What the FDA Actually Cleared When UpDoc Became the First Patient-Facing LLM Medical Device: A Stanford Insulin-Titration Trial, a Drug-Dose-Calculator Predicate, and the Real Shape of Clinical AI
Video Preview
🎧 Podcast episode for paid subscribers only. Also available on Spotify.
To listen to paid episodes in Apple or Spotify, link your Substack subscription via the show settings on those platforms (instructions inside the Substack app under Subscriptions → Podcast).
Abstract
UpDoc, a Palo Alto company spun out of Stanford in 2023, announced on June 25, 2026 what it called the first FDA-cleared clinical AI platform with a patient-facing LLM, alongside $18M in seed funding and deployments at four major health systems. Strip away the press release and the actual regulatory artifact is narrower and far more interesting: a 510(k) (K253281, cleared Dec 23, 2025) for a prescription SaMD that manages basal insulin in adults with type 2 diabetes, cleared as substantially equivalent to a drug-dose calculator (Hygieia d-Nav, K181916), under product code NDC, regulation 21 CFR 868.1890. The clinical evidence base traces to a 32-patient randomized trial (MIVA, NCT05081011) published in JAMA Network Open. This piece walks the gap between the category-defining narrative and the bounded device that actually got cleared, why the boring regulatory path was the smart one, who funded it (Lilly, Mayo, the ADA, Section 32, Polaris, Pear, Cathay, Oxeon), and what the precedent does and does not unlock for everyone trying to build the next one.
Quick hits:
FDA cleared a calculator that can talk, not an autonomous doctor.
The LLM sits outside a deterministic dosing core.
No clinical testing was required for the clearance.
A PCCP was authorized, with limits.
The strategic cap table tells you who thinks this category is real.
Table of Contents
Two stories, one company
It started with thirty-two patients and a voice bot
Read the clearance, not the headline
A chatbot bolted onto a calculator
Why the boring regulatory path was the smart one
Who wrote the checks and what they bought
What the rest of the field should take from this
Two stories, one company
There are two versions of UpDoc circulating right now, and they do not really sound like the same product. The first version is the one that hit the wire this past June, the one with the phrase clinical AI in bold energy, defined as artificial intelligence that delivers care historically requiring a licensed clinician. In that telling, UpDoc arrives as the company that did the hard thing responsibly while everyone else raced to slap the word doctor on a chatbot, and the FDA blessing is the proof. Andy Conrad, the former Verily chief and Section 32 partner, supplied the canonical quote about UpDoc not just redefining the category but establishing what it means to do it responsibly. The CEO, Sharif Vakili, supplied the line about clinical AI being held to the highest standard. The whole thing reads like the opening of a new chapter in medicine.
The second version is the one sitting in the FDA’s public database, and it is a good deal more modest. That version is a prescription software device that helps adults with type 2 diabetes titrate their basal insulin, cleared because it is substantially equivalent to a dosing calculator that has existed for years. Both versions are true at the same time, which is the part worth sitting with. The interesting thing about UpDoc is not that the marketing oversells it, because all health tech marketing oversells. The interesting thing is the specific, deliberate engineering of the gap between what was built and what was claimed, and how cleverly the regulatory strategy threaded a needle that a lot of better-funded companies have face-planted on. For an audience that has watched a decade of digital health companies confuse a press release with a business, the UpDoc clearance is less a milestone and more a case study in how to get a large language model across the FDA finish line without pretending it does something it does not.
