Four companies can now put generative artificial intelligence products in front of American patients without first winning the Food and Drug Administration’s permission to sell them. That is not a loophole. It is the point.
The agency’s TEMPO pilot, short for Technology-Enabled Meaningful Patient Outcomes, reached four participants in late August when the FDA added the London-based mental health company Limbic and Denver-based Sondermind. They join the diabetes technology firm Dexcom and Cadence, which builds a hypertension management tool. All four are exempt from premarket authorisation requirements while they collect real-world data on how their devices actually perform.
A regulator learning by watching
The FDA has spent the better part of two years trying to work out what to do about medical software that generates its own output rather than following a fixed algorithm. A cleared device is supposed to behave predictably. A large language model, by design, does not. In August the agency put out a request for public feedback on how it should regulate generative AI-enabled devices at all, an unusual admission that the existing framework does not obviously fit.
TEMPO is the practical response. Rather than waiting for a finished rulebook, regulators get to watch the technology operate in clinics and living rooms, with patients they can track. Companies get a market. STAT reported on the expansion on 3 September, framing it as the agency buying itself hands-on experience with cutting-edge tools in real-world settings.
There is a payment mechanism attached, which is what makes the pilot more than a research exercise. Participants can offer their products through a new Medicare model that rewards providers for improving patient outcomes in chronic conditions. The FDA built TEMPO in part to feed devices into that programme, known as ACCESS.
What the four are actually building
Limbic’s product is called Unpacked, and it is the most striking of the group. It delivers cognitive behavioural therapy for anxiety and depression over the telephone, through an AI voice agent, with a clinician overseeing the work. Chief executive Ross Harper has said the service is meant to address the shortage of mental health professionals. The company describes itself as the first AI-led mental healthcare firm selected for the pilot.
The guardrails are narrow and explicit. Unpacked is limited to people who speak English and have access to a telephone, and it is contraindicated for several conditions. Sondermind’s entry, a smartphone application intended to reduce anxiety and depression in adults alongside therapy or medication, carries its own exclusions: the device is contraindicated for people experiencing suicidality, manic episodes, psychosis or PTSD.
Cadence takes a different route into the same idea. Its HypertensionOS tool helps physicians and advanced practice registered nurses manage and adjust blood pressure medication for adults with Stage 2 hypertension, a condition where the clinical pathway is well understood and the bottleneck is follow-up rather than diagnosis. Dexcom, the first participant named, brings continuous glucose monitoring into the same outcomes-based payment experiment.
The objection nobody has answered
Patient advocacy groups have already pushed back. During the FDA’s Digital Health Advisory Committee discussions on digital mental health technologies, several called for broader oversight of generative AI rather than less. Their argument is straightforward: a pilot that lets unauthorised software reach vulnerable people is exactly the situation premarket review exists to prevent.
The counterargument is that the alternative is worse. Mental health chatbots and symptom checkers are already widely available, sold as wellness products that sit outside medical device regulation entirely. TEMPO at least puts a defined set of them inside a monitored programme with contraindications, clinician oversight and a data collection requirement. The evidence gathered will help when these companies eventually seek authorisation, and it will help the FDA write rules based on something other than speculation.
Whether that trade holds depends entirely on what the data shows and how quickly anyone acts on it. Four devices is a small sample, and none of the participants has published outcomes yet.
The more interesting question is what happens to the companies that did not get in. If TEMPO becomes the standard route to market for generative AI in medicine, selection into the pilot turns into a competitive advantage handed out by the regulator. The FDA has not said how many more participants it plans to admit, or on what basis. For more coverage of AI in healthcare, visit Mylistingo.
