• Blue Cross says hospital AI coding added $942M without changing care. About 70% of it came from secondary diagnoses, and many of those can be pulled from a single lab value. More below.

• Basalt Health raised $20M to replace the 70-page post-acute fax. At Lifepoint, median referral processing fell from 8.5 minutes to 1.2. It’s heading to 111 markets by year end.

• Oracle put ID.me identity proofing into its EHR. Controlled-substance e-prescribing credentialing is live now. Patient intake is still “rolling out.”

• North Carolina launched a shared AI evaluation network for rural hospitals. UNC and Duke will share “frameworks, vendor evaluations, lessons learned.” The funding is $4.4M over three years.

• 🎧 Podcast: The 229 — Derek De Young on Epic Agent Factory. Epic’s Agent Factory lead says it’s “still in what we’re calling the development preview.” Early adopters open at the end of October, and general availability is a Q1/Q2 goal.

🧭 The Curbside

“OpenAI just released MentalHealthBench. Can I use it to test our behavioral health bot?”

• Short answer: Yes, as a floor. Don’t treat it as a pass.

• What changed: It’s an open benchmark. More than 80 licensed clinicians wrote the rubrics, and it covers adults, teens, caregivers and clinicians at three acuity levels, up to emergencies. The conversations are synthetic, so you don’t need PHI to run it.

• Watchout: The benchmark came from a company that sells a chatbot, and it grades the kind of conversation that company’s product has. 😤 Haters: “So the vendor wrote the exam.” Yes. That’s why it’s useful as a floor and not as proof. Your own ugliest ten conversations are still the real test.

🔬 The Big Thing

If the code went up and the transfusion didn’t, what does the chart actually prove?

The Blue Cross Blue Shield Association published its analysis Thursday. From early 2023 to late 2025, the share of inpatient claims coded as medically complex rose from 37% to 40%, with no matching change in patient acuity.

The breakdown works out to $942 million, and $653 million of it came from 55,000+ extra secondary-diagnosis cases. In major bowel procedures, the share coded at the highest complexity level rose from 10.2% to 22.7%.

The day before, CMS Administrator Mehmet Oz told Oracle’s summit that “short term, AI is going to be inflationary.”

The payer doesn’t have to prove the code was wrong. It only has to show that the treatment didn’t move with the code.

Think about a hemoglobin of 9.8 on post-op day two. That’s a lab value. Whether it’s “acute blood loss anemia” is a clinical judgment. AI is very good at finding the lab value. It can’t tell you whether anyone thought about it.

That’s the whole fight. Blue Cross’s example is more anemia diagnoses without more transfusions. BCBSA’s Razia Hashmi, MD, put it bluntly:

“If it was worth coding, there should have been something done.” — Razia Hashmi, MD, BCBSA

Hospitals tell the mirror-image story. HCA’s CFO says hospitals are “behind the payers.” NYC Health + Hospitals’ revenue chief calls it a “rock ‘em, sock ‘em robot” fight that nobody wins.

😤 “Documenting a real diagnosis isn’t upcoding.” True. Even BCBSA’s Luke Chalker calls some of it “the stuff that hospitals should bill for.” The fight is over conditions that nobody treated. The methodology also has a limit: it’s claims-only, and BCBSA says so.

😤 “Payers use AI to deny. Why can’t hospitals use it to code?” They can. Both sides are running the same arms race.

😤 “This is a payer PR document.” It is. It’s also a preview of the audit letter.

❓ Every one of these disputes turns on one question: did the diagnosis change what the team did? I think there’s a product in linking each coded condition to the order, med or note that shows it was acted on. It wouldn’t be a coding tool or a CDI tool. It would be a “treatment-concordance” layer. Would anyone buy that before the first payer clawback, or only after?

📡 Builder’s Radar

The payer’s rulebook is becoming an API

Ensemble Health Partners is embedding Penelope Health’s policy engine into its revenue cycle platform. (Both are Thoreau companies: Thoreau agreed to buy Ensemble in June and put $100M into Penelope on Sept. 18.) Penelope turns payer coverage rules into structured data: 15,000+ procedure and drug codes, and policies covering 200M+ insured Americans.

The point is to check medical necessity before the encounter, not after the denial.

Once the rules are machine-readable, the fight moves upstream, to the moment the order gets placed.

SNOMED has 350,000 concepts. Your chest pain still maps to R07.9.

John Lee, MD, argues that coded terminologies squeeze out clinical meaning. His example is a patient with known coronary disease whose new burning chest pain gets the same “unspecified” code as the vague discomfort she had before. LLMs can read the nuance the code throws away.

Read that next to the Big Thing. A machine that reads nuance well will also find more to code.

🔮 Where this lands: The next fight won’t be over structured versus unstructured data. It’ll be over who gets to cite the free text in a dispute, and by 2027 payers will be asking for the note, not just the code.

North Carolina is trying to build a shared evaluation bench

NC CHAIN links UNC Health, Duke Health, the Sheps Center and Duke-Margolis to help rural and critical access hospitals choose and govern AI tools.

$4.4M over three years works out to about $1.5M a year for a whole state. That’s a few staff, not an evaluation lab.

😤 “It’ll produce a framework and a webinar series.” It might. The test is whether a published vendor evaluation ever comes out of it. If one does, it’ll be the most-read document in rural health IT.

⚡ Quick hits

• Anthropic and OpenEvidence are offering free clinical AI in about 100 countries. It’s a localized version of OpenEvidence for clinicians in low- and middle-income countries, including Uganda, Haiti and Mongolia. The US product doesn’t change, and neither company disclosed financial terms.

• Memorial Health in Savannah is holding off on about 50 Epic AI features and plans to turn them on a few at a time, with close monitoring. Having a feature available isn’t the same as deciding to use it.

🎙️ From the Pods

🎙️ Stanford AIMI Grand Rounds — Maya Yiadom, MD, “AI in the Loop”

Yiadom’s move is to treat the model’s threshold like shared decision-making, except the health system is the patient. Simulate the options on site data, show the trade-offs, then have governance sign off on the parameter, the acceptable and unacceptable trade-offs, and the conditions that trigger a reassessment. In her words, “it’s not just the algorithm alone.”

💡 Builder take: Write the threshold authorization before you write the alert.

🔇 Speaker Blindspot: Generalizing from a resource-rich site. Her governance table has about ten roles in it, from ECG techs to health-equity experts. She allows it could be just a few stakeholders and says automating governance lowers the bar, but that’s still a bench a critical access hospital doesn’t have, which is why NC CHAIN exists.

🎙️ The Nocturnists — “The Lost Aura” with John Lantos, MD

Lantos gave a chatbot his own cardiac data and asked for the pessimistic take, then the optimistic one. His point is that patients can now “design our own doctors” and get as many opinions as they want.

💡 Builder take: A “show me the other read” button may do more for shared decision-making than another summary.

🔇 Speaker Blindspot: An illusion of independence. Two framings from the same model on the same data aren’t two opinions. They’re one opinion wearing two outfits.

💡 BTW: John Lantos was asking this question almost thirty years before ChatGPT. His 1997 book was titled Do We Still Need Doctors?

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Franciscan’s CTO: “I have to wait until the bill comes in”. Cloud spend is forecastable. Agent spend isn’t, even at a system with a full-time cost analyst. The Big Thing is below.

• An FDA panel backed GRAIL’s Galleri 7–2, with one abstention. It is the first multi-cancer blood test to reach a panel, but effectiveness passed only 6–4.

• Allina’s hospital doctors reached a tentative first contract after a four-day strike and more than 66 sessions of bargaining over pay, sick leave and AI. The AI terms haven’t been made public.

• Pelago made its clinical AI the front door for all behavioral health. Sona now assesses every member first, under licensed clinical supervision.

• 🎧 Podcast: The 229 Podcast — “They Thought They Had 25 AI Tools. They Had 250”. Christian Boucher of Island says one customer’s count of AI tools went up tenfold once every browser was instrumented.

🔬 The Big Thing

If you can’t price an agent before you turn it on, who can?

Chuck Christian, CTO of Franciscan Health, created a Cloud Cost Analyst role in 2023. Every Wednesday at 7:30 a.m. he gets a report on spend to date against budget.

Moving one set of workloads onto right-sized servers will save about $50,000 a month, he said.

AI broke that system. Microsoft had to build a separate portal just to show token usage. “If we stand up another agent… I can’t predict,” he said. “I have to wait until the bill comes in before I know how much it’s going to cost.”

An organization that can forecast its servers to the month can’t forecast an agent at all.

The same day, Mass General Brigham CFO Niyum Gandhi said: “We are investing in it net.” When ambient documentation came back as a seven-figure annual line item, he didn’t wait for an ROI case.

Meanwhile a Bain × KLAS survey of 303 executives found many organizations have no formal ROI threshold for AI. Those that do most often target 3 to 3.9 times the investment.

So the buyers want a return on something none of them can price.

Here’s the part that belongs to us. What an agent costs isn’t really a property of the model. It depends on the work:

• how long the note is;

• how many chart pulls the prior auth needs;

• how often the agent retries when the payer portal times out.

A hospitalist can tell you that a CHF readmission note runs far longer than a cellulitis note. The finance team can’t.

The meter finance needs is built from workflow facts only a clinician has.

😤 “Vendors will just bundle it into the license.” As far as Christian knows, Epic already does. He doesn’t assume that will last.

😤 “This is a FinOps problem. Hire a FinOps person.” Franciscan did, in 2023. That analyst works from compute, storage and memory, all of which behave predictably. Tokens don’t. A spreadsheet can’t forecast a variable nobody has written down.

❓ Is there a product that prices a clinical workflow before anything is switched on: tokens per encounter type, the way an actuary prices a member month? Christian also can’t meter the outside agents pulling his data. I think those are the same product, and I can’t quite see its shape yet.

🧪 Try the interactives — both built with real CMS data:

A — The Stay Sets the Bill — 520 Medicare DRGs, one dot each. Plan an inpatient AI agent on the typical stay and the bill arrives about 31% higher, because stays are skewed.

B — Price the Agent — An inpatient AI agent cost simulator over all 520 Medicare DRGs. Set runs per day, context size, chart growth, retries and token prices, and see which diagnoses drive the bill.

📡 Builder’s Radar

Oracle announced five agents, and a disclaimer that they might not ship

Oracle Health’s revenue-cycle release covers prior auth, documentation integrity, charge capture, pro-fee coding and appeal packets. All five are “planned for general availability in the coming months.”

The fine print says the release “is not a commitment to deliver any material, code, or functionality.”

😤 “Point solutions are dead, then.” Only if the buyer is fine with a date that isn’t a commitment. Most buyers aren’t.

Can a 6–4 vote support a screening label?

Galleri reads cell-free DNA methylation patterns with a machine-learning classifier that claims to detect more than 50 cancer types. The panel voted 10–0 on safety and 6–4 on effectiveness.

In an advisory vote, safety and effectiveness get split. Your marketing page doesn’t do that.

A classifier can be safe and still not clearly work. That’s the gap four panelists voted on.

💡 80/20: Write your tool’s intended-use sentence before the feature list: who it’s for, what it finds and what someone does next.

Pelago’s AI now sees every behavioral health member before a human does

Sona is the front door for substance use, mental health and behavioral addiction care. It all runs on one contract and one member record, with pricing tied to outcomes. Pelago says every assessment happens “under licensed clinical supervision.”

Pelago ran two observational studies covering more than 9,000 conversations. Nearly half of participants improved by at least five points on the PHQ-9 or GAD-7.

Read who that “nearly half” is: participants who started in the clinical range and completed repeat assessments.

😤 “Completers always look good.” Yes. The number to ask for is how many never completed a second assessment.

⚡ Quick hits

Claude Opus 5.5 launched with per-token prices cut about 20%; Anthropic says typical workloads cost about 40% less than on Opus 5. That per-task figure is the vendor’s, and your task isn’t theirs, so rerun your own workloads before you rebudget.

65% of hospital nurses surveyed feel watched or tracked by AI in a Black Book survey of 202 nursing professionals, and 63% say AI added tasks without removing old ones. Tools that watch people and tools that help them aren’t the same product.

🎙️ From the Pods

🎙️ Digital Health Leaders (CHIME) — “From Automation to Autonomy: Reinvesting Capacity to Build Healthcare’s AI Workforce”

Innovaccer’s Sandy Gupta told CHIME CEO Russ Branzell that a prior auth that took 50 minutes now takes under three. He said the bigger gain is faster surgical scheduling, not a smaller team.

💡 Builder take: Measure throughput on the other side of the task, not just minutes saved.

🔇 Speaker Blindspot: Motivated reasoning. A platform vendor said the hype is “worth it” while citing an 85–90% pilot failure rate, and never counted what the failed pilots cost.

🎙️ Relentless Health Value — “Overpaying for Infusions Means Overpaying for Stop-Loss Coverage”

Jake Velie and Keith Hartman, RPh, explained how overpaying for the same infusion at a hospital outpatient department (site-of-care gaps can run 10x, host Stacey Richter noted) pushes up an employer’s stop-loss renewal. The place to intervene is pre-certification, with a hand-off to the prescriber instead of a denial fax.

💡 Builder take: Site-of-care routing at pre-cert is an unglamorous workflow tool that saves real money.

🔇 Speaker Blindspot: Anecdotal generalization. “Trend went flat” rests on one conversation, in an episode sponsored by a site-of-care vendor.

💡 BTW: Chuck Christian is a woodworker. He explains architecture review this way: “if you put too big a screw in a piece of wood, what happens? It cracks.” (healthsystemCIO)

📅 Upcoming: healthsystemCIO’s “Managing AI Token Budgets” webinar, Sept. 29. Details →

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Epic paused product development to harden security. Judy Faulkner says the company is focused on safeguarding its system against attacks. Expanded below.

• Mass General Brigham isn’t letting AI agents act autonomously across platforms. Claude Cowork is in a pilot there, but only ring-fenced inside MGB’s own environment.

  😤 Haters: “That’s just a big system being slow.” It’s a big system that lost track of its own devices during an outage and decided not to repeat that.

• Heidi raised $340M, and only $100M of it is equity. The rest is a $240M go-to-market investment from General Catalyst, pointed at “supervised action.”

• Abridge won a seat on a $775.72M VA contract. It’s a multiple-award ceiling, not a check.

• CMS posted preliminary 2027 lab rates, about 16% lower on average. The money math is in Money Plumbing below.

• 🎧 Podcast: Radio Advisory — “Inside Sutter Health’s playbook for scaling AI”. Sutter CMIO Veena Jones says the scale signal was pilot clinicians asking “when are my colleagues going to get this?”, and patient and family advisors graded the AI in-basket drafts.

🧭 The Curbside

“Patients are booking with us through Gemini and Amazon’s health AI now. What changed?”

• Short answer: Your calendar became an API that someone else’s agent can transact against.

• What changed: Zocdoc opened its Care Access Network to any provider on Tuesday. One connection makes a provider bookable across Amazon Health AI, Gemini, Yelp, Healthgrades and payer directories. There’s no booking fee for existing patients or for searches by provider name.

• Builder read: Your visit-type rules and intake questions now have to work for an agent, not just a person at the front desk. If a new-patient slot needs a referral, someone has to encode that.

  🔮 My bet: by mid-2027, “agent-bookable” shows up as a line item in practice RFPs, next to “EHR-integrated.”

🔬 The Big Thing

What does it mean when the EHR stops shipping features to go look for its own holes?

Epic has paused product development while it makes sure its systems can’t be exploited. Judy Faulkner framed it as preventing attacks.

The full story is paywalled. Scope and duration aren’t public yet, and I’m not going to guess at them.

The context is. Since spring, health system CIOs have been bracing for AI models that shrink the gap between finding a flaw and exploiting it. Corewell Health’s CISO, Scott Dresen, put it plainly: the new models don’t create new risks, they shorten the time between discovering a flaw and exploiting it.

The scarce resource at the biggest EHR company this fall isn’t engineers. It’s confidence that what’s already shipped can’t be turned against a hospital.

😤 “This is PR after the AI-feature backlash.” Maybe. But pausing a roadmap is about the most expensive press release a software company can write.

😤 “This slows everyone down.” Yes. For now, that’s the point.

🧪 Try the interactives — both built with real CMS lab-pricing data:

A — How a Lab Price Is Made — 17,372 private-payer prices for one lipid panel, and the single Medicare number they collapse into. The volume-weighted median, drawn.

B — The Unfinished Cut — 942 lab tests, one dot each: the private-payer weighted median from the first PAMA round (2016 data) against what Medicare pays in 2026. Most of the gap CMS is about to close was already there.

📡 Builder’s Radar

Only $100 million of Heidi’s $340 million is equity

Heidi’s $100M Series C values it at $900M. The other $240M is a growth investment from General Catalyst’s Customer Value Fund to pay for go-to-market. The stated goal is moving “from documentation into supervised action.”

Heidi says it will also invest in clinical evidence, quality management and regulatory submissions. That’s what agentic features in healthcare require, and it’s rarer to see it in a funding announcement than it should be.

“Supervised” is the word doing the most work in that release.

🔮 My bet: within 12 months an ambient vendor files an FDA submission for an agentic feature. Scribes that act start to look like devices.

Abridge won a hunting license, not a contract

The VA contract is a five-year, multiple-award IDIQ with a $775.72M ceiling shared by all eligible vendors. Each medical center or regional VA network places its own task orders. Abridge came in through a distribution partner and already runs at 75+ VA medical centers.

The moat wasn’t the model. It was running on VistA/CPRS and the Oracle-based Federal EHR at the same time, in the middle of a migration.

😤 “$775 million is a big win.” It’s a ceiling shared with competitors. The revenue is a hundred local decisions.

⚡ Quick hits

Benji Graham (FHIR architect, HL7 certified educator) says his FHIR Department forum is near 900 members and made the agenda at an HL7 working group meeting. A grassroots community for the people who actually write the resources (via LinkedIn).

💰 Money Plumbing

The weighted median: how a reporting sample becomes a national lab price

Under PAMA, labs report what private payers paid them for each test, and Medicare pays the volume-weighted median. The preliminary 2027 file covers 1,528 of 1,947 codes from 6,411 reporting labs and averages about 16% lower.

The spread is the lesson. Genomic sequencing drops 23% and molecular pathology 22%, but proprietary lab analyses (PLA codes) fall only 2.4%. Cuts are capped at 15% a year for 2027–2029, which compounds to roughly 38.6% over three years.

If your CDS nudges an order or your RPM program depends on a panel, those unit economics just moved. Comments go by email within 30 days of posting, and final rates publish in November.

🎙️ From the Pods

🎙️ The Gist — “Examining physician payments tied to GLP-1 drugs”

Washington Post data reporter Lydia Sidhom walks through more than $270 million that Lilly and Novo paid prescribers from 2018 to 2025, reaching roughly 270,000 of them. The limit she names is the useful part: Open Payments shows who got paid, not what they prescribed.

💡 Builder take: Open Payments is free and public. Joining it to prescribing data is a real weekend project, and nobody hands it to you pre-joined.

🔇 Speaker Blindspot: The bias blind spot. The highest-paid prescribers describe the money as peer education, and self-assessment is exactly where influence research says people are least reliable.

💡 BTW: Heidi CEO Thomas Kelly trained in vascular surgery, and before Heidi he ran a company tutoring medical students in clinical skills. He was a teacher before he was a founder.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Epic’s AI flood is forcing health systems to triage it themselves — Seattle Children’s stopped piloting and started sorting. The evaluation queue is the new integration queue.

• Inova’s CIO says a data moat is a losing strategy — Matthew Kull’s ChatGPT Health traffic into Epic has skyrocketed in two months; plan for 10X API fees.

  🔮 My bet: within 12 months a major EHR publishes a per-call price list for outside model traffic, and “who pays for the tokens” becomes a standard contract line.

• Microsoft retired the SMART on FHIR proxy yesterday — nine days ahead of Azure API for FHIR itself. If something of yours broke Monday, this is why.

• 🎧 Podcast: Health Tech Nerds Radio — “The Grand Roundup” — an AI assistant app is now messaging telehealth vendors on users’ behalf and negotiating discounts.

🧭 The Curbside

“The imaging vendor says FDA-cleared for 17 findings. Does it diagnose?”

• Short answer: No. a2z-Abdo-Triage is computer-aided triage — it flags suspected positives so radiology can reorder the worklist. Ten findings cleared September 21 on top of seven.

• 😤 Haters: “Seventeen findings is seventeen chances to be wrong.” It’s also seventeen chances to move the hemoperitoneum ahead of the surveillance chest.

🔬 The Big Thing

Who on your org chart is qualified to say no to 300 AI features?

Clara Lin, MD, VP and chief medical information officer at Seattle Children’s, said it plainly at Becker’s IT + Revenue Cycle conference: Epic is releasing “like 100, 200, 300 … just so many different AI things all at once.”

Her team stopped piloting on its own timeline. They sort the list by where a return is likeliest to show up.

The same week, Matthew Kull at Inova said walling off your data is a losing strategy now: patients ask a chatbot who’s available nearby, and the system that can’t answer inside that conversation loses the appointment. His ChatGPT Health API volume has skyrocketed in two months.

Both are describing the same shortage from opposite ends: demand for clinical judgment is arriving faster than anyone can supply it, and the supply is a person, not a product [unless Jev can do it.]

On this week’s Newsday, Bill Russell said CIOs at the front of ambient adoption are telling him EHR request volumes went “up 9000%” — because ambient vendors widened their scope from drafting a note to pulling the longitudinal record and checking rev cycle along the way.

Treat it as a rumor, because it is one: no system, no baseline, no window. But the mechanism is arithmetic — a tool that goes from one write to N reads produces a number like that whether or not anything is wrong.

😤 “This is just governance. We have a committee.” You have a committee that meets monthly and a vendor that ships weekly. That’s not governance, that’s a lagging indicator with catering.

😤 “Epic shipping features is a good problem.” It is. It’s also how point solutions die — not outcompeted, just never evaluated. If the review board has room for six items a quarter and the platform takes four, you didn’t lose on merit.

🧪 Try the interactives — both built from the FDA’s own 510(k) clearance database:

A — Watch the Ladder Collapse — 382 AI-pathway FDA clearances sort themselves onto five rungs of what the software is allowed to say. Four in five land on the two rungs whose safety case is “a clinician will check it.”

B — The Permission Ladder — 382 AI-pathway 510(k) clearances, one dot each, sorted onto the five rungs of what the software is legally allowed to say. The ladder collapses as you climb: 223 clearances at “measure,” 12 at “diagnose.”

📡 Builder’s Radar

FHIR grew a place to put the model card

At HL7 Connectathon 43 on September 19–20, the AI Transparency on FHIR track tested the implementation guide, which balloted in January and is now in reconciliation. It carries model cards inside FHIR resources — CHAI’s and Hugging Face’s both, with no format mandated — and Provenance tagging at the element level, not just the resource.

Element-level is the part to notice: “this sentence came from a model, that one from a human” becomes expressible in the chart itself.

If your tool writes into a record, the standard for saying so in machine-readable form is in ballot reconciliation right now, not two years out.

💡 80/20: Emit one Provenance resource for one thing your tool generates. If you can’t describe your own output in the standard’s vocabulary, that’s the gap.

Clinicians drew the line, and it’s in a strange place

Sarah Neville reported in the Financial Times on September 20, under the headline “Medical AI has a proof problem”, that clinicians accept AI in imaging and diagnostics while pushing back hard on everything past it — documentation, treatment recommendations, patient communication — because the performance data is thin.

The numbers underneath come from a March survey of 355 US doctors and nurses: 74% named deskilling, 74% hallucinations, 72% advertiser-driven bias, and only 27% reported any awareness of AI governance where they work.

The deskilling number is the operational one. Every monitoring plan I’ve read names “clinician review” as the mitigation. Three-quarters of clinicians just said the tool degrades the skill that review depends on. That’s not an ethics finding, it’s a broken control.

😤 “Doctors always say this and adopt it anyway.” They do. The interesting part isn’t the resistance, it’s where it stopped — at the boundary between finding something and deciding something.

The gray zone nobody wants to define

Ben Schwartz, MD, an orthopedic surgeon, writing September 20 on the models he uses: “nearly all of them tend to err on the side of caution and slight catastrophizing.”

That’s a specific, testable failure mode — and the opposite of the one everybody evaluates for. We build harnesses to catch a model missing badness. Almost nobody builds one to catch it manufacturing badness, which in an ED is the difference between a discharge and a $4,000 workup.

⚡ Quick hits

Oura’s $2.2B IPO isn’t a raise — 50 million shares at $40–$44, about $14.1B at the top of the range, but the company is selling only 13.5 million of them and Forerunner is unloading its entire stake. Seventy-three percent of the offering is a cash-out, not capital.

Addiction treatment keeps consolidating — seven closed transactions in H1 2026 against eighteen in H1 2025 — the prior consolidation wave having left fewer sizable targets to buy.

🎙️ From the Pods

🎙️ Health Tech Nerds Radio — “The Grand Roundup” (Sept 21)

A user told an AI assistant to get a restaurant reservation; it hit the Resy API roughly 200 times an hour until it got one, and the user got temporarily banned. Telehealth vendors now see the same assistants messaging them on patients’ behalf.

💡 Builder take: Your rate limits were designed for humans who get bored. Check your per-identity request ceiling and ask what happens when the requester is an agent on a five-minute cron.

🔇 Speaker Blindspot: Appeal to an interested authority. The bearish OpenEvidence read leans on a Doximity earnings-call claim that a major system is banning the product — from a direct competitor, relayed as “I think they said,” unnamed and unchallenged.

🎙️ The Heart of Healthcare — “Can AI Finish What The Quality Revolution Started?” with Dr. Azita Hamedani, Chief Clinical Officer at Bunkerhill Health (Sept 21)

Hamedani walks an agentic incidental-findings workflow end to end: chest CT shows coronary calcification, the agent reasons across contraindications and existing specialists, then auto-generates the cardiology referral, the patient letter and the PCP note.

💡 Builder take: The product isn’t detection. It’s the closure record — proof a flagged finding reached a human with authority to act.

🔇 Speaker Blindspot: False dichotomy, said out loud in four words — “it’s all upside.” Unfollowed findings versus AI follow-up deletes the third option, that some incidentals shouldn’t be chased. Also unasked across 37 minutes: who’s liable when the auto-generated referral is wrong.

💡 BTW

💡 BTW: Dr. Azita Hamedani, Bunkerhill’s new chief clinical officer, spent fifteen years turning emergency medicine at the University of Wisconsin from a division inside Internal Medicine into its own academic department, doubling clinical volume and raising a $10 million endowment along the way. She was its founding chair.

💺 Builder Seats

Just ones I found on LinkedIn that look interesting. No sponsorship. Use at your own risk.

Strategic Medical Director, Inpatient — Cohere Health · Remote (US) · $285K–$305K
Prior auth from inside the company automating it — a rare seat where clinical judgment is the product spec.
🔗 Apply on LinkedIn

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• The readiness score has no outcome data behind it — Eric Topol went looking for evidence tying wearable HRV or readiness scores to health outcomes and came back with none. Expanded below.

• PeaceHealth is cutting 150 IT jobs, its cybersecurity staff among them — day-to-day IT moves to Tech Mahindra and The HCI Group. The system says it still owns security. The people who did it are leaving anyway.

• Two-thirds of Angle Health’s $600M round is a cash-out — $200M of new money, $400M buying existing shares. And the secondary priced at $2.5B, not the $2.7B in the headline. That’s a discount for the past.

• Eight states turned rural health money into venture funds — Louisiana is writing $250K–$3M checks and has fielded 200-plus pitches. Delaware, Georgia, Massachusetts, Nebraska, South Carolina, Virginia and West Virginia are building the same thing.

• 🎧 Podcast: Thursday’s Tradeoffs — “911 Fails People in Mental Health Crises” — roughly 60% of 911 call centers are overseen by law enforcement, so the response is decided by the org chart before anyone picks up.

🔬 The Big Thing

Is the readiness score a measurement, or a product feature wearing one’s clothes?

Eric Topol read everything published on wearable heart rate variability and readiness scores, expecting to find a few good studies. He found none.

This ten days after Apple added a 0–10 Readiness score and increased HRV reporting 24-fold. Well over 100 million American adults now wear one of these sensors.

Here’s the gap. Every association between low HRV and mortality was measured from an ECG.

Your watch measures pulse rate variability off an optical sensor — related, not the same, and one study in over 900 adults called it “an invalid surrogate for HRV.”

The evidence everyone cites was collected with a different instrument than the one on your wrist.

It gets thinner. The most-cited concordance study ran on 13 healthy adults.

Apple reports SDNN while nearly everyone else reports RMSSD, so a 90 on one device means nothing about a 90 on another. A UK Biobank analysis of 46,000 people found genetically predicted HRV didn’t track mortality.

The readiness scores on top are proprietary composites — one review of 14 found HRV carrying 86% of the weight — that a company can rewrite overnight without telling you.

💬 Standout Quote

“If they believed and invested in the products they are selling, we’d not be in this position of not knowing.” — Eric Topol, on the wearable companies

[These are companies with the money and the users to run the trial. Not running it is itself a finding.]

😤 “This is anti-technology hand-wringing.” From the loudest advocate for consumer biosensors of the last twenty years? No. He’s separating what earned its evidence — activity, atrial fibrillation detection, resting heart rate — from what didn’t.

😤 “Patients find these numbers useful anyway.” Some do, and the long-COVID readers in his comments make that case better than I could. Same device, same conditions, trend over weeks — that’s the honest ceiling.

❓ So what sits between “raw sensor stream” and “proprietary score” — something that tells a clinician which metric a device is actually reporting, on what firmware, with what population gaps?

🧪 Try the interactives — both built from the real ClinicalTrials.gov registry:

A — He Found None. So Does the Registry. — 3,801 registered trials measure heart rate variability. Watch the field empty as you apply the four conditions that would make one of them evidence. Six survive. None used a consumer wearable.

B — Find the Trial That Doesn’t Exist — every registered trial that measures heart rate variability, all 3,801, plotted by size and year. Filter for a consumer wearable, a hard health outcome and posted results, and the field goes to zero.

📡 Builder’s Radar

PeaceHealth just cut the people who would have evaluated you

PeaceHealth is eliminating about 150 IT jobs across its hospitals in Alaska, Oregon and Washington — application developers, analysts, network and database administrators, data specialists, and cybersecurity workers — as day-to-day IT moves to Tech Mahindra and its US subsidiary The HCI Group on a five-year contract. The 150 is the Washington WARN figure, 119 of them in Clark County; Becker’s reports fewer than 250 affected systemwide.

The company is explicit that it keeps technology decisions and responsibility for cybersecurity and patient data in-house.

Hold both of those at once. The accountability stays. The staff go.

Find out whether the person evaluating your product is still a health system employee. That’s an afternoon of email and it’s the difference between a Q4 pipeline and a Q4 mystery.

😤 “So it’s harder to sell into them now.” Maybe the opposite. A contracted evaluator has a written standard instead of a personality, and for six months they have no favorites.

Ultra-shorts

An ambient scribe that never opens a socket — WrenScribe, pre-launch for Mac and iPad, runs Whisper plus a small local model entirely on-device: transcript, note, suggested ICD-10/CPT, PIN-encrypted, optional read-only Medplum connector. ⚠️ Their table lists “BAA / vendor risk review: Not applicable” — a design claim from a product that hasn’t shipped, not a compliance determination.

Michael Hobbs, MD (pediatrician, AI-in-medicine educator) posted part three of “Anatomy of an AI Clinical Error” on Friday — pediatric vignettes held constant across 10 frontier models (N=300) and 7 commercial clinical AI platforms (N=126), toggling only the framing: “what do you think?” versus “draft the discharge instructions.” The prompt posture changed the errors.

🎙️ From the Pods

🎙️ Tradeoffs — “911 Fails People in Mental Health Crises. Here’s a Plan to Fix It.” (Sep 17)

Rebecca Neusteter, who runs the Health Lab at the University of Chicago Harris School of Public Policy, names three structural problems with 911. The first is a data model: about 60% of call centers are overseen by law enforcement, and the caller gets a three-item menu — police, fire, ambulance.

💡 Builder take: your dropdown is your policy. Any triage taxonomy you ship decides most outcomes before a human reads the case, and “other” is where the hard patients go to disappear.

🔇 Speaker Blindspot: False dichotomy — three structural problems named, only the funding one priced. Rewriting the call-taker’s option list is software, not the $10-billion-plus she quotes for the technology upgrade.

🎙️ The 229 Podcast — “56% Fewer IV Harm Events. 2,000 Hours Back to the Bedside” (Sep 17)

Jennifer Jones walks through IV pump interoperability work that cut adverse infusion drug events 56% and returned roughly 2,000 nursing hours. Her argument, with Kay Burke and Rachael Hill: informatics belongs in the room before the technology is chosen.

💡 Builder take: if the buyer’s informatics lead meets you for the first time at the demo, you’re already inside the failure mode they describe.

🔇 Speaker Blindspot: Survivorship bias — three nursing informatics leaders who got a seat at the strategy table, explaining that a seat at the strategy table works. The interesting episode is the one with the CNIO who asked and got told no.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

In the spring of 2000, Scott Pioli — the Patriots’ personnel director — put two photographs on his desk.

One was a sixth-round quarterback out of Michigan. Taken 199th, after six other quarterbacks and most of the league’s attention.

The other was Dave Stachelski, a tight end out of Boise State. Pioli had taken him earlier, because he liked him more.

He kept both where he had to look at them. Not as a trophy. As evidence that he could easily have gotten it wrong.

The quarterback was Tom Brady.

David Halberstam recorded both photographs in The Education of a Coach.

Years later, Seth Wickersham asked Brady the same question twice, a decade apart: how close had he come to never making it.

The first time, Brady was in his twenties. He answered easily. If a different team had taken him one pick earlier, he said, he’d probably be out of the league. He could still see that version. It was close enough to touch.

The second time, he was in his late thirties. He leaned back, stared at the ceiling, and got defensive. He’d done the work. The others hadn’t.

Nothing about the draft had changed.

The story had won.

That’s the part nobody names. Success doesn’t just reward you. It deletes every version of events where it didn’t happen, until the way it went looks like the only way it could have.

We have a word for this. Outcome bias.

And medicine did something strange about it: we built a room. A group convenes, on a schedule, to ask whether a decision was good — including when the patient walked out fine.

We’re not good in that room. Matthew Syed wrote a book about how much better aviation is at this, and he’s right.

But we built the room. Almost nobody else did.

You know why we bothered.

The patient you sent home who did fine — you don’t know whether you were right or lucky.

The first-pass intubation that went clean — you don’t know whether that was your hands or that airway.

The consult you didn’t call, on the guy who turned out to be nothing — that one’s still open too.

You were trained to keep the second picture on the desk. That training is rarer than you think.

TL;DR

Success erases the alternatives.

Keep the picture of the one that didn’t work.

What makes a clinician valuable on a software project is assumed to be clinical knowledge. It isn’t. Knowledge is the cheapest thing in the building now.

What’s expensive is someone professionally trained not to let the outcome grade the decision. Who looks at a dashboard of green and asks what the green is made of. Who has sat in that room and watched a good outcome get taken apart anyway.

You didn’t learn that in a course. You learned it with experience.

So put the other picture on your desk.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

New here? Every Saturday I take the week’s best idea in each of three buckets and cut it down to something you can read between patients.

Might take you 5 min to read.

I read the week so you don’t have to.

Let’s go.

— Kevin

1. TIME

The pilot was never the expensive part.

This week, on a stage in Chicago, Lee Schwamm, MD — a stroke neurologist who is now senior vice president and chief digital health officer at Yale New Haven Health — described the rule his system runs on:

“We’re not evaluating launching a pilot unless the executive sponsor commits to paying for it if it’s successful.”

Read that twice.

It isn’t a funding rule. It’s a time rule.

A pilot costs a quarter. A deployment costs years — retraining, version bumps, the day the vendor changes an endpoint, the workaround nobody tells you about.

Nobody signs up for that in advance. So it lands by accident, on whoever is still standing.

Here’s the number that makes it real. Drew Smith, chief data and analytics officer at ChristianaCare, said at a panel this week that at a prior employer, only two of ten deployed AI tools were still in use a year later.

Two. Not two that failed. Two that survived.

And the discipline at the other end of that arc, from Kay Burke, who runs nursing informatics at UCSF Health — the question she asks of every click already in the chart: if we were designing this today, would we still require this?

Every one of those clicks was somebody’s good idea. That’s the whole problem.

Schwamm’s other line is the one I keep hearing: some of the most important interventions we can do in medicine don’t save money.

True. And they still need an owner in year two.

So before you build it, answer the boring question. Who is maintaining this in 2029, and what are they being paid to do instead?

A win.

(Words: 280. A minute, and a bit.)

“We’re not evaluating launching a pilot unless the executive sponsor commits to paying for it if it’s successful.” — Lee Schwamm, MD

Dig deeper: the ROI panel write-up. Read it for the maintenance burden nobody prices at procurement.

2. MONEY

A hundred claims became twelve million dollars.

This week the HHS Office of Inspector General posted an audit of Methodist Hospital.

Auditors pulled 100 claims worth $1,426,020. Seventy-three were fine. Twenty-seven weren’t. The actual overpayment they found: $256,926.

The refund they demanded: $12.4 million.

That’s extrapolation: you sample, find an error rate, apply it to the whole population. The same week, an audit of HumanaChoice sampled 220 enrollee-years, found 178 with unsupported diagnosis codes and $669,237 in real overpayment — and asked for $130.9 million.

Don’t read the multiplier as aggression. UnitedHealthcare of Wisconsin drew a larger sampled overpayment — $722,280 — and a smaller ask, $46.9 million. The ratio tracks how many lives sit behind the sample.

Now the part that should keep you up. Extrapolation is a bet on consistency. It works because a hospital that got 27 claims wrong in a random sample is a hospital with a habit. OIG said as much at Methodist: the root cause was that the hospital didn’t follow its own written billing policies.

Software doesn’t have a habit. Software has a defect. Perfectly reproduced, on every claim, at 3 a.m., without variance.

Which makes your book of business more extrapolable than it has ever been.

Mohanish Anand, therapeutic-area head for global trial management, asked six frontier models to pick a random number between 1 and 30.

All six said 17.

Confident. Plausible. Identical. That is what a fleet of agreeing models looks like from the outside, and exactly what an auditor’s sample catches.

So pull your own hundred. Grade them the way somebody hostile would. Multiply by your volume.

If the number frightens you, congratulations — you just found the product.

A win.

(Words: 280. A minute, and a bit.)

An audit doesn’t punish your worst day. It punishes your most repeatable one.

Dig deeper: the Methodist report. Skip to the sampling methodology appendix — that’s the part that scales.

Also the interactive at clinicians.dev

3. MOJO

She kept the badge.

Before she took the job, Margaret Lozovatsky, MD told her CEO there was one thing she wanted.

She wanted to keep practicing.

She’s a pediatrician. She’s also, since May, the chief digital information officer of Premier Health in Dayton — the person who owns technology strategy for a five-hospital system, in from the AMA by way of Novant and Cedars-Sinai.

And she still does a couple of shifts a month, with newborns.

She isn’t alone — the physician-executive seat is having a year. But the interesting part isn’t the résumé. It’s the negotiation.

She asked for the shifts before she had the title. She was protecting something she already knew she’d need.

Here’s what I think it is.

Every other signal reaching a chief digital officer has passed through three people who want you pleased. The dashboard is curated. The governance committee is diplomatic. The vendor’s QBR is a work of art.

The 2 a.m. newborn is not diplomatic.

A shift is the last unmediated feedback channel an executive has. Not nostalgia, not optics — instrumentation. The one place the thing you shipped has to work with nobody managing your impression of it.

Most of us drift the other way. The further we get from the room where it’s used, the more we call that promotion.

So the question this weekend isn’t what you’re building.

It’s: when did you last use it yourself, on a shift, with a real person waiting on you?

A win.

(Words: 243. A minute.)

You cannot manage your way back into a room you stopped standing in.

Dig deeper: the Becker’s piece. Read her paragraph as a job negotiation, not a bio.

🔭 THE ADJACENT POSSIBLE

One thing from outside health that’s about to be reimagined inside it.

On Tuesday a company called TypeSafe came out of stealth with $40M and a model that refuses to chat.

[You can test it out on openrouter right now]

Jev does three things: picks from a list you supply, scores on a scale you define, or returns the probability that a yes/no statement is true. Typed output, calibrated confidence, 70 to 500 milliseconds end to end. Founder Diogo Almeida worked on the research behind ChatGPT. This is the opposite bet. Input runs $0.042 per million tokens. Output is free.

Outside health it’s a router. Inside health it’s every gate we’ve been too poor to build: does this note support the code we billed? Should a human see it first?

Go back to the money section. An audit multiplies a hundred claims because sampling was all anyone could afford.

When checking costs a thousandth of generating, you stop sampling your own work. You check all of it.

One caveat: a confidence score is not a review. “0.91” tells a regulator nothing. The gate still needs a reason attached.

But the arithmetic of verification moved this week, and that doesn’t happen often.

Dig deeper: the model’s pricing page — it’s the clearest statement of what the thing actually is. Note that the calibrated confidence, not the speed, is the part you’d build on.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, almost completely AI generated and may contain errors)

• Memorial Sloan Kettering put OncoKB inside OpenEvidence — and put OpenEvidence inside its own Epic workflow. The curated annotation, not the model, is what changed hands.

• Connecticut wrote human review into its state employee health plan — Anthem, Cigna and Aetna agreed. No AI-only denials or down-coding for 270,000+ members from January 1.

• IDEXX bought CoVetAI — a diagnostics company bought an ambient scribe. Vet medicine reaches the answer about a year before we do.

• 25 health systems spent this week building clinical-judgment agents with Epic — NEC flags, refill protocol checks, med-history conflicts. Not scheduling anymore.

• 🎧 Podcast: NEJM AI Grand Rounds — “Beyond the Hype: Dr. Xiao Liu on Evaluating Medical AI” — under 5% of published AI-versus-clinician imaging studies met the methodological bar medicine uses everywhere else. The reporting guidelines ask you to disclose what you did, not to have done it perfectly.

🔮 My bet: within 18 months, at least three more academic centers license a curated clinical knowledge asset — a tumor board catalog, an antibiogram, a transplant eligibility ruleset — into someone else’s tool. The AI deal everyone signs in 2027 isn’t for compute. It’s for annotations.

🧭 The Curbside

“Can I point a cheap fast model at synthetic billing data instead of a frontier model?”

• Short answer: Maybe — if you treat the confidence score as the product, not the answer.

• What changed: TypeSafe released Jev on Tuesday — a model that doesn’t generate text at all. You hand it the allowed options; it returns one plus a calibrated confidence score ultrafast. Guideline adherence anyone?

• Builder read / Watchout: It’s waitlist early access, not GA, and nobody is using PHI with it.

🔬 The Big Thing

What exactly did Memorial Sloan Kettering just sell?

On Wednesday, MSK and OpenEvidence announced a two-way integration: OpenEvidence goes into MSK’s Epic workflow, and OncoKB — MSK’s precision oncology knowledge base — goes into OpenEvidence for every clinician outside MSK.

OncoKB isn’t a model and it isn’t patient data. It’s a catalog of what each genomic alteration actually means — oncogenic effect, level of evidence, the therapies attached — curated by people and granted partial recognition status by the FDA under its public human genetic variant database program.

The model was never the moat. The curation was.

Every frontier model can read a tumor sequencing report. Almost none can tell you whether a Level 3B alteration is worth a trial referral this month. That answer is an annotation — made by a human, dated, revised when the evidence moves.

Now the uncomfortable half. A Nature Medicine pilot published last month put an LLM decision-support system into a tertiary ED. No adverse events; reviewers rated 99 of 100 sampled outputs clinically appropriate.

Clinician use fell from 68% to 30% over four weeks — hardest as shift workload rose.

Accuracy wasn’t the failure mode. Attention was.

So the question for MSK isn’t whether OncoKB-in-the-chart is correct. It’s whether an oncologist on their eleventh patient of the afternoon opens it.

😤 “This is a licensing deal with a press release attached.” Mostly, yes. But look at what’s being licensed. Not a dataset, not a model, not a patent — a maintained opinion, with a version and a date on it. I haven’t seen that priced before.

😤 “Curation doesn’t scale.” Correct. That’s the point.

😤 “Great, so I should go build a knowledge base.” You already have one. It’s the stuff you re-explain to residents every July that exists nowhere in writing.

❓ If the annotation is the asset, who owns it when it goes stale inside somebody else’s product? OncoKB has a curation cadence and a team. The reference tool has a cache.

🧐 Interactives

One Lab, One Code

Memorial Sloan Kettering just licensed OncoKB — the annotations its curators have hand-maintained for a decade — into OpenEvidence, for every clinician outside MSK. The interpretation layer is now everywhere. Here is the layer it interprets: 330 molecular diagnostic codes in Medicare Part B, $2.80 billion, and 117 of them billed by exactly one entity in the entire country.

https://www.clinicians.dev/interactives/2026-09-17-one-lab-one-code-b

📡 Builder’s Radar

Connecticut turned “a human reviewed it” into a contract term.

The state employee plan’s carriers — Anthem, Cigna and Aetna — agreed to stop using AI as the sole basis for a denial or down-code, covering 270,000+ members from January 1. Officials will ask lawmakers in January to extend the same terms to every state-regulated plan.

A purchaser got there before a regulator did.

A diagnostics company just told you what an ambient scribe is worth.

IDEXX acquired CoVetAI on Wednesday — ambient listening and workflow software used by tens of thousands of veterinary professionals. Terms undisclosed.

The buyer wasn’t another AI company. It was the incumbent that already owns the lab, the instrument and the practice management system, buying the layer where a spoken sentence becomes an order.

Ambient documentation may not be a product. It may be an interface that attaches to whoever owns the workflow underneath it. The copilot categories have been collapsing into each other all year; this is what the bottom of that collapse looks like.

😤 “It’s dogs. This doesn’t transfer.” Veterinary medicine has no HIPAA, no prior auth and a two-year product cycle. It’s the fast-forward button, and it just landed on “the scribe is a feature.”

25 health systems spent this week building agents that make clinical judgments.

At Epic’s second Agent Factory Build-a-Thon, teams moved past the first wave’s transfer-and-scheduling agents. One reads notes for early signs of necrotizing enterocolitis in premature infants. Another checks refill requests against protocol and marks which are safe to approve. A third flags conflicts between a med history and new admission orders. Training is October; broad availability 2027.

The governance policy your system wrote in the spring almost certainly does not cover this class of agent.

❓ A NEC-flagging agent is a screening test with no published operating characteristics. What’s the sensitivity, and who computed it? I don’t know whether anyone in the room asked, and it seems like the only question that matters.

⚡ Quick hits

The fax outlived the FHIR rollout. MRO acquired Vyne Medical on Wednesday — paper, voice, fax and image intake for 800+ hospitals — pushing the combined footprint past 2,500. Interoperability is still mostly a document-processing business.

Sword Health is buying Headspace for up to $300M, all cash. 100M+ users, 20,000 employer accounts, sold at a ceiling of roughly a tenth of its peak private valuation. Distribution beat brand equity, and it wasn’t close.

🎙️ From the Pods

🎙️ NEJM AI Grand Rounds — “Beyond the Hype: Dr. Xiao Liu on Evaluating Medical AI”

CONSORT-AI and SPIRIT-AI are transparency requirements, not methodology mandates. You don’t need a perfect study to make a defensible claim — you need to say exactly what you did.

💡 Builder take: Write your eval’s limitations section first. If you can’t describe your test set honestly in five sentences, the eval isn’t done.

🔇 Speaker Blindspot: Is-ought conflation. Liu predicts clinicians will gain gatekeeping control of medical AI because of their proximity to the patient — a normative claim wearing a prediction’s clothes. In every other industry, gatekeeping power followed capital and platform control, not proximity to the work.

🎙️ Relentless Health Value — “How Stark Law, Stipends, and Noncompetes Drive Hospital Consolidation”

Eric Bricker, MD on site-neutral payment arbitrage: the same echocardiogram reimburses far higher inside a hospital than in an independent practice. If you sell into independent groups, you aren’t losing to a better product. You’re losing to a price difference your product can’t touch.

💡 Builder take: Before sizing an independent-practice market, check whether the procedure you’re built around is one the arbitrage is actively pulling into hospital ownership.

🔇 Speaker Blindspot: Circular causation. Bricker blames stagnant professional fees on balkanized physician lobbying versus a unified hospital lobby — but hospitals now employ most physicians, so both are downstream of the same arbitrage he just described. One cause presented as two.

🤝 Selling It

The utilization ledger — what actually decides your renewal

Boston Medical Center pays $200 to $600 per provider for an ambient AI license, and their chief data officer says that price tag doesn’t buy a permanent seat — falling utilization means retraining, or redeploying the licenses somewhere they’ll get used. Mayo Clinic’s Edwina Bhaskaran, on the same panel, asked the version with teeth: what do we decommission, and what have we failed at? Your first sale was an outcome pitch. Your renewal is a spreadsheet.

So build the spreadsheet yourself, from quarter one. Four columns: licenses provisioned versus actively used; utilization by cohort, never one blended number; the one outcome metric your buyer is personally measured on, with their baseline; and what you’re changing next quarter because of the first three.

Then say the part nobody says: “What would make you cancel. Tell me now and I’ll build the early-warning signal into this page.”

💡 Try this: Ask your pilot champion which of their numbers appears on their own performance review. If they can’t answer in one sentence, you don’t have a renewal — you have a trial.

💡 BTW: Daniel Nadler, OpenEvidence’s founder, published a collection of imagined ancient love poems — Lacunae: 100 Imagined Ancient Love Poems — with Farrar, Straus and Giroux in 2016, and is the youngest person ever elected to the board of the Academy of American Poets.

📅 Upcoming: MassAITC’s free “Understanding and Assisting Older Adults with AI” webinar, Tuesday September 22, 4pm ET.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Medicare is working out what an “AI physician” is worth — the internal number under discussion is 60% to 80% of the human rate for the same service. Full treatment below.

• CMS opened its ACCESS model to heart failure, COPD, substance use disorder and tobacco cessation — 160 participating organizations, payment tied to measured outcomes, new tracks spring 2027. Three of four people on Medicare already qualify for a track.

  🔮 My bet: within 18 months somebody builds the referral layer between the tracks, not another track. Heart failure plus COPD plus nicotine dependence is three vendor relationships today, and nobody owns the seam.

• The WISeR prior-authorization pilot went live on a rushed timeline and delayed care — per newly released documents, covering skin substitutes and epidural injections. Same agency, same month, opposite direction of travel.

• HCA crossed 72 hospitals on Meditech Expanse — two of 15 US divisions fully converted. If your integration roadmap assumes two EHRs, add a third.

• 🎧 Podcast: This Week Health’s Solution Showcase — “22 Agents in 12 Months,” with UTMB — Bunkerhill CEO Nish Khandwala’s pricing model: a fixed pool of AI credits the hospital buys at the start of each quarter and allocates across use cases itself, so a pilot that stops working can be defunded.

🧭 The Curbside

“Can I run any genuinely useful agents on my own laptop yet?”

• Short answer: For defined tasks, yes — and localhost is the one sandbox that needs no BAA.

• Evidence: A tuning walkthrough published Sept. 15: Gemma 4 12B fits in ~6.7GB of VRAM, and Ollama’s default 4,000-token context is nowhere near the ~64,000 agentic work wants. Simon P. Couch, a senior engineer at Posit: “A few months ago, any LLM that I could run on my Macbook scored 0% on an agentic coding eval I put together. [The April] Qwen 3.5 and Gemma 4 releases both scored 90%.”

• Watchout: Local still lags badly on long multi-step chains. But the gap has closed enough that there’s no excuse for prototyping.

  [BTW check out Bionic from LM Studio if you are in to running local agents (not just models) …. https://lmstudio.ai/ … gemma 4 and qwen 3.8 27B work ok, but kinda slow obv]

🔬 The Big Thing

What is a doctor’s judgment worth? Medicare’s opening bid is 60 cents on the dollar.

Federal health officials are weighing a new Medicare payment category that could reimburse companies for AI software that supports diagnosis or care. Nothing has been proposed. It’s a conversation.

Inside the administration, per a person involved in the discussions, the live question is whether AI physicians run by technology companies should be paid 60% to 80% of what a human earns for the same service.

A benefit category isn’t a press release. It’s a gravity well. Staffing, liability, build-vs-buy, what a residency is worth — all of it bends toward the rate once the rate exists.

The evidence underneath is what makes the timing strange. FDA has cleared more than 1,500 devices containing AI, and essentially all of them do one discrete thing — find the clot, flag the nodule. There’s no framework built for a chatbot that prescribes.

And Sam Ashoo, MD went through the numbers this week. In a published cohort of 1,357 of those cleared AI/ML devices, 2.5% had a prospective trial, and 0.2% — three of them — were studied against mortality, morbidity or readmission. Clear on substantial equivalence, deploy at scale, then set a price. Validation is the step nobody’s standing on.

Which is where a clinician who builds becomes the scarce input. Nobody at CMS is going to define what an AI internist may do when a patient on apixaban turns up with a hemoglobin of 6.4 and insists she feels fine. Someone who’s been in that room writes it down, then builds the test that proves the model stays inside it.

😤 “60% is a discount. It should be 5% — software has no marginal cost.” That’s the CFO’s argument and I think it eventually wins. But the first number anchors the next decade.

😤 “This is vaporware. No rule has been proposed.” Correct — a discussion, not an NPRM. I’m telling you anyway, because payment categories are where American healthcare’s real decisions get made, and by the time there’s a comment period the shape is set. Say “an AI can’t be a physician” to a CFO looking at a 1.4% operating margin and watch what happens.

❓ If Medicare pays an AI 70% of a physician rate, who signs the note? I keep circling the attestation layer — some artifact recording what the agent did, what a human reviewed, and what neither of them caught. I think there’s a product in there. I can’t see its shape yet.

🧪 Try the interactives — both built from the real CMS public-use files:

A — Two Targets, One Policy — WISeR sends one AI prior-authorization tool at skin substitutes and epidural injections; one chart of 13,943 Medicare billers shows they are not the same problem.

B — Sixty Cents on the Dollar — 60% to 80% of what, exactly? The 300 largest Medicare Part B service codes, $88.9 billion, with a rate dial and an eligibility switch. Drag it and watch which specialty pays.

📡 Builder’s Radar

The founder handed back the keys, and the company got two CEOs.

Terry Myerson passed the torch at Truveta after six and a half years; co-founders Jay Nanduri and Ryan Ahern take over as co-CEOs of the health-system-owned data collective.

🔮 My bet: provider-owned data collectives become the most contested asset class in health tech within two years. If Medicare pays for AI care, somebody has to hold the evidence that it works — and the systems that own the data are the only ones who can price it.

Quick hits

Evvy raised a $40M Series B to turn 100,000 patients’ worth of at-home microbiome testing into a research engine. Priyanka Jain sold a swab; what she built was a dataset in a domain that had almost none.

💡 80/20: find the question your specialty has no dataset for — the one where everyone reaches for expert opinion because the evidence isn’t there. That gap is the asset.

Karl Swanson, DO is leaving ambient scribing — his title at Doctronic is Clinical Context Engineer, which may be the most honest job title in health tech (via LinkedIn).

Aaron Neiderhiser of Tuva Health shipped a refreshed open-source guide to claims data; the Tuva Project is still the fastest path from raw claims to something queryable (via LinkedIn).

🎙️ From the Pods

🎙️ This Week Health, Solution Showcase — “22 Agents in 12 Months” with UTMB

Peter McCaffrey, MD, UTMB’s chief digital and AI officer, says the entry point wasn’t an agent — it was an FDA-cleared module flagging incidental coronary artery calcification. The flag created a brand-new problem: who reviews it, who acts, do you hire a navigator. The agents were hired to do work the AI invented.

💡 Builder take: every detection tool you ship generates downstream labor. Budget for it in the pitch, or the pilot dies on a people cost nobody scoped.

🔇 Speaker Blindspot: False dichotomy. Build-it-yourself versus buy-the-platform, with the middle option — an internal platform team — never priced. Worth remembering the vendor is paying for the episode. “22 agents” is a vanity numerator too: no retirements, no error rates, no outcomes.

🎙️ Lifers with Christina Farr — “Miriam Paramore on becoming a caregiver”

Paramore has 42 years in health tech and spent this episode as a daughter. Her father sat $500 over the Medicaid line, which cost him the dual coverage that would have paid transport for a 45-minute one-way dialysis run the family could not always cover.

🔇 Speaker Blindspot: Straw man. She rebuts AI by noting it can’t lift a dialysis bag or drive to the clinic — a claim nobody makes. The real ones, including the missing care-quarterback role she describes so vividly herself, go untested.

💡 BTW: Sean Kelly, MD — the emergency physician whose name is on this week’s research about agents acting on clinicians’ behalf — still practices, teaches part-time at Harvard Medical School, and does first aid at Fenway Park.

💺 Builder Seats

[These are just ones I found on LinkedIn that look interesting, no sponsorship or anything. Use at your own risk but look legit.]

Associate Partner, Clinical Technology Innovation — Chartis · Remote
Advisory side — you’d see twenty health systems’ AI roadmaps instead of one.
🔗 Apply on LinkedIn

Staff Clinical Product Specialist (MD) — SmarterDx · Remote
Physician product seat where the whole surface is documentation accuracy.
🔗 Apply on LinkedIn

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• The scribe moved into the mid-revenue cycle — Abridge’s pre-bill review compares coded diagnoses and DRGs against the documentation before the claim goes out. Full treatment below.

• The VA is scaling its AI scribe past primary care — behavioral health, PM&R and specialists, heading toward an enterprise contract. It did not provide note-quality data. A VHA-funded study already did.

• CMS put $16M into rural Kansas for “emerging technology” — fourteen rural providers buying AI diagnostics and remote monitoring, plus $780K for transport coordination.

• Advocate’s CIO says AI will raise total healthcare costs — Bobbie Byrne, MD, pushing back on stage against Penn Medicine’s Mitchell Schnall, MD, PhD, who’d just reported a ~30% efficiency gain in pharmacy prior authorizations.

  😤 Haters: “efficiency that costs more isn’t efficiency.” It is if the constraint was never dollars — it was people, and you can’t hire them.

• Oracle shipped its nurse agent — voice charting, search and acute summaries in the inpatient chart. More below on the one word that matters.

• 🎧 Podcast: Radio Advisory — “312: Are declining hospital volumes the new normal?” — visits fell in three of 2026’s first four months (−3% Jan, −3% Mar, −7% Apr) while same-site patient days fell 10–17% every month. Size a market off 2025 and you’re sizing off a watermark year.

🔬 The Big Thing

The scribe just started reading the bill. That’s not a feature — it’s a different product.

Abridge shipped a pre-bill review tool for CDI and coding teams on Monday. It compares the final coded diagnoses and the assigned inpatient DRG against what the documentation supports, before the claim drops, and shows the evidence behind every discrepancy it flags.

It runs on the same platform and clinical record as the ambient scribe. Reid Health, already a documentation customer, is the named launch reference — it hasn’t rolled this out to CDI yet.

It doesn’t change codes; the CDI team decides whether to hold, adjust or release. That’s the right design, and a tell about what the legal team thinks this is.

If ambient capture materially increases the documented complexity of an encounter, then the ambient scribe was never a documentation product. It has been a severity-coding product this whole time, and nobody priced it that way.

That changes who you are to a regulator. Minutes saved is a satisfaction metric. A DRG that moved because the machine heard something the physician never wrote down is a payment integrity question — audited years later.

😤 “This is just CDI software with a microphone.” CDI software works from the note. This works from the captured encounter, upstream of the note — and a Washington trial court held last month that a patient has no right to the ambient recording, treating it as administrative rather than part of the record. The audit trail thins out exactly where the money is.

😤 “Every vendor expands into revenue cycle eventually.” Sure. Nice to be right about something.

😤 “Nobody’s coding fraudulently here.” I don’t think anyone is, and that’s the point. A tool that reliably surfaces documentable severity raises severity on average, across every case, with no individual decision you could point to. The honest version and the bad version look identical in review.

❓ If the vendor holds the captured encounter, the draft note, the final note, and now the coded claim — who is positioned to build the independent check on any of it?

📡 Builder’s Radar

The largest clinical AI deployment in the country is expanding, and the one published quality number came out against it

The VA has already taken ambient scribes past primary care into behavioral health, PM&R and medical and surgical specialties, heading toward an enterprise contract this year on Abridge and Knowtex. Inpatient is explicitly not on the roadmap.

The department did not provide note-quality data. But a VHA-funded study already found AI-generated notes scored lower than human ones across five standardized cases.

So the number exists, it’s unflattering, and the rollout is proceeding anyway. That’s not a secret — it’s a judgment about which evidence counts.

😤 “Clinicians love it and burnout is real.” Both true, and neither is a note-quality measurement. Adoption and satisfaction go up whether or not the notes are good, which is exactly why they’re the metrics that got published.

“Discrete” is the word to watch in Oracle’s nurse release

Oracle announced US availability of its Clinical AI Agent for nurses in the inpatient chart — voice search, acute nursing summaries, and “voice-enabled discrete charting,” which the release glosses as structured documentation in near real time. BayCare is the named early adopter.

Physician ambient turns speech into prose, where a wrong word reads oddly. Nursing ambient turns speech into structured fields, where a wrong number is a wrong number in a flowsheet somebody titrates off.

⚡ Quick hits

Monday’s $16M in rural Kansas Emerging Technology Grants will land at fourteen providers with no CIO and no vendor-management function.

💡 80/20: find your own state’s Rural Health Transformation awards — usually posted by the state health agency — and read the project list. It names what the state committed to buying: a dated procurement roadmap almost nobody selling into rural health is reading.

NeuroFlow CEO Chris Molaro argues the PHQ-9/GAD-7 model is the wrong unit — episodic screening replaced by continuous longitudinal risk, on the strength of a NeuroFlow/Independence Blue Cross study of 13,000 commercial members: a 68% increase in the likelihood of receiving outpatient behavioral health care, $27.63 PMPM off total cost. Vendor’s own study. And nobody has yet agreed what an outcome record contains — a score with no clinical action attached is compliance theater with a timestamp.

Graham Walker, MD, co-founder and head of AI at MDCalc, is stepping away from clinical practice after 18 years in emergency medicine — explicitly not a rage-quit, no final terrible shift. He built MDCalc while working nights.

🎙️ From the Pods

🎙️ Health Tech Nerds Radio — “The Grand Roundup: Autonomous clinical AI is here…”

The most useful thing said about autonomous clinical AI this week was structural, not technical: the heart failure work clearing regulatory ground is “an FDA approved software as a medical device that’s based on a predicate device that was a digital therapeutic that’s been around for eight to 10 years.”

💡 Builder take: The shippable shape of “autonomous” is a narrow dosing loop with an existing predicate behind it. Find the predicate first, then design the loop.

🔇 Speaker Blindspot: Composition fallacy. One cleared dosing algorithm becomes “autonomous clinical AI is here” — never asking whether a grant program’s selection process is itself a filter for use cases already known to work.

🎙️ The Heart of Healthcare — “Anthropic’s Bet on the Future of Healthcare”, with Sarah Russell, MD, Anthropic’s healthcare and life sciences lead, who still works express care shifts at Stanford.

HIPAA’s technical safeguards assume a human opening a chart. “Right now we track users across the ecosystem. What if we tracked data across the ecosystem?” For agentic tooling on PHI, the auditable object is the data’s path, not the session. On whether AI has changed healthcare’s business model — as opposed to pajama time, where she says we’re doing nicely — “I would say we’re, like, still in the warm-ups.”

🔇 Speaker Blindspot: Appeal to motive. Asked whether Anthropic will compete with the startups building on it, the answer is that the mission isn’t healthcare — intent, not incentive. The Google analogy breaks too: Google entered as a customer-side competitor; a model vendor sees its customers’ usage.

💡 BTW

💡 BTW: Chris Molaro, who runs NeuroFlow, came to behavioral health through a platoon in Iraq — West Point, five years in the Army, then home to find that the system waiting for his soldiers didn’t work.

💺 Builder Seats

[These are just ones I found on LinkedIn that look interesting, no sponsorship or anything. Use at your own risk but look legit]

Senior Physician Executive — Abridge · Remote
Read today’s Big Thing, then read this posting.
🔗 Apply on LinkedIn

SVP, Clinical Guidance — Highmark Health · Remote (US)
Payer-side, which most clinician-builders skip. It’s where coverage criteria get written — the rules your tool argues with for life.
🔗 Apply on LinkedIn

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• A court says the ambient recording isn’t part of your chart — one of the first US rulings on ambient AI: using the DAX audio to draft the note serves an “administrative purpose,” so it’s exempt from disclosure. Full treatment below.

• Six health system leaders on ARPA-H’s autonomous heart failure agent: who carries the liability? — nobody in the $62.7M architecture has an answer, and one of them noticed who holds the patent on it. More below.

• Ambient AI arrives at the nurses’ station — Mercy cut flowsheet documentation time 22% per shift. Structured fields, not narrative, so it’s a different product.

• A year after “kill the clipboard,” CMS has eight more workgroups and no published progress data — Intermountain’s 769,000 app activations all came from inside its own walls, the part never blocked.

  🔮 AI’s bet: the first enforceable piece ships as a condition of participation, and it’s patient matching.

• 🎧 Podcast: HIMSSCast — “Prenosis CEO on using AI to detect sepsis” — Bobby Reddy Jr. says their data shows roughly eight distinct biologies under what we call sepsis. If your alert fires on a syndrome that’s really eight diseases, the false positive rate is a label problem, not a tuning problem.

  🔇 His oncology analogy paid off because each subtype got a different drug; sepsis subtypes only pay if they imply different treatment, which the episode never visits.

🧭 The Curbside

“CMS is asking whether the AMA should still own CPT. Comments close today — worth twenty minutes?”

• Short answer: If you’ve ever paid to put a code set inside your own product, yes.

• What changed: Buried in the CY 2027 Physician Fee Schedule proposed rule is a request for information on the harms flowing from the AMA’s monopoly over CPT-4 licensing, what alternatives to CPT and the RUC exist, and whether procedural services could be paid off ICD-10-PCS. Comments close today.

🔬 The Big Thing

The recording isn’t the record. That’s the whole ruling — and it’s load-bearing.

A Clark County, Washington judge ruled that a patient has no right to the ambient AI audio her clinic used to draft her visit note. The case is Raphael v. Mantei; the order is dated August 7, and it reached the national trade press only last week.

The patient wanted the DAX recording of a telehealth visit at the end of which her physician withdrew from her care. The clinic argued it was a drafting aid, not a record. The court agreed.

💬 Standout quote

“The court finds that the use of the DAX recording to assist physicians in preparing the written record of a patient encounter is an ‘administrative purpose.’ Thus, the DAX recording is exempt from disclosure under the Uniform Healthcare Information Act.”

The AMA Litigation Center and two Washington associations filed an amicus brief making the same distinction: the intermediate materials are temporary drafting tools, and only the physician-authenticated note carries professional judgment.

The line isn’t between audio and text. It’s between the record and the process of producing the record — and what separates them is a human signature carrying professional judgment.

Sit with what that means for everything else you build. Retrieval traces, intermediate drafts, tool-call logs, your model’s first answer before the guardrail rewrote it — all of it sits on the administrative side of that line, and this is the first authority saying so. It holds for exactly as long as a clinician authenticates the output. Remove that step and the artifact becomes the record itself.

😤 “This is a trial court in one county.” Correct, and it’s the first thing your general counsel will cite, because it’s the only thing there is. The AMA doesn’t file amicus briefs in cases it thinks will stay local.

😤 “So patients can’t hear what was said about them.” She wasn’t curious. She wanted the recording of the visit where her doctor dropped her. The ruling is coherent and does nothing about the fact that the audio exists, on somebody’s retention schedule.

😤 “Our vendor deletes the audio.” Then it isn’t retained, and this never comes up.

📡 Builder’s Radar

Nobody in the $62.7 million architecture will say who is liable

ARPA-H’s six ADVOCATE awardees split the work: Atman Health, Tempus AI and Updoc build the patient-facing heart failure agents, Stanford builds a supervisory AI to watch them, Duke and Kaiser deploy. FDA package due in 24 months.

Six health system leaders were asked the obvious question and none could answer it. The Christ Hospital’s Joy Oh put it best: “If a supervisory agent fails to detect an inappropriate recommendation that results in patient harm, who is ultimately responsible: the developer of the worker agent or the developer of the supervisory agent?”

Baptist Health’s Brett Oliver, MD, read the award as a regulatory signal: human-in-the-loop “may no longer be the primary safety design model... but rather a transitional governance model.”

🔮 AI’s bet: the supervisory agent gets its own FDA submission inside eighteen months — device that acts, device that watches — and that split becomes the template.

Ambient AI moves to the nurses’ station, and it isn’t the same product

Mercy is the first system to deploy the Android version of Microsoft’s ambient tool for nurses — three pilot units, a 22% cut in flowsheet documentation time per shift, reaching every med-surg unit by June 30, 2027. Northeast Georgia became the fifth system on Epic’s “Chart with Art: Nursing” in August.

Physician ambient AI turns speech into prose. Nursing ambient AI has to turn speech into discrete structured fields — a harder extraction problem with a much worse failure mode.

A wrong adjective in an HPI reads awkwardly. A wrong value in a Braden score or an I/O row propagates into a quality measure, an acuity calculation, and possibly a sepsis alert.

💡 80/20: Take one nursing flowsheet you know and mark which rows are free text, which are coded, and which feed a downstream score. Anything in that third bucket needs a confirmation step regardless of the accuracy number. Blank template, no patient data.

⚡ Quick hits

Dario Amodei’s “We Must Pace the Frontier” landed Saturday. The part that matters here isn’t the speed limit — it’s the commitment underneath: permanent, employee-like access to his own systems for an embedded outside evaluation team, with METR named as the kind of outfit he means. Sam Altman said the same day that OpenAI would do it too. Governance committees have been asking vendors who verified this other than you? and getting a self-attested framework with circles and arrows. If the labs concede outside verification, “we audit ourselves annually” stops being defensible one tier down.

Rollin “Terry” Fairbanks, MD (chief quality and safety officer, MedStar Health) flagged reporting that AHRQ funding has been quietly throttled — the agency behind most of the patient safety evidence base that clinical AI frameworks cite as prior art. Via LinkedIn.

Sanjay Dowerah on why medical ontologies never get adopted: curated by experts with no mechanism for reaching the data pipelines, and the tooling needs its own specialty.

🎙️ From the Pods

🎙️ How I Doctor — “Medicine Is Headed Toward Semi-Autonomous Care. Are Doctors Ready?”

Rishi Khakhkhar, MD, an emergency physician at Mount Sinai and CMO of Counsel Health, draws the line: AI takes the full history, but diagnosis, prescribing and referral stay with the physician — and the decision to escalate stays with the patient.

💡 Builder take: The escalation decision belongs to the patient, not the algorithm. A consent argument built into the design, cheaper than the governance machinery it replaces.

🔇 Speaker blindspot: False dichotomy — his Waymo analogy assumes the supervised phase generates the evidence that licenses the transition. Waymo’s supervised years produced an adversarially-mined disengagement dataset. Nobody asks whether anyone here is capturing override data.

📅 Upcoming - not free but good: HL7’s 40th Annual Plenary, Working Group Meeting and FHIR Connectathon, September 19–24 in Rockville, MD. If you write code, Rockville beats everything else on the fall calendar.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

Satchin Panda’s lab ran an experiment I think about often, even if it is pseudoscience and superseded by GLP-1’s.

Two cages of mice. Same chow — high fat, the rodent equivalent of a gas station at midnight. The same number of calories going into each cage. They measured.

One group could eat whenever it wanted.

The other could only eat inside an eight-hour window.

The free-feeding mice got fat. Fatty liver. Glucose intolerance. Inflammation. The whole syndrome.

The eight-hour mice, eating the same garbage in the same amount, stayed lean.

Nothing about the input changed. Only when it arrived.

He wrote a book about this called The Circadian Code, and the thing most people take from it is the eating window. The part I can’t stop turning over is the mechanism underneath: the same molecule is a different signal depending on what time the body thinks it is.

You already believe this about medicine.

A potassium of 7.2 that lands while the patient is still in the room is a calcium.

The same 7.2 that lands forty minutes after discharge as someone didn’t notice it was missing from the BMP is a phone call and a drop in your stomach.

Same number. Same analyzer. Same accuracy. Different actions given different timings.

We build as if the information is the product.

The model that’s 94% accurate and fires after the disposition is decided isn’t 94% anything. It’s a report.

The deterioration score that was right at 3 a.m. and surfaces in Monday’s dashboard is an audit.

The best differential in the world, hours late, is documentation.

We measure accuracy because accuracy is easy to measure. You can put it on a leaderboard.

We measure latency too — but as an engineering number. Milliseconds to response. Not the clinical version of the question, which is whether it arrived while the decision was still open.

I keep circling the feeling that there’s a whole kind of scoring in there that nobody does. Not how often it’s right. How often it’s right in time.

Here’s the part I’d defend in a room full of people who disagree.

You cannot learn the window from the data. It isn’t in the data.

The window is a lived thing. It’s knowing that between the moment a resident says “I think we’re sending him home” and the moment the note gets signed, there are maybe ninety seconds where a new fact can still change what happens — and after that it can only change the wording.

You know those ninety seconds. Nobody who hasn’t worked a shift knows they exist.

TL;DR

Same information. Different moment.

You don’t need a better model.

You need the right eight hours.

Accuracy you can buy by the API call. It arrives the same for everybody who pays.

The window doesn’t. It’s the residue of a few thousand shifts, and there’s no way to hand it to someone who hasn’t stood there while one closed.

It’s the one input you already own.

Most of us are spending it on the wrong question.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

New here? Every Saturday I take the week’s best idea in each of three buckets and cut it down to something you can read between patients.

I read the week so you don’t have to.

I’ve talked to many of you offline and you know I love Dan Hewitt’s Three Wins, so I thought to try to use that format on Saturdays for a bit.

Let’s go.

— Kevin

1. TIME

The part that didn’t get faster.

On Friday the FDA published three pages in the federal code — 21 CFR 870.2380 — describing what a cardiac AI has to prove.

The device that started it, Viz.ai’s Viz HCM, was authorized on August 3, 2023.

Three years. Three pages.

Now look at what’s actually in them. A test set independent from training. At least three geographically separate sites you didn’t train on. Performance reported per demographic subgroup, per site, per acquisition device. Human-factors testing for whether a tired person over-trusts the output.

None of that is engineering time.

That’s contracts, IRBs, enrollment, adjudication, waiting. Calendar time. It has not moved in twenty years and it is not going to move because you have a better model.

Here’s the uncomfortable arithmetic. The build got 100x faster. The evidence got 1x faster.

That gap is the whole opportunity, and it’s also the whole trap.

The trap is obvious once you say it out loud: you can now build fifty prototypes in the time it takes to validate one. So you will. Fifty weekends, fifty demos, nothing that anyone is allowed to use on a patient.

The opportunity is the same fact, held the other way round. Everyone can build it. Almost nobody will spend three years proving it.

So the scarce resource isn’t the idea anymore. It’s the willingness to still be working on the same idea in 2029.

Pick the one thing you’d stay with that long.

Then build it on a Saturday.

A win.

(Words: 251. A minute.)

“The test dataset must include a minimum of three geographically diverse sites, separate from sites used in training of the model.” — 21 CFR 870.2380

Dig deeper: the order itself. It’s three pages. Read it once and you’ll never write a validation plan the lazy way again.

2. MONEY

You’re not competing with the problem. You’re competing with a line item.

This week Mayo Clinic licensed CareCast — a workforce-demand forecaster it built inside its own walls — to the staffing company Trusted Health, for sale to other health systems. Mayo disclosed a financial interest.

Hold that next to something Ernest Grant, PhD, RN said the next day. Grant was the first man elected president of the American Nurses Association in its 130-year history, now a vice dean at Duke’s nursing school.

Nursing, he said, is “folded in on the room and board.”

Every other profession there bills for what it does. Nursing is priced into the bed.

Which explains a decade of confusion: nurse-facing software is the most obviously needed software in the hospital, and the hardest to sell.

No line item, no buyer.

Mayo didn’t solve that. Mayo went around it — selling the forecast to the company whose line item is labor, instead of a CFO with nowhere to book it.

That’s the move.

The pressure is real. Russ Branzell, who runs CHIME and hears from more health-system CIOs than almost anyone, said last week that systems have “run out of options from a financial pressures perspective” — several are working out how to offshore, or “even lay off hundreds or thousands of people.” His guest Andy Smith, who founded Impact Advisors, gave the other half: zero to 360 associates in Mexico City in three years.

So “saves a nurse thirty minutes” isn’t priced against a nurse. It’s priced against a nearshore associate at a third the cost.

Before you build, name the line that goes down.

A win.

(Words: 271. A minute, and a bit.)

Nobody buys a solution. They move money from one line to another and hope you were worth it.

Dig deeper: the Mayo/Trusted announcement. Read it as a distribution story, not a staffing story.

3. MOJO

All the work nobody asked for.

In 2008, Stephen Smith — an emergency physician at Hennepin County Medical Center in Minneapolis — started posting ECGs on a blog.

Not the obvious ones. The subtle occlusions — where the artery is blocked, the machine says normal, and the patient goes back to the waiting room.

Nobody assigned it. There were no RVUs. He just kept noticing the same thing getting missed, and kept writing it down.

Eighteen years. Thousands of cases, one at a time.

Last week the FDA authorized a device built on that judgment — Powerful Medical’s Queen of Hearts, which Smith trained to catch what the standard criteria miss. It came through de novo: the long door, the one only a handful of AI devices go through in a year.

Here’s what I keep turning over.

He wasn’t building a dataset. If you’d told him in 2008 that the blog was the moat, he’d have laughed, because in 2008 the moat was the model and the model didn’t exist yet.

He was just a guy who couldn’t stop noticing.

That’s the part you can’t buy, can’t prompt, and can’t catch up on. Compute is rentable. Engineers are hireable. Eighteen years of paying close attention to one specific way that people get hurt is not available at any price.

So the question isn’t what you should build.

It’s this: what have you been noticing for years that nobody asked you to notice?

Write that down. Publicly. Badly, at first.

You are already ten years into something. You just haven’t called it a dataset yet.

A win.

(Words: 265. A minute, just about.)

The stuff you do because you can’t help it is the only part of you that isn’t for sale.

Dig deeper: the blog. Start anywhere. It’s still going. And the authorization write-up, if you want the regulatory side of the same story.

Love this format for Saturdays? Hate it? Thinking “I wish it were _______”? Let me know at kevin@clinicians.build — I read every one.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• HTI-5 is reportedly finalizing without its information-blocking fixes — including the proposed language that would have written autonomous AI systems into the definition of “access.” More below.

• Your BAA does not cover the connector — Anthropic’s own coverage table says data sent to third parties through MCPs or connectors falls outside the agreement. Most clinical agents are mostly connectors.

• Inspiren raised $70M past a $500M valuation — its platform now powers care at communities across more than 80% of the largest senior housing investors, fall detection included. Nobody calls it clinical software, but a system that detects a fall and routes it to a caregiver is making an acuity judgment in a building with no physician on the floor at 3 AM.

• Healthcare beats open banking on the rules — Brendan Keeler ran it: we have a standardized API mandate and no-fee patient access. They have a §1033 rule their own agency is now asking a court to vacate, and banks charging aggregators.

🔬 The Big Thing

Your agent was about to become a “user.” ONC is having second thoughts.

HTI-5, the fifth Cures Act interoperability rule, proposed revising the “access” and “use” definitions at 45 CFR 171.102 — in the preamble’s own words, to emphasize that they include “automated means of access, exchange, or use of EHI—including, without limitation, autonomous AI systems.” The proposed regulatory text names it directly: agentic artificial intelligence under “access,” autonomous artificial intelligence systems under “use.”

Ruth Reader reports at Second Opinion that the rule is now finalizing without the information-blocking changes, with ONC head Thomas Keane confirming a couple-of-weeks timeline at CTA’s Health AI+ event Wednesday. The rule is at OMB. Treat that as reporting until the Federal Register says otherwise.

Pushback is on the record. The American Hospital Association asked ASTP/ONC in February to hold off on the AI definitions until it issued broader guidance on how information-blocking exceptions apply to AI, and to keep the third-party-modification exception. Second Opinion reports the objection was jurisdictional rather than substantive: ONC isn’t the enforcement body for information blocking, the HHS Office of Inspector General is.

So whether an autonomous agent asking for a chart has the same standing as a human asking for a chart stays unanswered — during the exact eighteen months when every serious builder is shipping one.

If your tool requests EHI on a patient’s behalf and the response is slow, partial, throttled, or refused, the information-blocking framework is your only lever — and it was written for a world where a person clicks a button. Without the update, “we don’t serve bots” sits in an unlit room: not obviously permitted, not obviously prohibited, not obviously anyone’s to enforce.

So the rule you’re building against for two years is whatever each EHR vendor writes into its developer terms. That isn’t regulation. It’s a terms-of-service regime with one vendor at the center.

😤 “This is deregulation. Deregulation is good for builders.” Not when the thing deregulated is the rule that gave your software standing. Lower certification burden helps whoever has to certify; a narrower access definition helps whoever controls the endpoint. Only one of those is you.

😤 “ONC overreached.” Possibly — OIG does hold the enforcement pen. It’s still worth noticing that the entity with the most to lose from agentic access is reported to have made the argument that carried.

😤 “Nothing’s final, calm down.” Watch the Federal Register.

📡 Builder’s Radar

Your BAA stops at the connector

Steinlin’s four gaps are all post-signature. AI logs become a second, wider PHI repository — prompts, responses, tool calls, sometimes whole clinical narratives, readable by anyone with a debugging reason. Feature-level coverage is not vendor-level coverage. 42 CFR Part 2 records didn’t get simpler because February’s amendments simplified consent. And every vendor around the model — vector store, log sink, connector, MCP server — is another PHI recipient.

Steinlin’s line is the one to keep: “A BAA that covered last quarter’s chat workflow may say nothing about today’s autonomous agent.”

💬 Standout Quote

“If you cannot say which passage came from where, you cannot show which rule applied.” — Mat Steinlin, Aptible

💡 80/20: Draw the data-flow map, not the BAA folder. One page: user input → retrieval source → model → tool call → log destination → reviewer → retention. Synthetic Synthea case, zero PHI.

Tennr bought the cold open

Tennr named Toby Cosgrove and Cindy Hundorfean industry advisors — thirteen years running Cleveland Clinic, and seven running Allegheny Health Network.

Read it as procurement, not personnel. A marquee advisor doesn’t shorten the path from departmental pain to signed contract — it raises the altitude you enter at. The CEO’s office instead of the forty-deep CMIO inbox.

Nursing informatics leaders say nurses have to govern the AI rewriting their workflows — co-designers of the tool, not end users who meet it at go-live.

H1 acquired Defacto Health — third buy in two years, landing as CMS tightens on Medicare Advantage “ghost networks” and makes plans attest to directory accuracy. Provider data is becoming a compliance product.

Rohan Ramakrishna, MD (neurosurgeon, co-founder of Roon) posted the uncomfortable version with Adam Rodman: what if keeping a doctor in the loop makes the AI’s answer worse? Nobody designing a review step has an answer that survives that question.

Jung Hoon Son, MD is running Kimi K3 as his primary coding driver, explicitly as a hedge against depending on one closed provider. Make your eval harness portable before your model choice is permanent.

🎙️ From the Pods

🎙️ CEO Pajama Time — “In the Business of Reinvention”

Oron Afek, CEO and co-founder of Vim, on the failure that made the company: BookMD texted patients an Amazon-style MRI shopping flow — deductible waived, free Uber, photos of the facility. Open rates were great. Almost nobody booked. On callbacks the reason wasn’t loyalty to UCLA. It was “my appointment is already scheduled.” Health systems were staffing free prior-auth processing into community PCP offices, capturing the referral by fax, and scheduling before the auth reached the payer.

💡 Builder take: Find the timestamp where the decision actually locks, and who locks it. Afek on the rest: “The tech part was easy.”

🔇 Speaker Blindspot: False dichotomy. He sells against “probabilistic” agentic AI by promising “100% confidence that a write-back is a write-back and an order is an order” — while his headline launch lets customers vibe-code apps that write to the EMR. The determinism lives in the transport layer. A wrong order delivered with perfect fidelity is still a wrong order.

💡 BTW

💡 BTW: Toby Cosgrove — the surgeon Tennr just named an industry advisor — got three D-minuses and a D at Williams College, applied to thirteen medical schools and got into one. His grades stayed mediocre until he hit the clinical years, where he finally started pulling A’s. He didn’t learn he was dyslexic until adulthood. He went on to perform more than 22,000 operations and run Cleveland Clinic for nearly thirteen years.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• ARPA-H commits $62.7M to partially autonomous heart-failure agents — agents that assess symptoms, prescribe drugs, order labs, with FDA authorization as the target. More below.

• The AMA released the CPT 2027 code set — 299 new codes, 10 of them AI, effective January 1. Your tool has a pipe or waits until 2028.

• FDA now has a Deputy Commissioner for Technology and AI — Jared Seehafer, the first to hold the role, named eight days after the agency’s longtime digital health lead walked.

• Stanford says walking away from AI vendors over data terms is routine now — CTO Christian Lindmark has dropped vendors over data and contract terms and calls it routine; CHOP has done the same.

• 🎧 Podcast: Podnosis — “How video games are changing medical education” — Sam Glassenberg, now EVP at Relevate Health Games, compressed a months-long diabetes caregiver curve into hours of gameplay. ~50,000 downloads.

🔮 My bet: the supervisory layer becomes a product category before it becomes an EHR feature. By Q3 2027 at least two health systems will buy agent-monitoring separately from the agents — because the vendor selling you the agent is the worst possible party to grade it.

🧭 The Curbside

“Should I fine-tune an open model for a specialty, or just prompt a frontier model?”

• Short answer: Fine-tuning wins on narrow, bounded tasks, and the evidence just got specific. EviNurse, a Qwen3-32B fine-tune on nursing evidence, hit 91.45% on a 3,438-question benchmark and beat general-purpose models in expert review (npj Digital Medicine, Sept 9).

• Builder read / Watchout: It was built and evaluated in mainland China — 316 nurses across every province — so scope of practice and documentation standards don’t transfer. The recipe does: a 32B open model on consumer hardware beating frontier models on a bounded vocabulary. Steal the method, not the weights.

🔬 The Big Thing

Who watches the agent that writes the prescription?

ARPA-H committed $62.7M to a program called ADVOCATE to build partially autonomous, FDA-authorized AI for heart failure — agents that assess symptom severity, prescribe medications, and order labs.

Atman Health, UpDoc, and Tempus AI are building them. $33.7M lands in year one.

Then there’s the part that changed how I read it.

Stanford’s award isn’t for an agent at all. It’s STEWARD — a disease-agnostic supervisory system that monitors the clinical agents after deployment, catching unsafe recommendations and out-of-distribution behavior and producing inspectable per-claim rationales.

The government separated the agent from its supervisor, funded them as two different jobs, and handed the supervisor to a different institution.

That’s not a technical detail. That’s a statement about who gets to grade the work.

Duke validates across five health systems and rural sites on both Epic and Oracle Health. Kaiser Permanente deploys across 21 medical centers and 260+ clinics inside Epic workflows.

So the path is real, it runs on the EHRs you already live in, and the safety layer has its own accountable party.

😤 “Autonomous prescribing is a lawsuit with a user interface.” Partially autonomous. Every one of those actions already happens under a protocol, a standing order, or a pharmacist with a collaborative practice agreement. What’s new isn’t the autonomy — it’s that the autonomy has to show its work continuously instead of once, at a 510(k) moment.

😤 “ARPA-H funds moonshots that die at the demo.” Then watch Duke, not the agents. Multi-site, dual-EHR validation infrastructure outlives whatever gets built on it, and the private market has never once paid for it.

❓ The supervisor is disease-agnostic; the agents are disease-specific. So the supervisor is the one with the general market. What is that product — a service, a library, a certification body? There’s a company sitting in that gap and I can’t name it yet.

📡 Builder’s Radar

AI now has 43 CPT codes. Yours probably isn’t one of them.

The AMA released the CPT 2027 code set on September 9 — 299 new codes effective January 1, ten of them AI-related, bringing the total to 43. The Editorial Panel also updated Appendix S, the taxonomy sorting AI into assistive, augmentative, and autonomous.

That taxonomy isn’t academic. My read: autonomous is the only tier where the operating entity has a plausible path to billing without a physician encounter. Everything else is a line item you have to justify inside someone else’s fee.

If your service isn’t in the 2027 set, your next shot is the February panel — submit by November, effective January 2028.

The FDA hired a technology chief before it replaced its digital health lead.

HHS named Jared Seehafer the FDA’s first Deputy Commissioner for Technology and AI on September 8. The role is commissioner-level and agency-wide, which puts AI strategy above CDRH rather than inside it.

Read that as centralization, not staffing.

🔮 Where this lands: the test is the generative AI docket (FDA-2026-N-7874), open through October 19. If that yields a real review framework by spring instead of another discussion paper, the office has teeth. If not, it’s a layer.

A journal is recruiting collaborators to write the validation standard.

Nature Medicine published a correspondence on September 9 calling for collaborators on the “Validation Accords,” a consensus framework for validating generative AI before clinical deployment.

Bright Huo leads it and Gordon Guyatt is the last author — an international working group, but the standard it produces will land on US deployments too. Their argument: generative AI gets validated on benchmarks and rarely gets tested in the setting where it will actually run.

Rare one where the ask is literally in the title — “call for collaborators.” If you’ve wanted a clinician-builder in the room while the evaluation rules get written, that room is open right now.

⚡ Quick hits

Cleveland Clinic partnered with Luminai to read inbound external referrals — many still arriving by fax — into the EMR as structured data. A 2015 problem with a 2026 price tag, and still the highest-volume unstructured-document workflow in American medicine.

Meta shipped Muse, a consumer agent that runs inside a dedicated “Secure VM” with its own browser, with a Confidential VM coming later this year that encrypts the whole machine under a key only the user holds. Untrusted agent, hard sandbox, containment as the headline feature — the same instinct ARPA-H just funded.

Dr. Josh Au Yeung co-founded Vitruvia Labs with a clinical + ML + privacy/cryptography founding team, pitched as going beyond “applying llm wrappers.” The privacy-engineering seat on a founding team is the tell.

Brian Fung is running a second cohort of his build-with-AI bootcamp for clinicians — two weeks, live, starting September 28.

🎙️ From the Pods

🎙️ Podnosis — “How video games are changing medical education”

Sam Glassenberg’s team replaced static type 1 diabetes caregiver education with a decision game — his framing is that it trains the brain to manage diabetes in hours instead of months. Roughly 50,000 downloads since launch.

💡 Builder take: Patient education is the most under-designed surface in medicine and the one with the lowest regulatory bar. Your first shippable artifact is here, not in the EHR.

🔇 Speaker Blindspot: Survivorship bias — every example is a game that worked, and the claim that efficacy holds “no matter how old you are” rests on studies recruiting doctors who volunteered to play a game.

🎙️ Relentless Health Value — “EP528: The Current State of Primary Care—Inevitable or Fixable? Will It Be Private Equity to the Rescue?”

Yashaswini Singh, PhD on PE-acquired primary care: 20% more preventive care delivered, no detectable rise in low-value care, prices negotiated 8–10% above independent practices.

💡 Builder take: The lever that moved was administrative capture, not clinical quality. Build for the preventive-care gap and the coding of it, not the ownership model.

🔇 Speaker Blindspot: Moving the goalposts — asked three times whether a bad outcome is “inevitable,” she offers a theoretical best case, concedes the data can’t confirm it, then admits she suspects the worse reading.

💡 BTW

💡 BTW: Gordon Guyatt — on the byline of that Validation Accords call — originally wanted to name the movement “scientific medicine.” In his telling, the basic scientists at McMaster were “so enraged that I was calling what we were doing ‘scientific medicine’” that he went back to the drawing board and returned with “evidence-based medicine” instead. That was 1990. (JAMA/BMJ oral history, via AMA)

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Forus raised $150M at a $3B valuation to give every prescription its own AI agent — prior auth, benefits verification, appeals, financial assistance, specialty pharmacy routing. Tripled since May.

  🔮 My bet: within 18 months a major payer buys an administrative-agent company outright. The cheapest way to kill a toll booth is to own it.

• Trinity Health is cutting 557 IT jobs and outsourcing technology support across a 23-state system.

• Five AI drug discovery companies: $458.7M in revenue, $5.58B in losses since 2021 — and the one with an FDA-approved product is in Chapter 11.

• FDA’s digital health leader is out after 21 years — more turnover in the office writing the rules you build against.

• 🎧 Podcast: Lifers — Microsoft’s Peter Lee and Seattle Children’s Chris Longhurst — Lee cites a Nature finding that general frontier models beat the healthcare-specific models built on top of them.

🧭 The Curbside

“My agent reads a note it didn’t write. Can that note tell it what to do?”

• Short answer: Yes, and most clinical agents have no boundary between data and instruction.

• What changed: A walkthrough of prompt injection through untrusted tool output went up this week. The attack isn’t in the user’s message — it’s in what the tool hands back.

• Builder read: A clinical agent’s tool output is a FHIR narrative field, an OCR’d fax, a scanned outside record, a patient-entered chief complaint — all text somebody else wrote. Treat every tool return as hostile string data, not as a turn in the conversation.

😤 “Nobody is attacking my little internal tool.” Not yet.

“How do I prove my coding agent actually ran the checks?”

• Short answer: Make the proof a signed artifact instead of a claim in a log.

• What changed: Vouch shipped a small CLI that gates a coding agent’s Stop hook behind an HMAC-signed receipt proving named checks passed against the current git HEAD. The agent can’t say “done” without it.

• Builder read: “The tests passed” is a sentence. A signed receipt tied to a commit is evidence.

🔬 The Big Thing

What’s the gap between a prescription and a patient worth? Three billion dollars.

Forus — formerly Tandem — raised a $150M Series C led by Bain Capital Ventures at a $3B valuation. That’s triple its May valuation, four months later.

Try to find the medicine in the product description: prior authorization, benefits verification, appeals after denial, financial assistance enrollment, specialty pharmacy coordination, pharmacy routing, patient communication. Providers in all 50 states, reaching 85% of US residential zip codes.

Every step in that list happens after the clinical decision is already made. The drug is chosen. The patient is waiting. Three billion dollars is what the market thinks the wait is worth.

Here’s the part I keep circling. On October 1, UnitedHealthcare removes prior authorization from roughly 1,700 service codes — the largest single-payer cut of the year. If administrative friction is the product, the product just lost inventory.

I don’t think it matters, and that’s the interesting part. The gate is one of maybe fifty steps between the order and the drug. Remove it and the specialty pharmacy still routes wrong, the assistance application still expires, the appeal still needs a letter. The bottleneck was never a single gate. It was that no one person has ever seen the whole path.

😤 “This is a $3B valuation for a fax machine with a language model on it.” Partly, yes. And the fax machine was load-bearing.

😤 “Automating a broken process cements the broken process.” The strongest objection, and I don’t have a clean answer. Every agent that makes prior auth survivable makes prior auth cheaper to keep. The counter is that UHC dropped 1,700 codes precisely because the administrative cost exceeded the denial savings. Ask me in a year.

😤 “Forty percent of agentic AI projects get cancelled before 2027.” Probably true. The survivors will be the ones where somebody could name the exact step being automated and who gets paged when it fails. “AI agents for healthcare” is a cancellation. “The letter that goes out when a specialty drug is denied for step therapy” is a product.

❓ What does the second wave look like? Forus sells the automation to the practice. Somebody is going to sell the same visibility to the patient — a live map of where their prescription is stuck. I think there’s a real product there and I can’t yet see who has the right to build it.

🧪 Try the interactive: Nobody Has Seen the Whole Path — 931 molecules across all 328 Medicare Part D formularies, and 495 of them where some plans demand prior authorization and others do not. Built with real CMS data.

🧪 Try the interactive: Same Drug, 328 Answers — you write the prescription, and what happens next depends on which of 328 Part D formularies your patient happens to be on: one drug, five different answers. Built with real CMS data.

📡 Builder’s Radar

A 23-state health system just outsourced the people who answer when your integration breaks

Trinity Health filed to lay off 557 employees at its Livonia headquarters between October 25 and November 29, moving technology support to an outsourcing arrangement — service desk, database administrators, network security analysts, applications engineers. The system cited federal Medicare and Medicaid cuts.

If you sell into a health system, your implementation partner is now a contract, not a colleague.

😤 “Outsourced IT is still IT.” It is. It’s also a vendor with a scope of work, and your integration escalation is either in the scope or it isn’t.

AI drug discovery has produced $458.7M in revenue and $5.58B in losses

A tally of five AI-native drug discovery companies published Monday puts cumulative revenue since 2021 at $458.7M against $5.58B in losses. BioXcel — the only one in the group with an FDA-approved product — filed Chapter 11 in late August, Teva stalking-horse at $57.5M upfront.

Approval was never the hard part. Distribution was.

Benchmarks grade the first answer and never grade the doubt

John Ferguson, MD, FACS argues that medical AI benchmarks score first-answer accuracy and never score calibration — so a model that’s right 70% of the time and unshakeable on all of it is more dangerous in a clinical workflow than one that’s right 60% and flags its own uncertainty.

FDA’s digital health leader leaves after 21 years

A long-tenured digital health official is out at FDA, continuing a run of turnover in the office that writes guidance for AI-enabled devices. Twenty-one years of institutional memory about how a submission actually gets read.

⚡ Quick hits

Benjamin Tran, MD mapped the LA aesthetic market from public data — 2,000+ practices tagged with toxin used, injector headcount, real ownership, devices carried. Not a pitch deck. A dataset nobody else had.

GenHealth.ai raised a $16.5M Series A led by Flare Capital for agents doing intake, eligibility, prior auth and claims inside existing EHR and payer systems.

🎙️ From the Pods

🎙️ Lifers with Christina Farr — “Microsoft’s Peter Lee & Seattle Children’s Chris Longhurst”

Lee cites a Nature finding that general frontier models outperform the healthcare-specific models built on top of them, and argues that fine-tuning on narrow medical data doesn’t reduce hallucination — it produces brittle savants. Longhurst’s counter is that the model was never the lever: outcomes come from workflow integration, and health systems on grocery-store margins can’t afford frontier inference at scale anyway.

“We shouldn’t be comparing AI against a perfect gold standard.” — Chris Longhurst

💡 Builder take: Wrap a general model, spend your engineering budget on routing and cost, and put the domain knowledge in the workflow rather than in the weights.

🔇 Speaker Blindspot: False analogy. Lee maps healthcare’s coordination friction onto software teams, where coding agents collapsed specialists into full-stack generalists. Healthcare’s friction is licensure, liability and reimbursement — the ER example he’s answering was blocked by scope-of-practice rules, not by anyone failing to be full-stack.

💡 BTW

💡 BTW: Before he was the face of Microsoft’s medical AI work, Peter Lee ran experimental undersea datacenters and a program to store digital data in DNA — and made his academic name co-inventing “proof-carrying code,” where a program ships with a mathematical proof that it’s safe to run. Microsoft Research bio.

📅 Upcoming: Stripe’s health tech AMA on payment infrastructure for digital health startups — billing, fraud, build-vs-buy — Thursday Sept 10, 12:00 PM ET.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Nature Medicine wants a different scoreboard — Kristina Lång, writing from one of the first randomized trials of AI in medicine: judge human–AI systems on patient outcomes, not algorithms on clinician-matching.

• Two internists put the accountability question in writing — Andrew Parsons (UVA) and Adam Rodman (Beth Israel Deaconess) on how risk and oversight get allocated between clinician and model. Published this morning.

• The surrogate moved and the outcome didn’t — Novartis’s Lp(a)HORIZON readout: pelacarsen lowered Lp(a) as designed across 8,323 patients and missed its primary cardiovascular endpoint. The lead below.

• CMS opened the FY2028 NTAP town hall docket this morning — applications close Oct 5 at 5:00pm ET via MEARIS, well before the Dec 9–10 meeting.

  🔮 My bet: the Sept 30 Breakthrough-designation cutoff catches more device teams than Oct 5 does — it closes a pathway rather than a window.

• 🎧 Podcast: Radio Advisory — “The biggest ASC trend isn’t growth. It’s strategic variation” — Rae Woods on why ownership structure, not specialty, decides what a surgery center wants your software to do.

🧭 The Curbside

“Our patient app reads Fitbit data. Do we have a problem this month?”

• Short answer: Yes, and it’s not a code problem — it’s a consent problem.

• What changed: The legacy Fitbit Web API turns down at the end of September; cloud access moves to the Google Health API, where 120-plus endpoints collapse into 31 data types. OAuth tokens do not transfer. Every user re-authorizes before the cutoff, and intraday access now needs separate approval that isn’t instant.

• Builder read: The failure mode is silent. Your app won’t throw a 500 — it will just stop receiving syncs. Same shape as Carequality’s delegation-of-authority enforcement, where non-compliant queries return empty results instead of errors. Alert on data-flow cessation, not HTTP status.

• Watchout: While you’re in there, check whether your aggregation layer hands you the raw measurement or the vendor’s processed score. Build on the measurement.

🔬 The Big Thing

Your model got better at the part that was never the bottleneck.

On Friday, Novartis reported that pelacarsen lowered Lp(a) exactly as designed in 8,323 patients with established cardiovascular disease — and did not reduce heart attacks, strokes, or cardiovascular death.

The mechanism worked. The target moved. Nobody got better.

On Monday, Nature Medicine published a comment from Kristina Lång, who ran one of the first randomized trials of AI in medicine, with the same shape: the first generation of medical AI was judged on whether algorithms could match clinicians, and the next should be judged on whether human–AI systems improve patient outcomes.

Same finding, two fields. We are extremely good at moving the step we can measure, and we keep discovering it wasn’t the step that decided anything.

This isn’t an argument that models don’t work. It’s an argument about where the variance lives.

In drug development, AI compresses discovery — roughly five percent of R&D cost — while Phase 3, more than half the cost, is untouched. In clinical software, the model reads the chart in two seconds and then waits behind a governance committee that meets monthly and reviews three things a meeting.

If your eval scores the model, you are measuring the fast, cheap, already-solved part of your own pipeline.

The builder version is concrete. Your differential-generator hits 92% top-3 accuracy.

Does the attending open it? Does the recommendation survive the nurse’s read-back? Does the order get placed, or die in a soft-stop nobody has looked at since 2019?

Every one of those is measurable. None of them are on anyone’s model card.

😤 “This is just outcomes research wearing a hoodie. We’ve known about surrogate endpoints since CAST.” That’s the point. Medicine learned this in 1989 with encainide and flecainide, which suppressed PVCs beautifully and killed people — then AI showed up and we cheerfully rebuilt the same mistake with AUROC. The lesson isn’t new. We had it and spent it anyway.

😤 “Outcome studies take three years.” You’re not running one. You’re instrumenting the step after your model — the handoff, the click, the order. Weeks, not years.

😤 “Our accuracy numbers are what get us in the door.” Sure. Then you’re in the door.

❓ What product sits between the model and the outcome? Not another scribe, not another CDS panel — the thing that watches whether a recommendation turned into an action and tells you where it died.

📡 Builder’s Radar

“By 2030 the software will read every chart before I do”

Doug Fullington, MD, an internist 25 years into primary care, published a forecast on Sunday: software reads every chart, result, and message first, and the physician keeps the ambiguous diagnoses, the exam, the contradictions.

His sharpest point isn’t the timeline. It’s legal. Delegating a task to a person under standing orders is settled law. Delegating it to software is not a recognized legal category at all.

Texas already noticed — SB 1188 and HB 149 now require disclosure when AI is used in diagnosis or when a patient interacts with an AI system. Build the disclosure string in now: two lines today, a compliance project in eighteen months.

“Physician” is not one thing, and the 4x is the story

AMA CEO John Whyte and Zeke Emanuel went at each other on video over an August JAMA analysis arguing AI could outperform physicians on many clinical tasks.

The number that mattered was buried further down — Garner Health’s Nick Reber citing a 4x spread in complication rates between best- and worst-quartile physicians inside the same brand-name system.

The AI-versus-physician frame assumes the comparator is uniform. The variance inside the profession is bigger than the variance being argued about.

😤 “So replace the bottom quartile with a model?” No — if you can’t see the 4x you can’t route around it either, and routing is a far easier product than replacement.

A one-person shop built the tracker for the FDA’s genAI docket

Arvita Tripati, a regulatory and quality engineer with two decades at AliveCor, Vineti, and Clip Health, shipped a public tracker mapping every submitted comment on the FDA’s generative-AI device discussion paper to the specific question it answers. The docket closes October 19.

Reading what everyone else is arguing beats filing your own comment, and takes twenty minutes.

⚠️ Verify: a solo practice’s side project, not an official mirror. Cross-check anything load-bearing against regulations.gov.

“The transcript is a claim. I wanted a receipt.”

Kyle Redelinghuys shipped Vouch, a Go CLI that won’t let a coding agent fire its Stop hook — the “I’m done” signal — unless the named checks actually ran and passed against the current git HEAD, emitting an HMAC-signed receipt bound to that commit.

Swap “tests passed” for “the model reviewed the med list” and it’s the same problem with a worse blast radius.

He’s honest about the limit: the secret is repo-local, so it’s personal attestation, not third-party audit.

🎙️ From the Pods

🎙️ Radio Advisory — “The biggest ASC trend isn’t growth. It’s strategic variation”

Outpatient joint replacement went from about 5% of cases five years ago to 1 in 5 today, half in some markets. The pearl is that a physician-owned surgery center and a hospital-owned one want opposite things from the same building — throughput and equity value versus offloading low-acuity cases to refill the hospital OR.

💡 Builder take: Ask who owns the site before you demo. The same tool is a scheduling accelerator to one buyer and a case-routing filter to the other.

🔇 Speaker Blindspot: Appeal to authority — CMS removing total knee and hip from the inpatient-only list is treated as having settled the safety question, and the one clinical objection raised gets a reductio (”we’re not doing open hearts in ASCs”) instead of outcomes data. No clinician or patient voice appears.

💡 BTW

💡 BTW: Adam Rodman — co-author of the Annals editorial on AI accountability — is also a medical historian who started a narrative podcast about the history of medicine as a second-year internal medicine resident, essentially because he wanted a Radiolab for clinical medicine and nobody had made one.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• AI-generated code needs an AI-powered reviewer — Redesign Health built an internal system around one observation: AI makes telltale mistakes, not random ones. (The Big Thing, below.)

• Microsoft announced Project Zenith — Windows for 64GB+ developer machines, built to run 30B+ parameter models locally with no usage metering. The word that matters is metering.

• UW Medicine is going back to a base Epic build — customer number 30, live since 1996, stripping thirty years of customization to adopt new features faster.

  😤 Haters: “That’s a budget cut wearing a strategy costume.” Sure. It’s also the reason that happens to be true.

🧭 The Curbside

“Should I fine-tune a small model instead of paying frontier prices?”

• Short answer: Only once the workload is boring — high volume, bounded, measurable.

• Evidence: Ben Dickson’s weekend breakdown of Shopify’s pipeline: a 0.8B specialist beat a frontier model on one narrow task, cutting serving cost from ~$27M/year to ~$1M/year. The hard part wasn’t training. It was building the judge.

• Builder read: Shopify’s production data held no examples of a correct refusal, because refusals never looked like successes. Yours won’t either.

  🔮 My bet: the first clinical workloads distilled this way aren’t diagnostic — they’re de-identification, chart summarization and prior-auth drafting, inside health systems, before any vendor sells it.

🔬 The Big Thing

Your AI wrote the code. Who reads it back?

Aron Szanto, head of technology at Redesign Health, spent Friday’s HIMSSCast on the part of the vibe-coding story health tech coverage keeps skipping: what happens after the model writes the software.

His team built an internal reviewer called Argus, aimed at AI-produced code — code that can look completely functional while carrying problems nobody spots.

Then he said the thing I can’t stop turning over. “When an AI makes a mistake in coding, they’re not uniformly random ... They make telltale mistakes.”

Patterned failure is a gift. Random failure you can only catch by reading everything. Patterned failure you can build a detector for.

That distinction is load-bearing if you’re a clinician who builds things.

Every argument against clinician-built software ends in the same place: you’re not an engineer, your code will be unsafe, no security review will pass it. Fair, when the only alternative was a senior reviewer you couldn’t afford and couldn’t recruit.

Much weaker when the reviewer is a specialist that runs on every commit and already knows the seventeen ways an LLM gets authorization wrong.

None of which makes the regulatory surface disappear. A BAA is still a contract, SOC 2 is still an audit with a human on the other end, and no reviewer catches a design decision that should never have been made. But annual and continuous are different regimes, and continuous is the one agentic review makes cheap. On that single axis, a clinician-built tool with a harness on every commit is in better shape than the enterprise vendor with a pen test each October.

😤 “Redesign Health has an interest in saying this.” They’re a venture studio, not a code-review vendor. Argus is internal tooling built because the bugs were real. Building for your own portfolio is a better signal than launching a product, not a worse one.

😤 “You still can’t read what it wrote.” Neither can the staff engineer at whatever vendor your system just signed, at the volume they’re shipping. Pick your unread code.

❓ If coding failures are patterned and enumerable, the same should hold for clinical failures — per model, per task, per specialty. Hallucinated citations. Anchoring on the chief complaint. Silent unit conversions. Nobody is shipping the Argus for clinical output, and I suspect the taxonomy is the product rather than the detector. I can’t see the shape of it yet.

🧪 Try the interactive: Telltale, Not Random — two views of the same idea, built with real FDA MAUDE adverse-event data.

A — Telltale, Not Random — every device class the FDA heard about in 2024, placed in a triangle by how it fails: malfunction, injury, death. The points pile into the corners. Press one button and see where they would sit if failure were random.

B — The Shape of a Failure — 560 device classes and 2.47 million adverse event reports plotted by volume against harm. 61% of classes are more than 90% a single event type. Random would be 2%.

📡 Builder’s Radar

A Windows box that runs a 30-billion-parameter model and never phones home

Microsoft announced Project Zenith Friday: a preconfigured Windows developer setup for machines with 64GB+ unified memory and 250GB/s+ bandwidth, aimed at running 30B+ parameter models locally without usage metering. Lenovo’s ThinkCentre X Ultra starts at $3,699 in November.

The interesting number isn’t the price. It’s the metering.

Everything you can’t put on a metered API becomes possible on a machine in a room you control.

💡 80/20: If you’ve been running Ollama locally and hitting the wall around 14B, this is the hardware class that unblocks you.

Epic’s 30th customer is ripping out its own customizations

UW Medicine CFO Jon Alford — customer 30, first go-live 1996 — says the system is heading back to a base Epic platform. “Organizations need to get back to a base platform to be able to implement new technology faster as it starts to come at us.”

Thirty years of legacy build is thirty years of somebody’s very good idea, each one now a reason a new feature won’t turn on.

🔮 Where this lands: by the end of 2027, “AI readiness” is the stated rationale for de-customization at most large Epic shops. The unstated one is that nobody can staff the legacy build.

🎙️ From the Pods

🎙️ 229 — “ChatGPT Did Not Come Through Epic’s Front Door”

Bill Russell takes the ChatGPT/UCSF integration apart: read-only, the USCDI subset rather than the full record, through the standard Cures Act FHIR path — not a privileged Epic partnership. OpenAI’s own terms say not to use it for diagnosis, which is the exact thing everyone is excited about.

💡 Builder take: “Connected to Epic” now means at least four different things. Ask which one before you build a roadmap on a press release.

🔇 Speaker Blindspot: Composition fallacy — arguing this changes nothing because systems “already have this through our ambient vendor” treats an enterprise contract as if it were a capability. That access is a business arrangement between two companies. It isn’t something the physician on shift can invoke, or a clinician-builder can build against.

🎙️ HLTH — “How Health Plans Can Close the Gap Between AI Ambition and Data Reality”

A survey of 103 health plan executives on the distance between AI ambition and provider-data quality. The line that landed: “AI doesn’t fix a bad foundation. It actually scales it.” A manual error costs one bad phone call; the same error learned by a model repeats at machine speed across adequacy filings, credentialing, member search and contracting.

💡 Builder take: Before you price the model, price a wrong record. If you can’t put a number on what one bad row costs downstream, your ROI math runs backwards.

🔇 Speaker Blindspot: Sponsored-survey framing — the research was commissioned by a provider-data company, and the finding is that provider data is the bottleneck. That doesn’t make it false. It does mean the ranking of causes belongs to the sponsor.

🧰 Builder’s Tip

Tool spotlight: query the FDA’s device failure database before you design around a device.

openFDA exposes the MAUDE adverse event database as a free JSON API. No key for light use, no PHI, no BAA, thirty seconds from your laptop:

curl "https://api.fda.gov/device/event.json?search=device.generic_name:\"infusion+pump\"&count=event_type.exact"

That returns the distribution of event types — malfunction, injury, death — for a device class. Swap in pulse+oximeter, ventilator, glucose+monitor, or whatever your tool takes a feed from.

MAUDE’s failure taxonomy is the closest public thing to a spec for what your error handling has to survive. We design the happy path because the unhappy path is invisible. This makes it visible, free.

📅 Upcoming: Using AI Intelligently to Solve Pressing Healthcare Challenges (AHIP/Engagys), Tue Sep 8, 2 PM ET, free.

💡 BTW

💡 BTW: Aron Szanto, who built the AI code reviewer in today’s Big Thing, studied computational economics and mathematical philosophy at Harvard. The person who decided AI errors are enumerable came out of the field that asks what a formal system can and cannot prove about itself. Redesign Health

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

You know the first half of the sentence.

Fortes fortuna iuvat. Fortune favors the bold. It’s on gym walls and roughly four hundred thousand LinkedIn posts.

Almost nobody knows it’s only half of what the man said.

Late summer, 79 AD. Vesuvius is throwing ash across the Bay of Naples. Pliny the Elder is commanding the Roman fleet at Misenum, on the far side of the water. Everyone who can leave is leaving.

His first impulse is curiosity. He’s a naturalist — he has spent his life writing things down, thirty-seven volumes of it, an attempt to record everything anybody knew about the world. He orders a fast boat because he wants a closer look.

Then a message arrives. A woman named Rectina is trapped under the mountain. Her villa is below it. The only way out is by sea.

So he stops looking and launches the warships.

By the time they’re close the ash is falling hot on the decks, the shore is blocked with debris, and the helmsman tells him to turn back.

That’s when he says the line.

Fortes fortuna iuvat. Pomponianum pete.

Fortune favors the bold. Head for Pomponianus.

Pomponianus is a friend. A person. With a name, at an address, in trouble.

That’s the half nobody quotes.

I’ve watched a lot of clinician-built projects die, and almost none of them died from a shortage of boldness.

They died pointed at a category.

Ambient documentation.
Clinical decision support.
Care coordination.
Patient engagement.

Those are volcanoes. You can sail at one for two years and never arrive, because there is nothing there to arrive at.

The ones that live have a name in the second half of the sentence.

The pharmacist who told you she couldn’t hire fast enough.
The night nurse who rewrites the same four things on the same whiteboard every shift.
The resident who asked you twice why the discharge instructions print out of order.

You don’t need a bolder idea. You need someone to steer toward.

Here’s the part I won’t soften.

Pliny doesn’t make it. He gets ashore, stays the night, and dies there. The letter that tells us any of this was written by his nephew, who was seventeen, stayed home with his mother, and lived to write it down.

Fortune favors the bold. It does not guarantee the outcome.

Your thing may die in governance anyway.

That was never the part you controlled. What you control is whether the sentence has a second half.

TL;DR

Boldness aimed at a category is just daring.

Boldness aimed at a person is a direction.

Pick the person. Then be bold.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)

• Peterson says fee-for-service and autonomous clinical AI are a bad combination — pay per unit and you’ll get more units. The same logic decides who keeps the half hour an ambient scribe hands back. (The Big Thing, below.)

• The House VA Committee voted 19–0 to subpoena Larry Ellison after the VA’s EHR ceiling went from $10B to $27B with 17 of 170 sites live.

  🔮 My bet: the number that ends up mattering isn’t $27B, it’s cost-per-site — and someone reads it into the record before Thanksgiving.

• Northwell moved subspecialty match rates from under 50% to over 85% and time-to-intervention from 74 days to 30. The metric changed before the model did.

• 🎧 Podcast: The 229 Podcast — “If I Took It Out Today, Would an Outcome Change?” — three informatics leaders on the scorecard that decides whether your pilot scales, and why a clean 12-month ROI isn’t a required dimension.

🧭 The Curbside

“Payers have to expose a Provider Access API in January. Does that give me anything?”

• Short answer: Yes — a data source you’ve never had, four months out.

• What changed / Evidence: CMS-0057-F requires impacted payers to stand up a FHIR Provider Access API by January 1, 2027, letting in-network clinicians pull their patients’ claims and clinical data from the plan. Four APIs, one deadline. The commercial signal that the plumbing is real is running the other direction first: on September 1, 1upHealth launched a product that lets plans pull clinical data out of provider EHRs over FHIR — sold on Star Ratings, with Capital Health Plan first in.

• Builder read / Watchout: Your EHR-embedded app has never seen the claims history or the prior-auth record. After January it can. But “payer has an endpoint” and “endpoint returns something usable” are separated by a claims-to-FHIR mapping layer — and the plans that are already good at pulling clinical data across that seam are the ones most likely to serve something back.

😤 “Payers will ship a compliant endpoint that returns almost nothing.” Probably, at first. Build against it anyway — the ones who under-ship in January are the ones Star Ratings force to fix it by summer.

🔬 The Big Thing

Who actually keeps the thirty minutes?

The Peterson Health Technology Institute published a warning this week that reads like a footnote and isn’t one: run autonomous clinical AI through fee-for-service rules and you may raise spending without improving outcomes. Pay per unit, get more units.

That’s the whole question the ambient scribe category has mostly avoided.

Scribes are sold on giving clinicians time back, and the early data support it — burnout fell from 51.9% to 38.8% within a month of adoption in a six-system study.

But finance doesn’t buy well-being. They buy conversions: minutes into visits, or minutes into richer coding.

We already ran this experiment. Human scribes spent twenty years inside fee-for-service, and a meta-analysis of 39 studies found throughput up about 0.3 patients per hour and RVUs up 0.55 per hour.

Efficiency that meets a volume incentive gets absorbed as volume. It’s the Jevons paradox — today’s saved thirty minutes becomes the baseline tomorrow’s schedule is built on.

Here’s the part to sit with. One vendor publishes a case study claiming roughly $13,000 in incremental annual revenue per clinician. The strongest measurement to date — a UCSF study of 1,565 physicians across 1.2 million encounters — found about 1.81 RVUs per week, or roughly $3,044 per physician per year.

Fourfold. The time side is wider: a five-center JAMA study measured documentation time down about 16 minutes a day against the 60 to 100 minutes the vendor playbooks assume.

A system that paid expecting $13,000 and is seeing $3,044 has every reason to chase the difference — and once the vendor is paid, closing the gap is the buyer’s problem.

The ratchet bites at one intersection: fee-for-service payment, employed clinicians on productivity targets, documentation-heavy specialties. Remove any one and it weakens. Which is exactly why PHTI is arguing about payment design and not about models.

The number that will matter is burnout the quarter after the targets reset. Nobody has scheduled that measurement.

😤 “This is anti-AI concern trolling. Burnout went down. Take the win.” The burnout number is real and I’d deploy on it tomorrow. The argument isn’t that the tool doesn’t work — it’s that the tool works and the contract decides who collects.

😤 “Nobody is re-baselining anyone’s schedule.” Not yet.

😤 “What’s a builder supposed to do about compensation design?” Instrument for it. Does your dashboard report minutes saved — which finance will convert — or minutes protected, which somebody has to defend? Most tools have never been asked to tell those apart.

📡 Builder’s Radar

Congress just subpoenaed the man whose company runs the VA’s chart

The House Veterans’ Affairs Committee voted 19–0 to subpoena Oracle chairman Larry Ellison and CEO Mike Sicilia after the VA moved to extend its 10-year, $10 billion EHR contract by another $17 billion. Oracle declined to send anyone to testify. Seventeen of 170 medical centers are live since 2020.

Twenty-seven billion dollars, and the software was never the expensive part.

😤 “Government IT is always like this.” Sure. And every builder pitching a health system right now is quoting an implementation timeline built on the same assumption the VA made in 2018.

The agent layer doesn’t dissolve the system of record. It cements it.

Two pieces landed the same day from opposite directions. Brendan Keeler argues antitrust doctrine structurally can’t reach a system of record — market definition is unsettled, B2B share data unobtainable, and the SSNIP test collapses when switching costs dwarf any price increase. His read: information-blocking enforcement is already doing more real work than the FTC’s inquiry will.

John Lee, MD makes the product-side version: when the AI layer becomes the interface, something still has to be the auditable source of truth — and owning that base layer beats owning the interface on top of it.

If both are right, the agent era makes the incumbent harder to displace and easier to regulate. Neither is what the disruption thesis predicted.

They stopped optimizing for the next open slot

Northwell piloted subspecialty matching that moved match rates from under 50% to over 85% and cut time-to-intervention from 74 days to 30 — by measuring “time to appropriate care” instead of “time to next available.”

The model is the boring part. Changing the denominator is the intervention.

🔮 Where this lands: the first health system to publish a mismatch rate next to its wait time will make everyone else look bad on purpose — and I think that happens within a year.

⚡ Quick hits

Thyme Care closed over $125M past a $2B valuation and spun up a parent entity for separate operating businesses — biosimilars and trial accrual first. Navigation was supposed to be a service line. The bottlenecks it routes around turned out to be businesses.

Boston Scientific resumed shipping most products more than a week after the cyberattack that halted distribution. Almost nobody can name which clinical workflow has a single-vendor device dependency until the boxes stop arriving.

🎙️ From the Pods

🎙️ The 229 Podcast — “If I Took It Out Today, Would an Outcome Change?”

Informatics leaders from Children’s Hospital of Philadelphia, Banner Health and ThedaCare describe scorecards where financial ROI is one dimension among four — and where a safety intervention can scale without one.

💡 Builder take: Ask your champion what evidence would make them stop. If they can’t answer, your pilot has no ending, only an expiration.

🔇 Speaker Blindspot: Survivorship bias — every framework described comes from a system mature enough to have one. The systems drowning in vendor pitches are drowning because they don’t.

🎙️ Relentless Health Value — “Why Should a Plan Sponsor Care About the 340B Charity Program?”

Shawn Gremminger walks the plumbing: roughly $68 billion runs through 340B, second only to Part D — and no statute requires hospitals to reinvest the spread into charity care, nor tracks whether they do.

💡 Builder take: If you touch drug acquisition cost, the discount-versus-rebate fight determines your data model. Build it model-agnostic; the rebate pilot has been enjoined once already.

🔇 Speaker Blindspot: Composition fallacy — “hospitals” is treated as one actor, but a rural disproportionate-share hospital and an academic system with contract pharmacies run opposite programs under the same statute.

💡 BTW

💡 BTW: The motion to subpoena Larry Ellison came from Rep. Maxine Dexter, MD — a pulmonary and critical care physician who practiced at Kaiser Permanente Northwest for fifteen years and was the first woman to chair the Northwest Permanente board. The person asking Oracle why 17 of 170 sites are live has personally used an EHR.

You have a unique combination of skills, experience and values. So do great things! … and tell me about them at kevin@clinicians.build.

— Kevin & AI

(please verify content for yourself, partially AI generated and may contain errors)