The Opening
For a couple of years, AI in medicine mostly talked: chatbots answering questions, scribes drafting notes for someone else to sign. This week's issue is about what happens when it starts doing.
Autonomous agents built into the EHR are now prepping medication verifications at one of the country's largest health systems and writing transfer summaries at one of its most rural. The FDA has cleared software that adjusts insulin between visits, and major health systems are piloting it. Mayo Clinic is quietly running more than 150 AI models. And nurses, who document more than anyone, are finally getting ambient tools built for flowsheets instead of paragraphs.
None of it looks like the movies, and all of it touches real patients. So this Sunday we're asking the unglamorous questions: who signs off, who reads the audit log, and who pays the token bill.
The machines are on the floor. Let's meet them.
- Troy, Ray, and Ibrahim
The AI agents just clocked in
For two years, agentic AI in medicine lived mostly in keynotes. This month it showed up on the schedule. In a two-part report in late July, Healthcare IT News detailed the autonomous AI agents that Advocate Health and ECU Health quietly deployed in May, built on Epic's no-code Agent Factory platform and running inside the EHR itself.
At Advocate, agents assemble the background a pharmacist needs to verify a medication order, prep infusion charts for chemotherapy patients, and message patients through MyChart before visits. At ECU Health, a nine-hospital rural network in North Carolina, an Epic-built agent writes three-sentence transfer summaries for the transfer center and chart digests for about 160 case managers. ECU says the transfer tool has saved roughly 20 hours of chart review a week, with no hallucinations it knows of.
Why it matters: This is what agentic AI looks like when it touches real patients: not diagnosing, but quietly reshaping who reads what before a human decides. Both systems enforce a hard human-in-the-loop rule. As Advocate's chief data and AI officer Andy Crowder put it: "I don't allow it to order, I don't allow it to diagnose."
The catch: "Zero hallucinations" is a self-reported grade a few months in. Both organizations also warn about data drift, a shortage of AI expertise, and a line item nobody used to budget for: token costs, the per-query LLM fees that can quietly outrun a health system's plans.
Bottom line: The agent era in hospitals will not arrive with a press conference. It arrives one verified medication order and one transfer summary at a time, with a clinician still signing the chart. Watch the audit logs, not the demos.
Autonomous AI agents are now working inside hospital EHRs. Where do you stand?
We're collecting stories. This week, one question.
This week's issue is full of software that acts on its own, so here's what we want to know:
Tell us about the first time an AI did something in your workflow without being asked: the summary it wrote, the reply it drafted, the order it teed up. Did you catch it being brilliant, or catch it being wrong? Either way, you learned something the vendor slide deck never mentions.
We want the real version, not the conference-panel version. Two paragraphs is plenty.
This is open to every corner of medicine: EMS, nursing, pharmacy, techs, registration, environmental services, and administration, all of it. You choose how you're named, whether that's full name, role only, or fully anonymous. We protect patients in every story we run. That is not negotiable and it never will be.
If we run your story, we'll send you a Consult mug as a thank you.
Dr. Sharif Vakili, Co-founder and CEO of UpDoc, Palo Alto
Vakili is a physician who spent years asking a narrow question with big consequences: can software safely adjust a medication between visits? At Stanford, he and co-founder Dr. Ashwin Nayak co-led the MIVA randomized trial, which used a voice assistant to help patients with type 2 diabetes titrate basal insulin under physician-set protocols. That patient-level evidence, not a demo, became the foundation of UpDoc, the clinical AI company he now runs.
Where things stand:
In the Stanford trial, patients using the voice AI reached their optimal insulin dose in a median of 15 days versus more than 56 days with usual care, and 81.3% achieved glycemic control, published in JAMA Network Open
UpDoc's device cleared the FDA in December 2025 (K253281) and was unveiled in June as the first cleared software medical device using patient-facing large language models
Raised an $18M seed round with backers including the American Diabetes Association, Lilly Ventures, and Mayo Clinic
Early deployments underway at Cleveland Clinic, UCSF Health, and Allegheny Health Network, which announced a type 2 diabetes pilot this July
Illustration: The Consult
Why they matter: The loudest version of clinical AI promises an artificial doctor. Vakili's version is deliberately smaller: a bounded agent that executes a physician's plan and escalates the moment a patient drifts outside it. If AI is going to touch treatment itself, this is the shape regulators just accepted, and every "AI doctor" that follows will be measured against it.
Inside the hospital running 150 AI models
CNN went inside Mayo Clinic, which now has more than 150 AI models working across its system. One of them, Record Time, sorts the mountain of outside records that complex patients bring, and internist Dr. Alexander Ryu says it saves him 5 to 30 minutes of prep per visit. The harder question Mayo keeps testing: which of the 150 actually change outcomes.
The FDA clears a chatbot that adjusts insulin
UpDoc's app is the first FDA-cleared medical device built around patient-facing large language models, helping adults with type 2 diabetes titrate basal insulin within limits their doctor sets. As STAT reports, the clearance quietly happened in December: the LLM handles the conversation while deterministic logic computes the dose. The line between chatting and treating just got thinner.
An ER's quiet adviser spots what's hiding
At Israel's Rambam Health Care Campus, a homegrown system called Shaked reads every emergency patient's chart in real time and flags the dangerous cases. Six months in, the hospital reports admissions are about an hour faster, specialist waits are down 25 minutes, and the tool helped catch a life-threatening potassium deficiency in a 13-year-old with muscle pain.
Ambient AI takes on the flowsheet
Ambience Healthcare expanded its nursing suite with Nursing Summary and Ambient Flowsheet Documentation, which turns a nurse's narrated bedside assessment into structured chart entries. Nurses log hundreds of discrete data points per shift, which makes this ambient AI's hardest test yet: structured data instead of prose, with a nurse confirming every value before it hits the record.
Twenty-five nominations for an overwhelmed ER
Earlier this month the Emmy nominations came out, and the most recognized show on television was once again The Pitt, the HBO Max drama that unfolds across a single shift in a Pittsburgh emergency department. Twenty-five nominations, more than any other series. When its first season won best drama last year, the cast dedicated the award to healthcare workers: respect them, protect them.
It's worth noticing what the show is celebrated for. Not miracle saves or futuristic machines. A waiting room that never empties. A charge nurse doing arithmetic on beds that don't exist. Doctors with no time to grieve between patients. Its realism is a realism of load.
Now set that against the issue you just read. Agents prepping infusion charts in the Midwest. A quiet adviser reading every ER chart in Haifa. Software adjusting insulin so a patient doesn't wait months for an appointment. Strip away the branding, and the honest pitch of clinical AI in 2026 is not brilliance. It's load-shedding: take the chart review, the flowsheet, the phone tag, and hand back minutes.
Which puts the two stories on a collision course. Sooner or later, the writers' room will have to decide whether an AI belongs in their fictional ER, because real ERs are deciding right now. And when the cameras find that story, what will read as heroic: the machine that flags the case, or the human who was still paying attention?
Twenty-five nominations say audiences already know medicine is drowning. The tools in this issue are a bet that software can lift some of that weight. The question neither the show nor the industry has answered yet: when the load lifts, will we give that time back to patients, or just fill it with more patients?
Until next Sunday,
For the people keeping medicine human.
