The FDA just cleared an AI that talks to patients. Now it's asking what 'practicing medicine' means.

The most consequential sentence in American health-AI regulation this year wasn't in a law. It was in a device clearance: UpDoc, a clinical AI company, holds FDA clearance for what it describes as the first Software as a Medical Device built around a patient-facing large language model — an app that talks with people with diabetes about their treatment, in natural language, as a regulated medical device. The clearance itself happened quietly in December; the company revealed it in late June, and the implications have been rippling through the clinical AI world ever since.

What was actually cleared — and what wasn't

Precision matters enormously here. The clearance is tightly bound to non-diagnostic tasks: the system operates inside a defined protocol, and the FDA is still drawing a very hard line at autonomous decision-making. No AI has been cleared to independently diagnose or prescribe. What UpDoc's clearance establishes is narrower but still historic — that a conversational model interacting directly with patients can pass the agency's bar at all. Every clinical AI team in the country just got an existence proof and a template.

The same fortnight brought a second signal: Pathway Labs announced FDA clearance on July 1 for EchoNext, an AI tool that flags structural heart disease from a routine 12-lead electrocardiogram — the kind of test performed millions of times a year, now with a second set of eyes that never tires. Different category, same direction: AI moving from back-office triage into the clinical encounter itself.

A plain-English decoder for the regulatory terms doing heavy lifting here, because the pathway is the story. A 510(k) clearance means the FDA agreed a device is substantially equivalent to something already on the market — it's the faster route, and it's how the first products have entered. The de novo pathway is for genuinely novel devices with no predicate: more demanding, slower, but it creates a new category that future products can then reference. That's why the agency engaging with a company on a de novo clinical AI submission matters more than any single product clearance — whoever clears that route first doesn't just get a product to market, they define the template every subsequent conversational clinical AI will be measured against. In regulatory terms, that's the difference between getting through the door and deciding where the door goes.

The existence proof matters more than the product. The question is no longer whether the FDA will clear conversational clinical AI — it's how far the envelope stretches.

The pathway fight coming next

A senior FDA official signaled this month that the agency will soon seek public input on AI that can 'practice medicine' — language that would have been unthinkable from the agency two years ago. The backdrop, per STAT's reporting: one product has already gone through the standard 510(k) pathway, and the agency is engaging with a company pursuing the more stringent de novo route. Meanwhile, a former FDA AI regulator now in industry argues publicly that biopharma is misreading the agency's guidance — treating flexibility as a trap and self-imposing caution the rules don't require. Whether that caution is prudence or paralysis is exactly what the public-comment process will fight over.

For patients, none of this changes tomorrow's appointment. For the industry, the sequence is unmistakable: cleared conversational AI, cleared diagnostic screening, and a regulator openly drafting the rules for what comes after both. The hard line at autonomous decision-making still stands — but the agency just told you where the negotiation starts.