The Food and Drug Administration has opened a new pilot program that allows a small group of generative AI-powered medical devices to reach patients, including Medicare beneficiaries, before the agency has granted them full marketing authorization. The program, called TEMPO, short for Technology-Enabled Meaningful Patient Outcomes, marks a notable departure from the FDA’s traditional approach of requiring premarket clearance before any device reaches real-world use, and it is being watched closely as an early test case for how regulators intend to handle a wave of generative AI health tools that behave less predictably than earlier software.
Four companies, four different chronic-disease bets
The FDA selected four companies for the pilot in July and August 2026: Limbic and SonderMind, both offering behavioral health technologies, Cadence Solutions, which makes a digital tool for hypertension management, and Dexcom, the glucose-monitoring company, which is contributing a program focused on blood sugar and metabolic health. Cadence’s offering, called HypertensionOS, is designed to help licensed clinicians with prescriptive authority initiate and adjust blood pressure medications for patients diagnosed with Stage 2 hypertension. Dexcom’s Glucose Health Program gives patients and caregivers real-time data and AI-generated insights on nutrition and screens for prediabetes and Type 2 diabetes. Each company’s tool targets a different chronic condition, but all four share the same regulatory shortcut: real-world deployment now, formal evidence review later.
Inside Limbic’s AI voice therapy sessions
The most closely watched participant may be Limbic, a London-based company whose device, called Unpacked, is intended for use within a structured outpatient behavioral health service to deliver cognitive behavioral therapy to Medicare beneficiaries diagnosed with clinically significant depression or anxiety. Unpacked works through an AI voice agent that conducts roughly 20-minute cognitive behavioral therapy sessions over scheduled phone calls, with a human clinician overseeing the process rather than sitting in on every call. That design lets a single supervising clinician support far more patients than one-on-one therapy would normally allow, addressing a well-documented shortage of mental health providers, particularly in the rural and underserved communities where many Medicare beneficiaries live.
Why the FDA is letting evidence follow deployment instead of precede it
TEMPO is explicitly tied to the Medicare ACCESS model, a broader federal experiment in paying for technology that helps beneficiaries manage chronic conditions, and it gives the FDA a mechanism to expand the number of eligible technologies available under that payment model. Rather than waiting for each company to complete the lengthy clinical studies traditionally required before marketing authorization, TEMPO allows real-world data on outcomes, safety and patient usage patterns to be collected while the device is already in use. That data collection, the agency argues, will actually help build the evidence base needed for eventual full authorization, effectively turning early deployment into a live clinical trial rather than a regulatory shortcut that skips evidence altogether.
The case for moving faster
Supporters of the pilot argue that traditional premarket evidence requirements were built for slower-moving hardware and static software, not for generative AI systems whose behavior can shift as underlying models are updated, and that forcing every AI health tool through years of conventional trials before any patient benefits from it needlessly delays access to care, especially for chronic conditions like hypertension and depression where waiting has its own real costs. They also note that Medicare beneficiaries, an aging population disproportionately affected by both cardiovascular disease and untreated mental health conditions, stand to gain the most from tools designed specifically to extend limited clinician time.
The case for caution
Critics, including some patient-safety advocates and clinicians who work directly with vulnerable elderly patients, have raised concerns about deploying an AI voice agent to deliver mental health therapy to a population that includes people with cognitive impairment, hearing difficulties or limited experience with AI systems, before the technology has cleared a full FDA evidentiary bar. They point out that behavioral health interventions carry different risks than a hypertension dosing tool: a therapy session that mishandles a moment of crisis has consequences that are harder to reverse than a delayed medication adjustment. The FDA has said clinician oversight built into Limbic’s model is intended to address exactly that risk, but the pilot’s real-world test will be whether that oversight structure catches problems before they reach patients rather than after.
What happens if the pilot works
The FDA has invited public feedback on its broader approach to generative AI-enabled devices through an open docket, with comments due by mid-October 2026, suggesting the agency views TEMPO as one input into a larger, still-unsettled regulatory framework rather than a permanent fixture. If the four pilot companies can demonstrate strong safety and outcomes data over the coming months, the program could become a template for how a much larger wave of generative AI health tools reaches patients, well before, rather than well after, the conventional evidence-gathering process concludes.
Photo: Tima Miroshnichenko / PEXELS via Pexels
