS model, a CMS initiative that reimburses technology aimed at helping beneficiaries manage chronic conditions. In practice, that means a company can put a generative-AI tool in front of real patients, collect outcomes and safety data in the field, and use that evidence to support an eventual, more traditional FDA authorization down the line, rather than waiting on the sidelines until that authorization is granted. The FDA says it plans to select as many as ten manufacturers for the pilot; four have been accepted so far, with two months separating the program’s initial announcement and its first participant list.
Who Made the First Cut
The four organizations chosen span behavioral health, hypertension management and metabolic disease. SonderMind Inc. is testing its SonderMind Adjunctive Care Application, a smartphone-based tool intended to reduce depression and anxiety symptoms in patients 22 and older as an adjunct to therapy. Limbic Inc. is deploying Unpacked, an AI voice agent that delivers cognitive behavioral therapy techniques for depression and anxiety directly to Medicare beneficiaries over the phone. Cadence Solutions Inc. is running HypertensionOS, which supports clinician-supervised initiation and titration of blood-pressure medications for patients with Stage 2 hypertension. And Dexcom Inc., the continuous glucose monitoring company, is contributing its Glucose Health Program, which layers AI-driven insights onto metabolic monitoring to screen for prediabetes and type 2 diabetes.
Why Chronic Disease Is the Testing Ground
All four products fall under what CMS calls early cardio-kidney-metabolic and behavioral-health conditions, the same chronic-disease categories the ACCESS model was built to address. That is not a coincidence. Chronic conditions such as hypertension, depression and prediabetes account for a disproportionate share of Medicare spending, and they are also conditions where continuous, AI-mediated monitoring and coaching could plausibly catch problems between office visits, when a patient’s blood pressure spikes or a mood dips, in ways a twice-a-year appointment cannot. By tying device access to the ACCESS model, the FDA and CMS are effectively trying to answer two questions simultaneously: does the AI tool work, and does earlier, cheaper access to it change what Medicare has to pay for downstream complications.
An Experiment for the Regulator, Too
FDA officials have described TEMPO as a chance for the agency itself to gain hands-on experience regulating generative AI in real-world settings, not just in the controlled conditions of a traditional premarket review. That acknowledgment matters. Generative AI systems, unlike the rule-based algorithms behind most of the more than 1,000 AI and machine-learning devices the FDA has already cleared or approved, can produce open-ended, sometimes unpredictable outputs, whether text, a voice conversation or a medication-titration recommendation. Building a regulatory track record for how those systems behave once they leave the lab and reach actual patients is, by the FDA’s own logic, more valuable than sitting on the sidelines waiting for a fully mature review framework to arrive.
Cautious Optimism Meets Genuine Concern
Supporters of the pilot argue that patients with poorly controlled hypertension or untreated depression are being harmed today by slow access to promising tools, and that a supervised, data-gathering pathway is preferable to either blanket approval delays or an unregulated free-for-all. Cadence and Limbic, in describing their selection, have framed participation as a chance to prove their technology’s value with rigorous outcomes data collected directly from the population, Medicare beneficiaries, that stands to benefit most. Critics, including some patient-safety advocates who have pushed back on earlier FDA moves to loosen AI device oversight, counter that ‘enforcement discretion’ is a euphemism for shipping unapproved medical software to a vulnerable, elderly population before its risks are fully understood, and that generative AI’s tendency toward unpredictable outputs makes chronic-disease management, where a bad medication-titration suggestion has real consequences, an especially high-stakes place to run the experiment.
What Comes Next
The FDA has left room to add as many as six more participants to TEMPO, and the agency is separately seeking public comment on how to regulate generative-AI medical devices more broadly, with submissions due October 19, 2026. How the first four products perform, whether Unpacked’s voice therapy sessions actually move the needle on depression scores, whether HypertensionOS’s medication recommendations hold up under clinician supervision, will likely shape not just the next wave of TEMPO admissions but the broader rulebook the FDA eventually writes for generative AI in medicine. For an agency that has spent years clearing narrow, single-purpose algorithms one at a time, TEMPO is a bet that watching AI work in the real world, on real Medicare patients, is now worth the risk of watching it fail there too.
