The FDA Just Let AI Doctors See Patients Without a License
The FDA is letting four AI medical devices treat real patients before proving they actually work. It's called the TEMPO pilot, and it might be the biggest regulatory experiment in modern healthcare.
The Regulatory Equivalent of "Move Fast and Break Things"
Imagine your doctor prescribes you a new medical device. It monitors your blood pressure, adjusts your care plan, maybe even delivers therapy through an AI voice agent. Sounds futuristic, right? Now imagine the FDA hasn't actually evaluated whether it works yet.
That's not a hypothetical. It's happening right now.
The FDA just greenlit four generative AI medical devices to reach real patients before full marketing authorization, under a brand-new pilot program called TEMPO. If that sounds like the agency is coloring outside the lines, well, it kind of is. And the implications for how we regulate AI in medicine could be massive.
What Is TEMPO, Exactly?
TEMPO stands for a mouthful of bureaucratic acronyms, but the concept is surprisingly simple. Think of it like a provisional driver's license for AI medical devices.
Normally, medical devices have to go through one of three FDA gatekeeping pathways before they can touch a patient: 510(k) clearance (prove your device is similar to one already on the market), De Novo classification (for novel, lower-risk devices), or PMA approval (the full gauntlet for higher-risk devices). Each one requires the company to show safety and effectiveness before launch.
TEMPO flips that script. Selected devices can be used in clinical settings while collecting the evidence that would normally be required upfront. The FDA calls this "enforcement discretion," which is a polite way of saying: "We're choosing not to enforce certain rules for now, as long as you gather data and eventually go through the normal process."
It's not a new approval pathway. It's more like the FDA pressing pause on its own rulebook for a handful of carefully chosen products.
The First Four Through the Door
So who made the cut? The FDA selected four devices across different clinical areas:
Cadence's HypertensionOS is prescription software designed to help clinicians manage medication for adults with Stage 2 hypertension. It's essentially an AI co-pilot for blood pressure management, and Cadence has some genuinely impressive clinical data backing it up. A study published in earlier this year showed a among Medicare patients using the platform. Another study in reported a .
Get tomorrow's biotech intelligence before your competitors.
Join thousands of biotech professionals who start their day with our free, daily briefing.
Mayo Clinic Proceedings: Innovations, Quality & Outcomes
$1,428 decrease in inpatient spending per patient per year
NEJM Catalyst
43% increase in blood pressure control rates
Limbic's Unpacked takes a completely different approach: it's an AI voice agent that delivers cognitive behavioral therapy (the gold-standard talk therapy for depression and anxiety) to Medicare beneficiaries under clinician supervision. An AI therapist, basically, for people who might otherwise wait months for a human one.
Rounding out the group are SonderMind's adjunctive care app for depression and anxiety, along with other selected devices.
Why Now, and Why Does It Matter?
The timing isn't accidental. These four devices are all tied to the CMS Innovation Center's ACCESS model, a Medicare payment program covering four clinical areas: early cardio-kidney-metabolic disease (eCKM), cardio-kidney-metabolic disease (CKM), musculoskeletal conditions, and behavioral health. TEMPO devices must fit one of those buckets, target outpatient care, and be supervised by clinicians.
But the bigger picture is this: the FDA has been wrestling with AI regulation for years, and the traditional playbook doesn't quite work. Conventional devices are static; you build them, test them, and they stay the same. AI devices learn. They evolve. By the time a traditional review is complete, the algorithm might already be three versions ahead.
The agency started addressing this in 2019 with proposals for a "total product lifecycle" approach. In December 2024, it finalized guidance on Predetermined Change Control Plans (PCCPs), which let manufacturers pre-specify certain future updates to their AI without going through a full new review each time. By late 2024, the FDA had authorized over 1,016 AI/ML medical devices through traditional pathways.
TEMPO goes further. Instead of trying to evaluate a moving target in a lab, the FDA is essentially saying: "Let's watch it work in the real world and learn together."
The Elephant in the Exam Room
Let's be honest about what's uncomfortable here. The FDA has explicitly stated that it has not evaluated the effectiveness of these devices for their TEMPO uses. That's not a footnote; it's the whole point of the pilot. These products are reaching patients based on a reasonable expectation of benefit and an assessment that they don't pose serious risks, not on proven results.
For generative AI specifically, the concerns are real. These systems can produce novel, unpredictable outputs. A blood pressure management tool that gives a slightly off recommendation is one thing. An AI therapist that says something clinically inappropriate to a vulnerable patient is another. The FDA knows this, which is why TEMPO requires manufacturers to collect, monitor, and report real-world data throughout the pilot.
The safeguards are meaningful: devices must be used under clinician supervision, they're limited to specific patient populations, and they can't present a "potential for serious risk." Companies are also expected to eventually pursue full FDA authorization using the data they gather. TEMPO participation doesn't guarantee future approval.
Still, there's an inherent tension in deploying AI tools on real patients to find out if they work safely. It's a calculated bet that the potential benefits (faster access, better chronic disease management, filling gaps in mental health care) outweigh the risks of an unproven technology.
What This Means for the AI-in-Medicine Gold Rush
TEMPO is a pilot, and the FDA plans to select up to ten manufacturers in each of the four clinical areas. That means up to 40 devices could eventually participate. For digital health startups, this is an enormous signal: the FDA is willing to create on-ramps for AI tools that can demonstrate clinical promise without requiring years of traditional review first.
Cadence, for its part, is already a unicorn. The company raised a $100 million Series C led by Spark Capital, with a reported valuation of $1.23 billion. It works with more than 20 health systems and over 100,000 active patients. For a company like that, TEMPO isn't just a regulatory shortcut; it's a chance to build a real-world evidence base that could accelerate full FDA authorization and, critically, Medicare reimbursement.
For investors and founders watching from the sidelines, the message is clear: the FDA is experimenting with regulation in real time, and the companies that can generate strong clinical data in live settings will have a significant advantage.
The Bottom Line
The FDA's TEMPO pilot is one of the most significant regulatory experiments in recent memory. It's not reckless (the guardrails are real), but it is unusually bold for an agency that typically moves with the urgency of a glacier. By letting AI devices prove themselves in the wild rather than solely in pre-market review, the FDA is acknowledging something the tech industry figured out years ago: sometimes the best testing environment is the real world.
Whether that philosophy belongs in healthcare is the billion-dollar question. The next year of real-world data from these four devices will start to answer it.
6 min read
The Cancer Therapy That Killed Three People Who Didn't Have Cancer
Three patients with autoimmune diseases died in Novartis CAR-T trials, forcing the company to pause eight studies and the entire field to confront an uncomfortable question: can cancer's most powerful therapy ever be safe enough for diseases that aren't trying to kill you?
