medAI Times
·6:48

Mayo Clinic Now Runs 150 AI Models — Jul 17, 2026

Show notes

Mayo Clinic is now running a hundred and fifty AI models at once.

Run time: 6:48

In today's episode:

  1. Mayo Clinic now runs 150 AI models
  2. Senate probes Medicare Advantage AI care denials
  3. Specialized tool beats GPT, Gemini, Claude
  4. Causal AI: the vasopressor call decides survival
  5. AI drugs pass safety, stall on efficacy
  6. NeuroVFM reads a hospital's own scans
  7. Claude adds self-serve HIPAA setup
  8. Moonshot's Kimi K3 lands at 2.8T params

TL;DR:

  • Mayo's 150 models plus a whistleblower suit show hospital AI's real fight has moved from capability to governance.
  • The Senate and the HHS inspector general are closing in on Medicare Advantage AI denials, and the GOP just blocked a bid to kill the WISeR AI prior-auth pilot.
  • Three peer-reviewed reads this week — OpenEvidence beating frontier LLMs, causal sepsis dosing, and the AI-drug Phase 2 wall — all land on the same point: a benchmark is not a bedside outcome.

Sources cited:

Subscribe: YouTube

medAI Times is for educational and informational purposes only. The content does not constitute medical advice, diagnosis, treatment recommendation, or professional clinical guidance. Consult qualified healthcare professionals and refer to official sources before making clinical, research, regulatory, or business decisions.

Transcript

Auto-generated from the episode audio. Click any timestamp to jump the player there.

Mayo Clinic is now running 150 AI models at once. Welcome to MedAI Times podcast, your daily update on medical AI. Don't forget to like and subscribe. This is your MedAI Times Weekly Wrap for Friday, July 17.

Here's the week in one breath. Mayo Clinic says it now runs about 150 AI models, and a former research leader is suing over how they're governed. Two US senators, one from each party, are hauling UnitedHealth, Humana, and CVS in over AI care denials, and the Senate just blocked an effort to shut down Medicare's AI prior authorization pilot.

A specialized clinical tool beat three frontier chatbots on real doctors' questions. A causal AI model pins septic shock survival on one drug decision. And AI-designed drugs keep sailing through safety trials while stalling on the one test that counts.

Last Friday, we watched OpenAI sign up eight hospital systems. This week, Mayo showed what living with 150 models actually looks like. Start with Mayo. In a CNN feature this week, Mayo Clinic said it has more than 150 AI models in use across the system, and it's building a healthcare-specific model with Microsoft and Scale AI on top of its own patient records and research data.

The showcase tool parses incoming records, tens of millions of pages a year, into chronological summaries a physician can actually search. Dr. Alexander Ryu, who runs innovation in the Department of Medicine, says it saves him between five and 30 minutes of prep per visit.

Dr. Matthew Kallstrom, who leads Mayo's generative AI program, points to catching pancreatic cancer years earlier than usual. So far, so shiny. Then the asterisk. Mayo's former director of research operations, Tracy Tomiko Edo, is suing the clinic, alleging she was retaliated against for raising privacy and oversight concerns about some of these AI systems.

That's the signal here. 150 models is an adoption milestone, not a safety one, and the fight has quietly moved from can it work to who is watching it. Signal for hospital IT and compliance.

Noise for now at the bedside. Washington spent the week on the other end of hospital AI, the money end. Senators Richard Blumenthal, a Democrat, and Josh Hawley, a Republican, sent letters to UnitedHealth, Humana, and CVS demanding records on how they use algorithms to deny rehab and post-acute care.

They point to two June reports from the HHS inspector general, finding those three insurers denied skilled nursing requests at about 12%, with nearly every appealed denial later overturned, and flagging UnitedHealth's NaviHealth tool.

The insurers have until July 28th to answer. And on Thursday, Senate Republicans blocked a Democratic move to end Medicare's Whizzer pilot, the six-year experiment that runs AI on prior authorization for things like knee scopes and nerve stimulators.

Critics say patients now wait four to eight weeks for care that used to take two. The through line when AI touches who gets paid and who gets care, the oversight fight gets loud, fast. Now the evidence tier, which had a busy week.

A team from Harvard, UCSF, and the University of Washington ran a blind head-to-head, 149 practicing physicians grading 620 real point-of-care questions. A specialized clinical tool, Open Evidence, beat CLAWD OPUS 4.8, Gemini 3.1 Pro, and GPT 5.5 on all five quality measures by 25 to 39 percentage points.

Here's the catch worth saying out loud. This is a pre-print, not yet peer-reviewed, and the questions came from Open Evidence's own platform, so treat it as home field advantage. But it's the mirror image of June's benchmark study that crowned the general models, swap exam questions for real ones, and the ranking flips.

From the ICU, a sharper result. Swiss researchers at IDSEA, working with critical care physician Maurizio Cecconi, built a causal AI model, not a black box, for dosing in septic shock, and tested it against real decisions in more than 3,000 patients.

When clinicians deviated from the model's vasopressor recommendation, the odds of dying in hospital were about five and a half times higher. Deviate on fluids, and it barely moved the needle. The lesson isn't trust the machine, it's that the model can tell you which decision actually carries the weight.

That's a pre-print too, with external validation, and it's the kind of cause and effect tool clinicians might actually trust. And the drug discovery reality check, because the hype needs it. A peer-reviewed analysis in drug discovery today, back in the news this week, on a fresh investment wave, found AI-designed molecules clear phase one at roughly 80 to 90%,

about double the historical rate. Then phase two success drops right back to the industry norm, around 40%. Read that plainly. AI is genuinely good at designing safe molecules, fast.

It has not yet moved the needle on whether the drug actually works. Still zero AI-discovered drugs approved. Phase two is where the story gets decided, and it hasn't yet. Two quick ones on the general side.

Anthropic added self-serve HIPAA configuration for its enterprise and API customers, business associate agreement review, and a one-step toggle, which lowers the paperwork wall for health systems that want clawed, near-protected data. And the model race didn't pause.

Moonshot AI dropped KimiK3, a 2.8 trillion parameter open model. The same week, OpenAI's GPT 5.6 family went from government-gated preview to broad release. Spotlight.

This week's Nature Medicine also carried NeuroVFM, a vision foundation model trained on a health system's own routine brain, MRI and CT scans built to improve diagnosis, report drafting, and triage across neuroimaging rather than one narrow task.

It's the quiet countertrend to the giant general chatbots, models grown from a hospital's real messy scans instead of internet text, and it's where a lot of the near-term clinical wins may actually come from. That's the wrap.

Every study, clearance, and letter is linked in the description with the evidence tier marked on each. One question for the clinicians and health tech folks in the room, and I mean it as the real thing to answer below. If your hospital ran 150 AI models, who at your shop would actually be tracking whether each one still works?

Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.