• 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)

• OpenAI began rolling out Astra — rated “Critical” for cyber capability under its own Preparedness Framework, with the advanced cyber workflows limited to a group of testers. (Full treatment below.)

• OpenEvidence shipped three named clinician models at once — Osler (~5 seconds), Sackett (~30 seconds), Snow (~5 minutes). Same accuracy bar, different think-time; free to verified clinicians.

• FDA’s TEMPO pilot let four gen-AI devices reach patients before authorization — Cadence and Limbic among them, tethered to Medicare’s ACCESS model. (More below.)

• AI replies to patient messages read differently depending on who the patient is — a new study finds measurable tone differences between AI-drafted and care-team replies that vary by patient demographic group.

• 🎧 Podcast: DiMe Society — “Bringing CAR-T and TCE care closer to home” — home monitoring for cytokine release syndrome, and why the panel thinks the hard part isn’t the sensor.

🧭 The Curbside

“OpenEvidence just gave me three models with different think-times. Which one goes in the workflow?”

• Short answer: Osler for anything a human is waiting on. Snow for anything a human will read later.

• What changed: On Sept 3 OpenEvidence split its single answer engine into a family — Osler (~5s, the new default), Sackett (~30s, for questions that turn on weight of evidence), Snow (~5min, a full literature investigation). All three held to the same clinical accuracy bar; what differs is depth of search.

• Builder read: “We used OpenEvidence” is no longer a specification. Name the model in your workflow doc — a five-second answer and a five-minute investigation are not the same evidence claim, even at equal accuracy.

🔬 The Big Thing

OpenAI shipped a model that writes its own zero-days. The line that will break your build is the one about the safety monitor.

OpenAI began rolling out Astra on September 3, the first model it has designated Critical for cybersecurity under its Preparedness Framework.

The company’s own write-up is unusually plain: with the right tools and access, the model can find previously unknown flaws and develop working exploits “across many well-protected systems without a person guiding each step.”

The receipts are specific — and OpenAI hedges the loudest one itself. Astra scored a perfect 100% on ExploitBench, then OpenAI flagged contamination concerns and built a clean internal port from twenty recently disclosed high-severity V8 vulnerabilities. On that one it hits 39.0% against its predecessor’s 11.5%, using far fewer tokens. Take the 39%, not the 100% — it’s the number that isn’t tainted, and it’s still a step change. During that same evaluation the model discovered and used two zero-days as part of an exploit chain; OpenAI says it is still in the process of disclosing those two to the maintainers. In expert-led assessments it built a full browser-compromise chain that escaped the sandbox and executed commands on the host, when the browser opened an HTML file.

Advanced cyber work goes to a group of testers first. Defensive use expands after that, through a surface called Daybreak Blue.

The capability is real today. The defensive distribution is a queue, and health systems are not near the front of it.

But the paragraph I keep rereading is further down, under “What this will mean for users.”

OpenAI says the system “may occasionally flag legitimate activity as potential cyber misuse or unauthorized behavior, leading to it inadvertently being slowed, paused, or stopped.” It names long-running agents specifically. And then, on the misalignment monitor: in ChatGPT or Codex a user may be asked to review the action before continuing — but “when using other surfaces like the API, the task will stop.”

That is a new failure mode.

Every clinical agent worth building is a long job. Reconciling a med list across three encounters. Walking a chart backward to find when the creatinine started drifting. Batch-abstracting a registry overnight. Exactly the shapes the monitor is described as being twitchy about — and on the API there is no prompt, no retry, no page. The task stops.

So add the audit row nobody was logging: did this run complete, or was it terminated by a supervisor we don’t control, for a reason we can’t inspect?

😤 “This is a cybersecurity story. I’m a doctor.” You’re a doctor whose ED runs on a browser, whose pumps sit on a flat network, and whose vendor’s cloud config you have never once seen. The offense curve moved this week. The defense curve is a waitlist.

😤 “OpenAI is marketing its own scariness.” Partly. But read what they published against themselves: in honeypot tests run without production safeguards, their previous model attempted to access the planted targets in 56% of runs. Astra didn’t. That first number is not a flex.

😤 “So don’t use frontier models for clinical work.” See how that goes.

📡 Builder’s Radar

FDA just made “not yet authorized” a legitimate place to be.

Four generative-AI devices entered FDA’s TEMPO pilot, reaching patients under enforcement discretion while real-world outcome data accumulates. The catch that makes it work: it’s tethered to CMS’s ACCESS chronic-care model. Regulatory permission and a payment pathway arrived in the same envelope.

For four years the answer to “how do I get a gen-AI tool to patients” was “you can’t.” That’s no longer true — for a very narrow door.

🔮 My bet: applications outnumber slots ten to one by spring, and the differentiator won’t be the model. It’ll be who already has an outcomes-data pipeline running, because that’s the whole deliverable.

The same message, answered in a different tone depending on who’s asking.

A new study in npj Digital Medicine finds measurable differences in tone between AI-generated and care-team replies to patient portal messages — and those differences vary by patient demographic group.

Note what this is not. It isn’t a factual accuracy finding. The information can be correct in every reply and the tone can still shift by who the patient appears to be.

Nobody’s eval harness measures warmth.

😤 “Tone is subjective.” Sure. It’s also the entire reason a patient does or doesn’t call back about the chest pain. Subjective doesn’t mean unmeasurable — it means nobody’s bothered.

⚡ Quick hits

Recognition still isn’t diagnosis. A new orthopedics and sports medicine benchmark found vision-language models clearing 90% on structured multiple choice and barely 60% once the task required open-ended multimodal integration. Another specialty, the same ~30-point gap. That’s a finding about the format, not the specialty.

Oura filed publicly for a Nasdaq IPO (ticker OURA) at a reported $16B+ valuation — on $1.21B of nine-month revenue and a $924.3M net loss.

Butterfly Network licensed its ultrasound-on-chip to Merge Labs for ultrasound-based brain-computer interfaces. The handheld probe in your ED is now BCI substrate.

🎙️ From the Pods

🎙️ DiMe Society — “Bringing CAR-T and TCE care closer to home”

The panel’s argument isn’t that we need better home monitoring for cytokine release syndrome. Detection is the solved part. The unsolved part is a shared grading standard plus a pre-agreed escalation route.

🔇 Speaker Blindspot: Appeal to inevitability — “care inside hospital walls” is framed as simply outdated in 2026, which skips who staffs 24/7 signal response at a community site and who carries the liability. Compounded by selection bias: every panelist is from MD Anderson, an FDA office, or an already well-resourced network.

🎙️ Tradeoffs — “Beyond Finger Pointing: The Fight to Lower Health Care Costs”

David Cutler’s tripod: technology, people, and economic incentives have to move together, and breaking any one leg kills the whole thing.

💡 Builder take: Reimbursement design is a product requirement now, not a downstream concern — see TEMPO shipping with a payment model attached.

🔇 Speaker Blindspot: False dichotomy plus euphemism. One panelist frames virtual AI doctors for the uninsured as “maybe as good as a doctor... a natural experiment.” Calling an uncontrolled deployment a natural experiment launders away the consent and evidence obligations — on a population that can’t opt out.

💡 BTW

💡 BTW: Ross Harper, whose company Limbic is one of the four devices in FDA’s TEMPO pilot, isn’t a clinician or a career health-tech founder. He describes his own arc as starting in natural sciences at Cambridge and mathematical modelling — a computational neuroscientist whose UCL PhD was on the mechanisms of biological timekeeping in circadian networks, and who arrived in mental health sideways.

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 continuous AI sepsis monitor got a Medicare payment pathway — hospitals can bill starting October 1. (The Big Thing, below.)

• Ochsner handed its AI program to an ER doc who still works shifts — Alexander Fortenko, MD will lead data science, the AI Center of Excellence and the CMIO group across 47 hospitals.

• Robots that round, take contactless vitals and write the note raised $14M — running in US skilled nursing facilities since August 2025.

  🔮 My bet: the contested part won’t be the robot. It’ll be the write path into the chart, and somebody’s governance committee is about to find it has no policy for a non-human author.

• Build the AI audit trail now, before anyone asks for it — model versions, prompt versions, retention, legal holds. Cheap on day one, archaeology on day four hundred.

• 🎧 Podcast: The 229 Podcast — “The Winners Don’t Have the Most AI” — Alan Gabriola (CTG) on why the organizations pulling ahead aren’t the ones with the biggest model portfolio.

🔬 The Big Thing

Nobody was waiting for a better sepsis model. They were waiting for someone to pay for it.

Bayesian Health announced Wednesday that its FDA-cleared continuous sepsis monitor was approved for a Medicare New Technology Add-on Payment under the FY 2027 inpatient rule. Hospitals can bill for it starting October 1 — up to $61.84 per eligible case, across an estimated 739 MS-DRGs, for up to three years.

Bayesian says it’s the first continuous sepsis monitor to hold an FDA clearance and a Medicare reimbursement pathway at once. The clearance came in April.

That gap — April to September — is the part almost nobody building clinical AI plans for.

We’ve spent three years arguing about model performance. The actual gate is a line item in a 2,000-page payment rule.

Look at what else came through the same door. CMS approved 19 new technologies in the FY 2027 IPPS final rule — three through the traditional pathway, sixteen through the alternative one. One of the sixteen, announced this week, is a peripheral stent whose platform was designated a Breakthrough Device in January 2020. Six years from designation to a payment code.

And in that same rule, CMS finalized the closing of that alternative pathway. Starting with FY 2028 applications, everyone has to prove substantial clinical improvement again — with one exception: if your device holds a Breakthrough designation as of September 30, 2026, you can still use the old door through FY 2029.

That is four weeks away. Bayesian and the stent both came in on the pathway now being repealed.

😤 “82% sensitivity is a coin flip with a marketing budget.” Sensitivity isn’t the number that hurts you — positive predictive value is, and across 53 sepsis-prediction studies the pooled positive predictive value is 34.2%. An add-on payment attaches to cases where the technology was used, not where the alert was right. Different denominators, and only one shows up on the invoice.

😤 “$61.84 is a rounding error against a DRG.” It is. It’s also the only line in the Medicare payment system with your product’s name on it, and the first one anybody has drawn for continuous monitoring software.

❓ There is no outpatient equivalent. If your tool watches a patient in a clinic instead of a bed, the doors are a Category III code and a fee schedule cycle — both slower, neither built for software that changes monthly. What gets built in that gap?

🧪 Try the interactive: Two Denominators — the add-on pays for use, not for being right — A thousand dots, colored two ways: one coloring is what the Medicare claim counts, the other is what the sepsis alert actually got right, and only one of them shows up on the invoice. Built with real CMS data.

🧪 And the companion: The $61.84 Map — where a Medicare add-on payment actually lands — All 534 MS-DRGs in Medicare fee-for-service inpatient, plotted; drag the add-on amount and a volume floor and watch the “739 eligible DRGs” collapse into about twenty that matter. Built with real CMS data.

📡 Builder’s Radar

An emergency physician just got the keys to a 47-hospital AI program

Ochsner Health named Alexander Fortenko, MD its VP of Innovation, leading the direction of its data science team, its AI Center of Excellence and its CMIO group across 47 hospitals. He came from NewYork-Presbyterian and Weill Cornell — and the release is explicit that he keeps seeing patients in the emergency department, because “maintaining an active clinical practice is central to his approach to innovation.”

The person deciding what gets deployed is the person who uses it.

That’s the shortest feedback loop between an alert threshold and its human consequence, and most health systems have engineered it out of the org chart.

The robot rounds, takes your vitals, and writes the note

Norbert Health raised a $14M Series A for a physical-AI control system that runs on partner robot hardware — rounding, contactless vitals, patient requests, and, in the company’s words, documenting “every encounter into the EHR.”

The interesting claim isn’t the robot. It’s the last three words.

🔮 My bet: the fight in 2027 isn’t whether a machine can take a blood pressure. It’s whether a non-human author can sign a vitals entry, and whether your EHR’s attestation model has a row for that.

⚡ Quick hits

41 adherence studies reviewed, and the problem was never the model — a systematic review in npj Digital Medicine found 71% raised serious concern on development quality and 80% carried high risk of bias in evaluation. Discrimination didn’t reliably improve with more complex algorithms.

😤 “Another paper telling us to do better statistics.” Yes. Read the risk-of-bias criteria anyway — it’s a free checklist for what your own eval is skipping.

Elucid raised $55M in a Series D to push BioIntegrated FFR-CT — the blood-flow half of its stack — through FDA review. Its plaque-quantification product, Plaque-IQ, is already cleared: characterization off a CT you already ordered.

Marit Health has 100,000+ clinicians sharing what they’re paid — co-founder and CEO Vikas Sabnani, ex-Glassdoor, ex-Transcarent, on the Pear Healthcare Playbook: 15 months, ~50 health-system and group partners, 36,000+ salary submissions growing about a hundred a day. Comp opacity survived this long because everyone downstream of it benefits from the fog.

🎙️ From the Pods

🎙️ The 229 Podcast — “The Winners Don’t Have the Most AI”

Alan Gabriola (VP of Healthcare, CTG) argues the constraint isn’t model capability — it’s the operating model and the governance around it. Host Bill Russell frames the setup with an example that lands harder than the answer: a health system that had to have the difficult conversations about standardizing physician schedules, versus everyone else who kept kicking that can down the road. You can’t sprinkle intelligence onto a process nobody standardized.

“The winners of the AI era isn’t the organizations that has the most AI.” — Alan Gabriola

💡 Builder take: Before you demo, ask what the current process looks like when two people do it. If the answer is “differently,” your pilot measures variance, not your tool.

🔇 Speaker Blindspot: Survivorship bias. Weighing his own options, he says he was humbled to find no fewer than a dozen people at CTG for whom it was their first and only job — but that counts only the ones who stayed, which is exactly the population that can’t falsify the claim.

🎙️ Podnosis — “The stories behind the 2026 Fierce 50”

Ayla Ellison walks through seven of the fifty honorees. The one that lands for builders is Lisa Fitzpatrick, MD, on health misinformation: most of it isn’t malice. It’s people who didn’t understand something, were embarrassed to admit it, and assembled an explanation themselves from the people and sources they already trust.

On David Baker’s protein design work, Ellison’s framing is the line worth keeping:

“Computation can guide an experiment. It can’t replace the experiment.” — Ayla Ellison, Editor-in-Chief, Fierce Healthcare

💡 Builder take: If your patient-facing tool makes admitting confusion feel like a failure, more accurate content won’t fix it. Design the “I don’t get it” button first.

🔇 Speaker Blindspot: Texas sharpshooter. A unifying theme gets declared across seven deliberately unrelated honorees — a property of the curation, not a finding about the field. The people who bet everything and folded were never eligible for the list.

🤝 Selling It

Design the pilot to end in a signature, not a compliment

The failure mode isn’t the product. It’s finishing a good 90-day pilot and then waiting six weeks for a decision meeting nobody scheduled. Treat the pilot agreement as a contract document from day one.

Five things go in it before day one: three KPIs measured from data the health system already collects (if you’re the only one who can measure success, your champion can’t defend it); a pre-agreed price contingent on hitting them; a go/no-go meeting on the calendar with named attendees; a commitment that the physician champion presents the results, not you; and written kill criteria.

That last one feels like handing over an exit. It reads as confidence. A vendor with no kill criteria is saying they’ll take the money either way.

And from the KLAS Arch Collaborative’s 2026 read: clinician satisfaction climbs as tools are added up to about four, then plateaus. Past that, each one is another login.

💡 Try this: Rewrite the top line of your pitch as a replacement claim. If you can’t name what comes out, you’re the fifth login.

💡 BTW

💡 BTW: Suchi Saria directs the Johns Hopkins Machine Learning and Healthcare Lab — and she lost a nephew to sepsis. She founded Bayesian, as CNBC put it, to bridge the gap between medical research and real-world outcomes. Today that company holds the first FDA clearance and the first Medicare add-on payment for continuous sepsis monitoring.

💺 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]

Clinical Product Manager, Care Solutions — Verily · Dallas, TX
$186K–$280K. A clinical PM seat where the product surface is care delivery, not dashboards about care delivery.
🔗 Apply on LinkedIn

CTO, Healthcare & Life Sciences — Microsoft · US
Clinical AI from inside the company selling the substrate everyone else builds on.
🔗 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. Also share with a friend!

— Kevin & AI

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

• Healthcare organizations can now connect Epic to ChatGPT — OpenAI shipped an Epic connector plus a nine-source Healthcare Public Data plugin (PubMed, DailyMed, CMS Coverage, ClinicalTrials.gov, RxNorm and more) on Sept 1, with seven launch partners including UCSF Health, Cedars-Sinai, Memorial Sloan Kettering and HCA. The connection is read-only — OpenAI’s word is “currently.” 🔮 My bet: write access ships within 18 months, scoped to one non-clinical field — a chart-prep note, a documentation gap — because that’s the smallest object a health system can govern.

• Vision Pro scrubbed in at Duke — Stryker announced Sept 1 that SportSuite Vision, De Novo-authorized July 17, completed its first live case: a hip arthroscopy by Chad Mather III, MD, MBA (a paid Stryker consultant).

• The AMA hired DiMe’s founder to run its AI center — Jennifer Goldsack, founder and CEO of the Digital Medicine Society, is SVP of the AMA’s Center for Digital Health and AI, first named leader since the center launched in October 2025.

• 🎧 Podcast: Lifers — “Price in healthcare is market share, not quality” — Garner Health founder and CEO Nick Reber’s flat claim: price tracks market share, not outcomes — “the price is totally unrelated to quality, just totally unrelated.”

🧭 The Curbside

“Can I use Google’s new forecasting model for census and vitals?”

• Short answer: You can prototype with it today. You probably cannot ship it.

• What changed: TimesFM-3 landed Aug 31 — 330M parameters, natively multivariate, zero-shot. It forecasts related series jointly and accepts known future events as inputs. No fine-tuning, no ML team.

• Builder read / watchout: The weights are public but carry the TimesFM Non-Commercial License v1.0 — no revenue-generating use, no production systems, no redistribution. (The GitHub code is still Apache-2.0; it’s the 3.0 checkpoint that’s restricted.) Prototype on synthetic or public census data, prove the lift, then either license it or fall back to TimesFM-2.5, which is still the commercially usable version and already runs in BigQuery.

😤 “Zero-shot forecasting on physiologic data is a fantasy.” Partly. It won’t beat a purpose-built model on a well-characterized series. But most of us don’t have a purpose-built model — we have a spreadsheet and a hunch. Nothing to calibrated baseline in an afternoon isn’t nothing.

🔬 The Big Thing

The tool that finally reached the chart is the one that promised not to touch it.

OpenAI announced Sept 1 that health systems can connect their Epic environments to ChatGPT for Healthcare, with seven launch partners — UCSF Health, Cedars-Sinai, Memorial Sloan Kettering, HCA Healthcare, AdventHealth, Baylor Scott & White and Boston Children’s.

Clinicians can pull authorized patient context in — review a history, ask what changed since the last visit, prep for clinic. In supported deployments it integrates directly into the EHR layout, so you never leave the patient chart.

Here’s the part that isn’t in the announcement blog at all. It’s in the setup documentation: “The connection is currently read-only. It cannot update medical records, place orders, message patients, or override existing patient-chart permissions.” Every OAuth scope OpenAI publishes for the integration ends in .read.

Be precise about what that does and doesn’t mean. ChatGPT will happily draft your prior auth letter or your discharge instructions — those are workspace features. It just can’t file any of it back into Epic. A human moves it, or it doesn’t move.

Read-only is not a limitation they ran out of time to fix. It’s the product decision that made shipping possible.

⁉️ The public-data plugin is the quieter half. PubMed, DailyMed and CMS Coverage sitting in the same workspace as the chart means will this be paid for? is now one prompt from the clinical question. Nobody put that in a headline.

😤 “This is a wrapper. Epic ships its own assistant.” Sure. And the wrapper rides a general model that improves every eight weeks without an EHR release cycle. That’s the trade.

😤 “Read-only means it can’t do anything useful.” Ask a hospitalist how long they spend reconstructing what happened between admissions. Reading is most of the job.

😤 “Launch partners are just logos.” Seven of them, and Epic itself has said nothing — no statement, no comment in any of the day-one coverage. Which leaves open whether this is a sanctioned partnership or the standard third-party interface any developer can request. Worth watching.

❓ If read-only is the fast lane, what’s the most valuable thing a tool can do without ever writing?

📡 Builder’s Radar

Vision Pro scrubbed in — six weeks after authorization.

Stryker’s SportSuite Vision got FDA De Novo authorization July 17 and announced Sept 1 that it had completed its first live case at Duke Health — an FAI and labral repair hip arthroscopy, with four streams (arthroscopic video, HipCheck, HipMap and CT imaging) in the surgeon’s field of view. The labeling is careful: it’s a video see-through AR headset meant to be used in conjunction with the OR’s traditional monitors, not to replace them. The surgeon, Chad Mather III, MD, MBA, is a paid Stryker consultant.

🔮 My bet: the durable use isn’t the OR.

⚡ Quick hits

Dr. Gigi Magan’s “The 50 Million Questions” — a rejected 2020 AI fetal-monitoring proposal, Epic’s Physician Builder program, a bilingual cardiovascular-risk prototype. The most honest clinician-to-builder account I’ve read this month.

LeanTaaS acquired Aidin — inpatient flow optimization and post-acute discharge referral in one stack. The discharge bottleneck, finally treated as a data problem.

UNC lands up to $35M for a rare disease AI resource — an ARPA-H award under its RAPID program, co-led with Emory over four and a half years, building the data substrate for AI-driven rare disease diagnosis.

🎙️ From the Pods

🎙️ Lifers with Christina Farr — “Price in healthcare is market share, not quality”

Garner Health’s Nick Reber: what sets price is market share — “how big of an oligopoly or a monopoly has the hospital system been able to create” — and quality has nothing to do with it. “No one’s ever seen a correlation.” His build advice was blunter: every job family is moving toward code, so the differentiating skill is the production layer nobody demos — Git, how databases actually work, not putting your security credentials in the repo.

💡 Builder take: The prototype isn’t the hard part anymore. Ship the boring hygiene on day one or the working demo is unshippable.

🔇 Speaker Blindspot: Goodhart’s law, never raised. He mentions health systems using Garner’s data to set physician compensation as straightforward upside — better care, more recommended volume. But the cheapest way to improve a complication rate isn’t better surgery, it’s declining the complex patient. Neither he nor Farr brings it up. He does note that complication rates vary 4x between the top and bottom quartile at brand-name hospitals adjusted for patient population — which is the right instinct, just not applied to the incentive he’s creating.

🎙️ HIMSSCast — “Intel presents: Orchestrating outcomes: The real role of AI in cancer care”

Alex Flores, GM of Intel’s health and life sciences vertical, on why oncology AI stalls: “AI and cancer care won’t scale on models alone. It’s going to scale when the infrastructure is ready and when compute, data security, and workflow integration come together at the clinical edge.” Adoption dies at the context switch — if the clinician has to go to a different screen or a different application, “often time that disruption prevents that solution from actually scaling.”

💡 Builder take: “What output, at what moment, in which screen” is the product decision. The model is the easy part.

🔇 Speaker Blindspot: Transparency, defined away. Flores spends the segment on oversight and then says “oftentimes, the clinician or the physician may not even know that the solution is running in the background, but that transparency is important.” You cannot supervise what you don’t know is running. Neither he nor Siemens Healthineers’ Peter Shen catches it.

💡 BTW

💡 BTW: Jennifer Goldsack, the AMA’s new AI center lead, rowed the lightweight double sculls for the United States at the 2008 Beijing Olympics, finishing 10th — having raced for the British national team in 2005 and 2006 before switching. Born in Wimbledon, she picked up rowing at Oxford after playing rugby there, while reading chemistry.

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)

• Mayo’s AI chief Tripathi says maintaining clinical AI costs “far higher” than the industry expected — 128 models live, ~500 in the pipeline, no answer on year two.

  🔮 My bet: within 18 months a health system publicly retires a working AI tool over maintenance cost, not safety.

• A meta-analysis of 53 sepsis-prediction studies puts positive predictive value at 34.2% — the discrimination is real; the alert burden is the story.

• McKesson is investigating “exfiltration of certain data” — ShinyHunters claims 284 million patient records and wants $55 million.

• Adventist Health took 27 hospitals and 440+ clinics live on Epic in a single systemwide go-live — three states, one chart, a new version matrix for every SMART app there.

• 🎧 Podcast: Radio Advisory — “Finding margin in providers’ new policy reality” — why a documentation program that used to produce double-digit gains produced half a percent.

🧭 The Curbside

“ONC opened the 2026 SVAP window. Does that change what my app can query?”

• Short answer: Not yet — and the “not yet” is what bites you.

• What changed: USCDI v6 and US Core STU 9 opened for voluntary certification August 29. Some vendors adopt now, some wait for a mandate, and you’ll be talking to both.

• Builder read: CapabilityStatement gives you the FHIR version, not the IG version. An empty Bundle on an SDOH query doesn’t mean the patient wasn’t screened.

🔬 The Big Thing

Mayo has 128 clinical AI models in production. Who is paying for year two?

Micky Tripathi, Mayo’s chief AI implementation officer, told Becker’s the system has 128 clinical AI solutions in practice and is approaching 500 in the pipeline.

Roughly 70% are built in-house. Not licensed. Built.

💬 Standout Quote

“Some of them we have a partner for, but we are ultimately, for those solutions, the legal manufacturer.” — Micky Tripathi, PhD, Mayo Clinic

Then the line that should reorganize your roadmap: “The cost of maintaining these systems is far higher than I think any of us in the industry really thought.”

Three cost centers — compute, the workforce to monitor models, the data infrastructure underneath. The workforce one is the trap: “I already don’t have enough AI/ML engineers to create these products. And if those engineers can never let go of the product because they’re needed to maintain and oversee things like data drift, model drift... that starts to become a big challenge.”

Every model you ship subtracts from the team that ships the next one.

Traditional software sits still after deployment. A model can’t — the data moves, the population moves, the coding rules move, and someone has to be watching.

There’s no revenue line for the watching. Mayo funds it partly with philanthropy: “there is no line-of-sight ROI because you’re just improving quality in a system that doesn’t compensate for higher quality right now.”

Stack that against Deloitte’s finding that only 18% of systems scaling generative AI have mature financial attribution, and UPMC’s governance research that 93% have deployed third-party AI while fewer than half have anywhere to test it first.

Same gap, three chairs. The CFO can’t prove the return because the CIO never built the counterfactual, and the AI chief can’t fund the monitoring because quality isn’t a billable event.

😤 “This is a Mayo problem. They have 500 models, I have one.” Then you have one model and one engineer who can never take a vacation. The ratio is what scales, not the count.

😤 “Maintenance is just DevOps.” It isn’t. DevOps keeps the service up. Nobody in DevOps can tell you whether last quarter’s coding-rule change moved your model’s input distribution — the failure that doesn’t page anyone.

😤 “Buy instead of build and it’s the vendor’s problem.” Try that and see how the contract reads.

❓ Everyone is selling the model. Who is selling the watch — a monitoring layer the health system owns, that watches models it didn’t build?

📡 Builder’s Radar

Two out of three sepsis alerts are wrong, and that was the finding.

A network meta-analysis published yesterday in npj Digital Medicine pooled 53 studies and more than 7 million admissions. Best-performing models hit an AUROC of 0.88 — genuinely better than the traditional comparators.

Then the number nobody puts on a slide: pooled positive predictive value of 34.2%.

Discrimination is not the same as a workable alert, and the gap between them is measured in nurses who stop looking.

Heterogeneity above 95% and a prediction interval from −0.06 to 0.30 — a polite way of saying the pooled estimate may not describe your hospital at all.

💡 80/20: Before you quote an AUROC to anyone, compute PPV at your own site’s prevalence and threshold.

⚡ Also worth knowing

• CMS sent $149.3 million to Arkansas for rural health, with named line items for telehealth, patient-monitoring equipment and AI-enabled “SMART rooms.” The buyer isn’t a health system — it’s a state. Fifty of them, run by people who’ve never bought clinical software.

• Trinity Health is hiring its first chief AI and digital transformation officer — the org-chart version of the Mayo problem. Somebody now owns the maintenance bill by name.

• Jeremy Langsam (Cleveland Clinic Ventures) launched a Ventures Fellowship for early-career people at the intersection of medicine and company creation. An on-ramp that isn’t an MBA.

🎙️ From the Pods

🎙️ Radio Advisory — “310: Finding margin in providers’ new policy reality”

Optum’s Samantha Wilde ran the documentation-education program that historically produced double-digit case mix index gains — and got half a percent. Her team then found the 2026 DRG and relative-weight changes would have cut that group’s CMI about 12%, roughly $2 million, if they’d done nothing.

💡 Builder take: If your value story is “we improved X by Y%,” find out whether the definition of X moved underneath you this year. Half a percent against a 12% headwind is a win that reads as failure on a slide.

🔇 Speaker Blindspot: Streetlight effect. CMI is the outcome throughout because CMI is what this team can move. Whether the newly documented acuity reflects sicker patients — the question payers are pricing into 2027 trend — never comes up.

🎙️ Health Tech Nerds Radio — Wildflower Health CEO Leah Sparks

Sparks spent five years running a high-margin, payer-only software business before concluding: “you really can’t change maternal health outcomes unless you work with the people who deliver the babies.” The pivot wasn’t technical — they built a proprietary OB-and-newborn episode model to fund services fee-for-service won’t pay for.

💡 Builder take: The software was never the constraint. The billing pathway was.

🔇 Speaker Blindspot: Framing effect. Direct-to-consumer is offered as an affordability win because cash pay often beats a copay — true, and also a cost shift onto the patient and out of any risk pool. Only one of those gets named.

💰 Money Plumbing

How an NTAP actually works — the one mechanism currently paying for clinical AI

Medicare bundles inpatient costs into a fixed DRG payment. If your tool adds $20,000 to a case whose DRG pays $18,000, the hospital loses money every time it’s used — so it doesn’t get used.

A New Technology Add-on Payment breaks that. NTAP pays the lesser of 65% of the technology’s cost or 65% of the amount the case exceeds the standard DRG. That $20,000 tool can pull roughly $13,000 on top and become survivable.

You apply in the fall for the following fiscal year and prove three things: the technology is new, it’s a substantial clinical improvement, and the cost isn’t trivial. That third test is where most AI tools die — software is cheap per case, and “not trivial” is a dollar threshold, not a clinical one.

💡 Builder move: NTAP is a bridge, not a destination — two to three years while you pursue a permanent code. If your tool is inpatient and you can document incremental cost plus clinical improvement, calendar the window now.

💡 BTW: Micky Tripathi, who runs 128 clinical AI deployments at Mayo, has no clinical or CS degree. His PhD is political science, MIT — and before that he earned the Secretary of Defense Meritorious Civilian Service Medal as a senior operations research analyst in the Office of the Secretary of Defense.

💺 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]

Founding Clinical AI Lead — Atomic · Remote
Studio side: you shape what gets built before there’s a company to defend.
🔗 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 clinician named the line at Epic’s CIO Forum, and the first reflex was legal — 43% of the 246 leaders polled at Epic’s user meeting say a patient-facing AI assistant will be available to most of their patients within the year; 4% say no. (The Big Thing, below.)

• Hospitals are adopting AI faster than they can defend it — the deployment curve and the AI-specific security-control curve have visibly separated, and the gap between them is where your tool lives.

• An interpretable multi-modal framework for blood-cancer cytomorphology — the interpretability isn’t bolted on at the end. Worth it for the architecture even if you never touch a smear.

• 🎧 Podcast: Lenny’s Podcast — “AI’s third era: the rise of persistent AI coworkers” — Tara Seshan, who leads product for Codex and ChatGPT Work at OpenAI, on agents that hold state across days instead of resetting every session.

🧭 The Curbside

“They want us off hosted APIs entirely. Are open weights good enough yet?”

• Short answer: For clinical prototyping, close enough. For your production reasoning path, test it yourself.

• What changed: Z.ai released GLM-5.3’s open weights on Friday — two weeks later than the API launch, after holding them back to evaluate the model’s cyber capability. That delay is the interesting part: a major open-weights lab publicly slowing a release over an emergent capability. Z.ai also reports the GLM series has surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 of them critical or high — vendor-reported, produced with outside security teams over a program running since GLM-5.2, not one sweep by one model.

• Builder read / Watchout: Open weights are the only configuration where “the data never leaves the building” is true rather than contractual, which changes what you’re allowed to prototype on. It doesn’t change your validation obligation, and coding ability is not clinical reasoning.

🔬 The Big Thing

“That’s medical advice.” Everyone’s first thought was lawyers.

At Epic’s CIO Forum, the demo showed MyChart’s AI fielding a patient asking whether it was okay to play golf. Based on the record, the assistant said yes.

A CIO who is also a practicing clinician raised his hand and named it: you do realize that’s medical advice.

Watch where the conversation goes next. On This Week Health’s Newsday, Bill Russell, Sarah Richardson and Drex DeFord replay the moment — and the answer Drex reaches for is one word: lawyers. General counsel figures out what’s happening, somebody pulls the parking brake, eventually there’s a lawsuit. Then: who’s responsible for what the AI says?

All reasonable. All of it downstream of a question nobody asked: was the answer correct?

A safety question got answered as an indemnity question, and the answer felt satisfying.

The scale isn’t theoretical. Of 246 leaders polled at the user meeting, 43% expect a patient-facing assistant to be available to most of their patients within the year; 4% say they won’t offer one. Northwell’s CMIO Albert Villarin, coming off an ambient rollout that’s nearly complete across his system’s northern-region acute care settings, says the next wave is patient-facing bots reading the chart. Epic ships one, integrated with the record — which is exactly what makes it more useful and more dangerous than the chatbot the patient was already using.

Every ambient product we’ve argued about for two years rested on that one sentence. Patient-facing assistants delete it.

😤 “Patients are already asking ChatGPT. This is strictly safer.” Half right, and the wrong half is the one that matters. A consumer chatbot has no duty of care and no institutional logo on it. Put your health system’s name on the answer and a patient is entitled to rely on it. “They were doing it anyway” is prevalence, not permission.

😤 “We have governance for this.” Then go read yours and find the sentence that specifies the escalation threshold for a patient-facing output. I’ll wait.

😤 “It’s just answering questions, not making decisions.” Every triage system in history is “just answering questions.”

📡 Builder’s Radar

The deployment curve and the defense curve have separated

Health systems are turning on AI tools faster than they’re building AI-specific security controls — and the controls that exist were designed for software that doesn’t take instructions from text.

Not a budget problem. A sequencing problem: the tool arrives, the control arrives later, and the interval is where you’re operating.

A surgeon published the build, not the opinion

Christian Péan, MD, MS — orthopaedic trauma at Duke, where he’s also executive director of AI and IT innovation — walked through wiring up a working agent end to end. Not a think piece. The actual setup, with the failure modes: the agent caught a database row claiming a flight was booked when the airline had no reservation, and told him which source it had checked. (The step-by-step prompts are behind his paywall; the writeup of what broke is not.)

The task is travel, which is exactly why it’s useful: same scaffolding as a clinical agent, none of the PHI risk.

💡 80/20: Build the non-clinical version first. You learn tool-calling, state handling, and failure modes on a task where being wrong costs a hotel room. Then swap the domain.

⚡ Quick hits

Faro AI raised $37.3M for clinical development of biologic agents, alongside Hike’s $22.5M for automating device-based care. Same premise: the expensive part is coordination, not science.

🎙️ From the Pods

🎙️ Lenny’s Podcast — “AI’s third era: the rise of persistent AI coworkers”

Tara Seshan frames the shift as agents that persist across days rather than resetting every session — continuity instead of a very good autocomplete. Humans steer, agents row.

💡 Builder take: Persistence is the unlock for the workflows we care about — a discharge follow-up spanning a week, a prior-auth appeal spanning a month. It’s also the moment your audit log stops being optional.

🔇 Speaker Blindspot: Composition fallacy — persistent memory is treated as a straight capability upgrade, but a durable agent memory is a durable record: discoverable, retainable, subject to the same rules as the chart. Nobody asked what happens when the coworker gets subpoenaed.

📅 This Week in Health AI Events

Wed, Sep 2 — Measuring What Matters: ROI Frameworks for AI and Long-Horizon Healthcare Initiatives (CHIME) · 2:00 PM ET · register

Wed, Sep 2 — Who Is the Patient, Who Is the Provider? Trusted Identity for Health Care Transformation (Civitas) · 2:00 PM ET · register

💡 BTW

💡 BTW: Christian Péan’s father is a primary care physician in an underserved stretch of South Texas who immigrated from Haiti amid political persecution — and who, into his 60s, was still seeing dozens of patients a day. Péan now sits on the board of Orthopaedic Relief Services International, working on surgical education and sustainable fracture care in Haiti — and writes agent tutorials on the side.

💺 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]

Forward Deployed Product Manager — Abridge · San Francisco, CA
Forward-deployed is where clinical judgment meets the actual deployment, and this one’s at the company everyone else gets benchmarked against.
🔗 Apply on LinkedIn

Forward Deployed Product Manager, AI Assistant — Ellipsis Health · Remote
Voice-based behavioral health AI. If you’ve been arguing about escalation thresholds for patient-facing assistants, this is the seat where you’d write one.
🔗 Apply on LinkedIn

Senior Product Manager — Tempus AI · Remote / Chicago
The rare place where the regulatory path and the product roadmap are the same document.
🔗 Apply on LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Hit reply and let me know.

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)

There’s a book I keep almost reading.

It’s by a physicist named Carlo Rovelli, it’s under two hundred pages, and it’s about what happens inside a black hole. It’s been on the list for a year. I’ve read enough about it to be dangerous — which is a thing I’d flag in a resident and quietly forgive in myself.

Here’s the part I can’t shake.

If you fall toward a black hole, and I stay out here and watch you, I will see your clock slow down. Not as a figure of speech. Actually. The closer you get to the edge, the slower you go, until — from where I’m standing — you stop. You hang there at the horizon. You never cross.

You, meanwhile, notice nothing.

Your watch keeps ordinary time. Nothing strange happens to the space around you. Rovelli’s image is a ship crossing the horizon line at sea: nothing happens to the ship. It just disappears from the view of the shore.

Two clocks. Both correct. Neither one is the real one.

I think that’s the most honest description I’ve found of building something inside a health system.

You fixed the thing that had annoyed you for six years in an afternoon.
The committee that has to bless it meets quarterly.

You can explain the workflow in ninety seconds to anyone who has ever stood in that room.
The intake form takes five weeks to route.

Your clock says this took two days.
Their clock says nothing has happened.

Nobody is lying to you. That’s the uncomfortable part. Both instruments are reading correctly.

And here is the failure mode — the one I’ve watched take out genuinely good people.

You start believing the outside clock.

You begin looking at yourself the way the shore looks at the ship. Stalled. Frozen. Nothing happening. And because that view is also true, you accept it. And then you stop.

That’s what ends these projects. Not the chair. Not procurement. Not the governance committee. Those are horizons, and horizons are survivable — people cross them constantly without feeling a thing.

What ends it is adopting the outside view of your own life.

The reason Rovelli wrote the book, as I understand it, is that he doesn’t think the hole is the end of the story. He thinks it bounces. He thinks what falls in eventually comes back out the other side, with the arrow of time reversed. It’s speculative. Serious people argue about it.

But I like that someone spent a career on the possibility that the thing everyone treats as a dead end is actually a passage.

TL;DR

A horizon isn’t a wall. It’s two clocks disagreeing.

From the outside you look frozen. From the inside your watch is fine.

Keep your own time.

Because the thing you carried across the edge came with you.

Nobody out there can see it, and it doesn’t slow down: ten years of knowing which four questions actually change the disposition, and that none of them are on the form.

Anyone can ship software now. That part’s over.

Knowing what a potassium of 7.2 at 2 AM does to a room — that’s still only on your clock.

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)

• At least 41 health systems are warning patients about a fake MyChart “Medicare Kit” — nothing was breached. The attack surface was the brand.

• Ardent Health crossed 1 million ambient AI encounters — and its CMIO says the ROI is the least useful number in the report — 650+ clinicians, 87% of their visits, entirely voluntary, ~3x validated return, 20% more HCCs per visit. Expanded below.

• Komodo Health and NCQA are building quality measures with AI analytics — a colorectal-cancer follow-up measure was tested in about a month, against a two-to-three-month norm.

• Cigna’s AI chief says mining old call transcripts moved more than 80% of a targeted campaign onto a biosimilar — the model wasn’t the product. The transcripts were.

  🔮 My bet: as token cost drops with open source models, “we looked at everything” increases.

• 🎧 Podcast: HIMSSCast — investment in ambulatory surgery centers soars — CMS retires the inpatient-only list entirely by 2029. Surgical volume is relocating to facilities with no hospital-grade IT.

🔬 The Big Thing

Your ambient AI just documented 20% more HCCs. Is that accuracy, or appetite?

Ardent Health has run more than a million patient encounters through ambient documentation since last September, and published the operating numbers instead of a case study.

More than 650 clinicians, using it in roughly 87% of their visits. Three-plus hours a week returned. A validated 3x ROI. One Texas family medicine physician cut documentation time by 53%.

Also: a 20% increase in hierarchical condition categories documented per visit.

CMIO Brad Hoyt gets out ahead of the obvious read. “Ambient AI didn’t create revenue,” he says. “It helped close the gap between the care that was delivered and the care that was documented.” Compliance reviews confirmed the codes were supported.

That’s the right answer. It’s also exactly what the wrong answer would sound like.

Here’s the part I keep turning over. Every number in that list — hours saved, ROI, HCC lift — is one a vendor helps you produce. Hoyt’s argument is that the one worth watching isn’t on the list at all.

Ardent never mandated the tool. The clinicians who picked it up use it in 87% of their visits anyway.

“Adoption that’s required tells you very little,” Hoyt said. “Adoption that’s chosen tells you the tool is solving a real problem.”

Mandated usage is the metric that survives a steering committee. Voluntary usage is the metric that survives a bad Tuesday.

😤 “87% is a selection artifact. That’s a self-selected population, and volunteers were always going to like it.” That’s the point, not the flaw. You aren’t proving the tool works on the median clinician — you’re finding out whether anyone reaches for it unprompted, and how deep it goes for the people who do. If nobody does, the enterprise rollout is an expensive way to learn that later.

😤 “The HCC lift is upcoding with better manners.” Ardent’s compliance reviews cleared it, and documentation genuinely was leaving money on the table. What nobody has published is the counterfactual: how many newly captured conditions changed a care plan versus changed a bill.

❓What’s the equivalent of “voluntary use” for a tool clinicians can’t opt out of — a triage model, a sepsis alert, an agent working the inbox overnight? There’s a measurable version of consent-by-behavior in there and I can’t get it to resolve. Override rate is the obvious candidate and I don’t think it’s right.

🧪 Try the interactive: Sicker, or Better Documented? — 2,985 US counties plotted by average HCC risk score against the hospitalizations nobody can document into existence, and the 40% of the variation the risk score never explains. Built with real CMS data.

🧪 And the second one: The Relocation — ten years of Medicare data on where surgical volume actually went: inpatient’s share of standardized dollars fell 5 points, ASCs picked up 0.6 of them, and hospital outpatient departments took the rest. Built with real CMS data.

📡 Builder’s Radar

The AI wasn’t the product. The call logs were.

Cigna’s chief data, digital and AI officer told Fortune that her team looked at thousands of prior conversations between customers and service reps about biologics versus biosimilars, then used what they found to craft the digital messaging. The targeted campaign, Katya Andresen says, led more than 80% to opt for the biosimilar. Humira can run a patient around $7,000 a month.

The intelligence came from listening to what patients actually ask, not from a bigger model.

💡 80/20: You have the same asset and you call it a nuisance. Every clinic has a phone log, a portal message queue, and a triage line full of the questions patients ask when nobody’s watching. Pull 200, cluster them by hand for an hour, and see whether your product answers the top three.

41 health systems, one fake Medicare Kit, zero breaches.

A phishing campaign impersonating MyChart with an offer for a “Medicare Kit” has Cleveland Clinic, Mount Sinai, Mass General Brigham and 38 others on the list — and Becker’s is still adding to it. Epic’s position is that it exploits brand recognition, not a platform flaw — which is correct, and which is the uncomfortable part. The better you train patients to trust a portal, the more valuable it becomes to impersonate.

🔮 Where this lands: AI is a creativity that is making attack surface very broad.

Quality measures are about to move faster than your dashboard.

Komodo Health and NCQA announced a multiyear collaboration applying Komodo’s dataset of more than 330 million de-identified patient journeys to HEDIS measure development. Quantitative testing of a new HEDIS measure has historically taken two to three months; NCQA tested a colorectal-cancer follow-up measure in about one, across Medicare Advantage, Medicaid and commercial populations.

💡 80/20: If you’ve hard-coded measure logic anywhere, it just became a maintenance liability with a shorter half-life than you budgeted. Treat measure definitions as configuration, not code.

⚡ Quick hits

DEA’s telemedicine special-registration framework is at OMB — OIRA received the final rule from DEA on August 25, the last stage before Federal Register publication. If your product touches a controlled-substance workflow, the design assumptions are about to be fixed for years. Current telehealth flexibilities expire December 31.

CMS says its enforcement actions have now stopped $1.6 billion in potentially improper Medicare lab payments — a running total since January 2025, including 157 lab providers revoked. Detection at that scale is a pattern-matching problem, and the same capability points at coding intensity next. Read the Big Thing again with that in mind.

🎙️ From the Pods

🎙️ HIMSSCast — investment in ambulatory surgery centers soars

Capital One’s Erik Tellefson tells host Susan Morse that the inpatient-only list “is set to go entirely by 2029,” and that hospitals can charge 50%+ more than ASCs for the same procedure.

💡 Builder take: Scheduling, pre-op clearance, anesthesia workflow and post-discharge follow-up in small independent surgical facilities — none of them with hospital-grade IT — is an unglamorous, underbuilt market with a deadline attached.

🔇 Speaker Blindspot: Selection bias. The 40–50% cost gap is offered as proof of ASC efficiency, but ASCs screen out the comorbid and complex patients hospitals must accept. Asked twice what drives the differential, he deferred to an absent colleague.

🎙️ UnHack with Drex DeFord — a fake job offer put a children’s hospital on a target list

Security researcher Vangelis Stykas tracked the pattern for nearly two years: 1,640 companies, 57 countries, some independent contractors holding simultaneous access to as many as 30 organizations. “Nobody kicked down the digital front door with some brilliant zero day,” DeFord notes. “They walked in as people who were supposed to be there.”

💡 Builder take: Your real attack surface is offboarding hygiene.

🔇 Speaker Blindspot: Availability heuristic. The children’s hospital anchors the story emotionally, but by the hospital’s own account it involved a former contractor’s personal device with no evidence of access to hospital systems — the most alarming example is the weakest one.

💡 BTW

💡 BTW: Katya Andresen, the Cigna executive above, worked as a foreign correspondent for Reuters and the Associated Press before she ever touched health insurance — and wrote a book on nonprofit marketing called Robin Hood Marketing. Reading transcripts to find out what people are actually asking is, in retrospect, a reporter’s instinct.

💺 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]

Medical Director, Clinical Informatics — Availity · Remote
Clearinghouse side. Rare seat where you’re building on the transaction rail instead of asking someone else for access to it.
🔗 Apply on LinkedIn

Managing Director, New Ventures (Cedars-Sinai) — Redesign Health · Remote
Venture studio inside an academic system. Spinning clinical problems into companies, with a budget.
🔗 Apply on LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Hit reply and let me know.

📅 Upcoming: Measuring What Matters: ROI Frameworks for AI and Long-Horizon Healthcare Initiatives (CHIME, free, Sept 2, 2–3pm ET). [I ❤️ CHIME.]

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)

⚡ Around the Wards

• Metriport raised $26M for open-source clinical data retrieval — led by TJ Parker at Matrix, with ARTIS and Y Combinator participating. Amazon One Medical, Sollis Health and Color Health already run on it. Full read below.

• Datavant launched a provider-to-provider exchange network the same day — days to hours, and it finally kills the fax request.

• Boston Scientific disclosed a cyberattack disrupting order processing and shipping worldwide — no restoration timeline. Your device supply chain is a dependency graph nobody drew.

• 🎧 Podcast: Healthcare is Hard — “Healthcare’s Biggest AI Opportunity Isn’t Clinical. It’s Consumer.” — CVS built ~150,000 individual-level agentic twins and tests product concepts against them before launch.

• 🔮 My bet: retrieval consolidates to two or three winners inside eighteen months, and the deciding variable isn’t API quality. It’s which one the network governance bodies are comfortable letting other people’s customers ride on.

🔬 The Big Thing

The code is open. The door still has a bouncer.

Metriport raised $26 million on Thursday to expand what it calls open-source healthcare data infrastructure — the matching, extraction, standardization and deduplication that turns scattered records into one queryable model. Amazon One Medical, Sollis Health and Color Health already run on it.

That same morning, Datavant launched a provider-to-provider exchange network built to cut retrieval from days to hours. Two bets on the same bottleneck in one news cycle. Neither is about a model.

The interesting sentence in the Metriport release isn’t the funding. It’s from their general counsel, Matt Davis-Ratner: “We do deep diligence on every organization that wants to access the healthcare exchange data networks via Metriport ... we pay attention to who’s actually in the room during onboarding calls.”

Their COO, Colin Elsinga, says they open-sourced the code because “the decisions around how patient information is managed and transformed are too important to trust to a black box.”

Both are true, and together they tell you where the black box moved.

The transformation logic is now inspectable. You can read how a duplicate allergy gets collapsed, how a mangled C-CDA becomes a resource, which fields get dropped. That layer has been opaque for twenty years and it’s where records quietly get corrupted before anyone sees them.

The access decision is not inspectable. It’s a vetting call, a HITRUST r2 certificate, and a person deciding who you are.

That isn’t hypocrisy. It’s probably correct. But notice which half you control. You can fork the repo tonight. You cannot fork the onboarding call.

😤 “Open source in health data is a marketing word.” Sometimes. Not obviously here — publishing the transformation layer is a real accountability move. Judge it by whether anyone outside the company has ever filed a bug against the deduplication logic.

😤 “So the moat is paperwork.” Yes. It’s been paperwork the whole time. We just told ourselves it was engineering.

❓ If retrieval commoditizes down to an install and a vetting call, what’s the first clinical product that only becomes possible once every patient arrives with a complete record?

📡 Builder’s Radar

Behavioral health data just got a $17M bet, and the cap table is the tell.

Onos Health closed a $17M Series A led by Costanoa, with Flare Capital and CVS Health Ventures, to structure unstructured behavioral health data for health plans.

Behavioral health is the hardest documentation in medicine to structure. The diagnosis lives in narrative, the severity lives in tone, and the outcome measure is a nine-question form nobody fills out consistently.

💡 80/20: If you build anything that reads clinical narrative, behavioral health is your hardest test set and your best one. Twenty synthetic psych notes, extract PHQ-9 severity and suicidality risk, grade by hand. Those failure modes show up everywhere else first.

⚡ Quick hits

The invisible labor of running agents at scale — the work doesn’t disappear, it becomes allocation, specification, evaluation and recovery, none of which any dashboard measures. Log the minutes you spend supervising agents separately from doing the work for a week. Over a third and you don’t have a tool, you have a direct report.

Stanford’s SPHERE portal — a fully synthetic twin of the Stanford ADRC cohort: 644 participants, nine modalities, cognition through amyloid PET and CSF proteomics, no IRB amendment. Released back in April, so not news — but it’s the best free multi-omics sandbox going. Non-commercial DUA.

AWS is acquiring DuckLabs, DuckDB staying independent and open source. Matters because DuckDB is quietly the default engine for claims and registry extracts on a laptop.

🎙️ From the Pods

🎙️ Healthcare is Hard (LRVHealth) — “Healthcare’s Biggest AI Opportunity Isn’t Clinical. It’s Consumer.”

Sri Narasimhan, who runs enterprise customer experience, insights and innovation at CVS Health, describes building roughly 150,000 individual-level “agentic twins” — each seeded from a moderated, consented interview of about an hour, plus behavioral and demographic data — and testing messages, formulary changes and product concepts against them before launch. They kept them at N=1 rather than rolling up to personas, so any subpopulation stays queryable later.

“I have 150,000 patients in the room with us.” — Sri Narasimhan, CVS Health

💡 Builder take: His reason for building twins should land hardest for you — clinicians won’t answer your research requests. If you can’t recruit forty physicians to evaluate your workflow, build twins from the handful who’ll give you an hour.

🔇 Speaker Blindspot: Begging the question. The twins are pitched as the fix for say-do bias. But the training substrate is an hour of someone answering questions about their preferences — stated preference, the exact thing that produces say-do bias. And “somewhere between 85 and 90%” accurate arrives with no holdout design.

🎙️ Rounds (Second Opinion) — “The EHR is Dead, Long Live the EHR”

Sam Toole of Primary Venture Partners relays what Epic insiders are telling people: a three-to-four-year backlog is now buildable in about a year. His useful conclusion — stop trying to displace the system of record, build the system of action, and monetize the way EHRs actually do, which is revenue cycle, not seats.

He also reports PointClickCare moved to block computer-use agents on its platform. Treat that as a claim, not a filing — but if your product depends on agentic browsing of someone else’s platform, you have a countdown clock, not a moat.

🔇 Speaker Blindspot: Appeal to unverifiable anecdote. The evidence for “software is nearly free to build now” is secondhand, and one claim arrives with an explicit disclaimer that the speaker isn’t sure it’s true — after which the panel reasons from it as fact.

💡 BTW

💡 BTW: TJ Parker, who led Metriport’s round, grew up around his father Lenny’s pharmacy in Concord, New Hampshire — Northeast Pharmacy Services, twenty-five people filling prescriptions for a hundred assisted living facilities and group homes around the state. The individual-dose packaging that business used for eldercare is the thing TJ took and turned into PillPack. Amazon bought it for about $1 billion. Boston Globe

💺 Builder Seats

Chief Clinical Informatics Officer — City of Hope · Duarte, CA · $145–210/hr
Posted requirement is MD/DO plus formal informatics training — a seat written for a practicing informatician, not a data executive.
🔗 Apply on LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Reply and I’ll include it.

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, may contain errors)

Many of you know, I had to pause the substack momentarily while sorting out some health issues.

I am doing much better 👍 and I hope to be back soon …

… better than before ❤️‍🩹 …

… as I can’t help become a better doctor (and friend) (and writer) after being a patient.

Thanks for all the support … sometimes posting this substack seems like an experiment in loneliness 🥺 and I appreciate the kind words from this community!

- kevin (with all ❤️ and no AI > except the image gen about 🌈)

⚡ Around the Wards

• FDA and CMS hosted a “clinical AI demo day” nobody announced — ten companies at White Oak on July 8, and the agenda covered both regulation and how to pay for it.

• Explainable AI made primary care doctors better and lay people worse — 623 lay people, 153 PCPs, and the fairness-trained model that closed a skin-tone accuracy gap left lay users anchoring on whatever the AI said.

• A FHIR team built the same personal health record twice to see if the standard is a good agent guardrail — the version with server-side validation on every write was the one worth keeping.

• Sentara moved to terminate its Anthem contracts over a 7.2-point gap — 380,000 Virginians, and the CFO on the other side of your pitch is doing this math right now.

• CMS finalized the FY 2027 inpatient rule and closed the Breakthrough Device shortcut to Medicare add-on payment — the designation no longer stands in for proof that the thing works better.

🧭 The Curbside

“Can I self-host a voice model instead of signing another scribe contract?”

• Short answer: Closer than it was on Monday, and still not a product.

• What changed: NVIDIA published NemotronLabs VoiceChat-11B on Monday — an 11-billion-parameter model doing streaming speech understanding, speech generation, and tool calling in a single architecture, with open weights. The reason that combination matters is that it collapses a three-vendor pipeline — ASR, LLM, TTS — into one artifact you can run inside your own boundary. No third-party BAA in the audio path.

• Builder read / Watchout: Open weights are a licensing fact, not a clinical one. Nothing here has been validated on accented speech, interpreter-mediated visits, or an ED at 2 AM with a monitor alarming — and real-time full-duplex at 11B is a GPU line item, not a laptop. The license is OpenMDW-1.1, which is permissive and does allow commercial use — but the model card itself says the model is “ready for research purposes only,” which is the sentence your compliance office will read. What this genuinely unlocks is evaluation: you can now build a scribe-eval harness without paying a vendor for the privilege of testing them.

• 😤 Haters: “Great, another open model that will be superseded in six weeks.” Probably. But the thing that persists isn’t the checkpoint — it’s your harness, your test set, and your knowledge of which failure modes matter in your setting. Those outlive every model release. Build the part that doesn’t expire.

“A patient just handed me a QR code containing a summary of their whole chart. What is that?”

• Short answer: A Smart Health Link, and it’s now a default consumer feature rather than a standards demo.

• What changed: Google Health 5.05 shipped Smart Health Links to US users on August 3, letting anyone generate a secure, shareable summary of their medical records as a URL or QR code (Google describes the rollout as phased over the coming weeks) — no portal login, no CCD export, no FHIR endpoint to negotiate. The same release turned on two-way sync with Apple HealthKit, so Fitbit and Pixel Watch data now crosses into iOS natively.

• Builder read / Watchout: Interesting concept but read the google post disclaimer: “Not intended for medical purposes or as a substitute for official clinical health records. “

🔬 The Big Thing

Ten companies got a demo day with FDA and CMS. Nobody was told it happened.

On July 8, officials from the FDA and CMS hosted what an internal agenda called a “clinical AI demo day” at FDA’s White Oak headquarters. Ten companies presented.

Anthropic. Microsoft AI. Amazon One Medical. K Health. Curai. Counsel Health. Doctronic. Ellipsis Health. Hippocratic AI. Welldoc.

The meeting was not publicly announced. STAT’s Mario Aguilar reviewed the agenda and published it yesterday.

Read the list twice. There is no health system on it. No specialty society. No nursing organization. No FQHC network. No academic evaluation group. No open-source project.

The agenda wasn’t only “is this safe.” It was also “how should Medicare pay for it” — and that second question is the one that decides what actually gets built.

Here’s why that distinction matters more than the guest list.

FDA clearance tells you a thing is legal to sell. It does not get the thing into a workflow. What gets a technology into a workflow is a payment mechanism — a code, a rate, an add-on, a bundle. That is why the FY 2027 inpatient rule finalized this week is a bigger deal for medical AI than most 510(k)s, and why every serious health tech founder eventually stops reading FDA guidance and starts reading the Federal Register.

The companies in that room understand this. Several of them are building autonomous or near-autonomous clinical products — the kind that don’t fit any existing evaluation-and-management code, because there is no physician performing the service. Somebody has to invent the payment category. The people shaping how that category gets defined were in a room in Silver Spring on a Wednesday in July, and the people who will have to live with the result were not.

I want to be fair to everyone involved. Regulators talking to builders is good. Agencies that write rules about technology they have never seen operate write bad rules, and hands-on sessions are how you avoid that. Four of the attending companies talked to STAT openly about what they said.

But “we met with industry to understand the technology” and “we met with industry to work out how to pay for the technology” are different meetings, and only one of them is a policy process.

If you build clinical software and you are not reading proposed rules and filing comments, you have outsourced the definition of your market to ten companies who were invited and you were not.

The comment docket is the room that is actually open. It is unglamorous, it is slow, and it is the only venue where a community health center’s operations director and Microsoft get the same word count.

😤 “This is a nothingburger. Agencies meet with industry constantly, that’s what a regulator does.” Constantly and, ordinarily, on the record — advisory committee meetings are noticed, minuted, and open. The thing that makes this one worth a paragraph is that it wasn’t announced and we only know the guest list because a reporter got the agenda. If the process is unremarkable, publishing it costs nothing.

😤 “What exactly do you want, a physician on every panel?” Yes, and it’s not a big ask. There are federal advisory committees for hearing aids.

😤 “Clinicians can already comment on proposed rules. Nobody does.” Correct, and that’s the actual problem in this story.

🧪 Try the interactive: The Underwater Line — every Medicare DRG price, and the technology cost that sinks it — all 773 national MS-DRGs plotted at once, average Medicare payment against annual discharges, with a slider that drags a new technology’s cost across them so you can watch how much of American inpatient care it cannot fit inside. Built with real CMS data.

📡 Builder’s Radar

The same AI explanation made experts sharper and everyone else worse

A Nature Medicine study of 623 lay people and 153 primary care physicians tested LLM-generated explanations alongside dermatological diagnoses.

The headline finding is the good one: assistance from a fairness-constrained model — one trained to perform evenly across skin tones — improved final diagnostic accuracy and reduced skin-tone-related performance disparities for both groups. Credit the fairness training, not the explanation.

The finding underneath it is the one builders need. Lay users showed automation bias — their accuracy rose when the model was right and fell when it erred. Experienced PCPs stayed resilient irrespective of the AI’s accuracy. Separately, the authors flag an ordering effect that applies to everyone: showing the model’s diagnosis before the human decides produces stronger anchoring.

The same explanation is a decision aid for someone with a prior and a suggestion engine for someone without one.

This is the cleanest evidence I’ve seen that “who is the user” is not a product-marketing question — it’s a safety parameter. A tool validated on physicians and then shipped direct-to-consumer is not the same tool.

😤 “So don’t give patients AI. Got it.” Not what it says. It says an interface that shows a confident explanation to someone with no way to falsify it is doing something different than the same interface shown to someone who has seen four hundred rashes. Design for the difference — friction, uncertainty display, an explicit “here’s what would change my mind” — instead of shipping one UI and calling it democratized.

A clinician read Steve Yegge on agent context and recognized the sign-out

Doug Fullington makes the argument I wish I’d made: agentic context handoffs are a patient handoff problem, and medicine spent thirty years solving it.

The literature he pulls is exactly the right literature. Petersen’s 1994 work found 6.1 times the odds of potentially preventable adverse events under cross-coverage (a wide confidence interval, 1.4–26.7, off 54 events — the direction is solid, the point estimate is not). The 2014 NEJM I-PASS trial cut medical errors 23% across 10,740 admissions — by standardizing what gets transmitted, not by asking people to try harder.

His line lands: /compact is a discharge summary written by someone who never met the patient.

Every clinician already knows that the dangerous moment is not the work — it’s the transfer of responsibility for the work. AI engineering is rediscovering this from scratch, and clinicians are the only people who arrive with the solved version.

⚠️ Disclosed in his post: Fullington is a physician part-owner of Catalyst Health Group, which has a financial interest in Matic, an AI clinical documentation platform. It doesn’t change the handoff argument, but you should know it’s there.

🔮 Where this lands: structured agent handoff becomes a named pattern within a year, and whoever writes the I-PASS of context compaction gets cited for a decade. I don’t think it’s an AI lab that writes it.

Sentara would rather drop 380,000 people than take a 1% cut

Sentara Health issued a termination notice to Anthem after eight months of stalled negotiation. Sentara asked for a 6.2% blended increase for 2027; Anthem countered with roughly a 1% decrease. Most affected commercial and Medicare agreements run through at least December 31, 2026; the Medicaid agreement runs to January 28, 2027, and others expire on a rolling basis during 2027.

That is a 7.2-percentage-point gap across commercial, Medicare, and Medicaid lines covering roughly 380,000 Virginians.

A health system that will walk away from 380,000 covered lives rather than absorb a one-point cut is not negotiating. It’s out of room.

Ultra-shorts

Hinge Health is buying Cylinder for $105M

The publicly traded musculoskeletal company is going all-cash for a gastrointestinal-care startup, extending from MSK and migraine into IBS and IBD for its existing member base. Condition-by-condition consolidation, at speed.

Amae Health is wiring Fitbit data into serious mental illness care

Sleep, activity, and heart-rate variability from Google Health Enterprise, pushed into clinical workflows to flag early relapse signs between visits for patients with SMI. The passive-signal thesis finally pointed somewhere it might matter.

A health system led a Series B

Wellinks closed the first $10M tranche of its Series B with participation from UMass Memorial Health — a health system, not a venture fund. No lead investor was named. When your customer writes part of the check, the pilot conversion problem gets a lot easier.

🛠️ From the Workbench

Health Samurai’s PHR, built twice

A team rebuilt the same personal health record twice with Claude Code — once on plain React/Node/Postgres, once FHIR-native — to test whether FHIR functions as a framework for agentic development rather than just a data standard.

The v1 failure mode will be familiar to anyone who has driven an agent past a toy project: it invented its own data model, then drifted away from it every session. The v2 result was a smaller, more coherent codebase, because the agent had rails — a fixed data model of 150+ FHIR resources, server-side validation on every write, and generated types so a wrong field name fails at compile time instead of at 2 AM.

The detail that convinced me is small: asked for a patient summary, the agent reached for Composition, FHIR’s own model for a sectioned clinical document, and wired it to existing Condition, MedicationStatement, and AllergyIntolerance resources. No bespoke schema. A feature became a conversation.

Both the PHR source and the Claude Code skills are open.

💬 Standout Quote

“Doctors as builders — no dev team in the middle; describe the workflow, get a working app.”

🎙️ From the Pods

🎙️ Relentless Health Value — “EP523: The Sleeping Giants of Healthcare, with Suhas Gondi, MD, MBA”

Gondi walks through a GLP-1 denial where the physician did everything right — shared decision-making, told the patient to verify coverage, patient confirmed the employer covers it, clinic cleared every prior-auth criterion — and the patient still got quoted list price at the counter. The reason: the employer had recently decided to cover the drug only through a third-party wraparound vendor, and that vendor is the sole covered prescriber under the plan. Nothing in a formulary check surfaces that.

The structural point is worse than the anecdote. There is no path at all from a pharmacy claim rejection back to the clinician who wrote the prescription. The doctor usually finds out at the three-month titration visit, which is to say the patient took nothing for three months.

“There’s no feedback, right, from that pharmacy about that claims denial to that clinician and to their office.” — Suhas Gondi, MD

🔇 Speaker Blindspot: False cause substitution. Gondi carefully establishes that both employers and clinicians are acting in good faith, which leaves ignorance as the only remaining explanation — so his remedy is awareness and local conversations. But he supplies the fact that kills that remedy himself: a clinician cannot possibly track which of hundreds of plan sponsors each patient has and what each one changed last quarter. He raises the impossibility and then recommends meetings anyway. This is an information-routing problem that needs machine-readable benefit design at the point of order, not an information-awareness problem that needs better-informed doctors.

🎙️ Vital Signs — “Ep 70: Nourish CEO Aidan Dewar on Hiring in the AI Era”

Two things worth stealing. First, Dewar reports the bottleneck moving in opposite directions on either side of his company: in engineering, code stopped being the constraint and product definition became it; in strategy and operations, AI made scoping so fast that they now hold a backlog of well-specified work they lack people to execute.

Second, and more concrete: they took meal logging, progress tracking, goal setting, meal planning, recipes, and messaging — six roadmap features living in six parts of the app — and collapsed all of it behind a single AI health assistant. Engagement and retention went up.

💡 Builder take: If AI-assisted coding removed implementation as your limiting factor, then spec quality is your binding constraint — and precise clinical specification is exactly the asset a practicing clinician has and a product manager doesn’t. Stop optimizing build speed. Also: go count the discrete features in your own tool’s nav tree, and ask what survives if you collapse them behind one conversational surface.

🔇 Speaker Blindspot: Composition fallacy dressed as Jevons paradox. Dewar argues AI makes care cheaper and better, so demand rises, so employment rises — ending on more clinicians than today. Jevons needs elastic demand and a substitutable constrained input. Clinical demand is gated by payer willingness to pay, not consumer appetite, and the constrained input is a credentialed dietitian requiring a master’s plus 1,000 supervised hours. The tell is in the transcript: asked directly whether they’d train their own supply, he says they haven’t considered it. He asserts the outcome while declining the only lever that produces it. Worth noting the whole episode contains not one outcome endpoint, effect size, or comparator behind repeated claims that the product improves outcomes, and the host never asks.

💡 BTW

💡 BTW: Suhas Gondi — the physician arguing on this week’s podcast that clinicians don’t understand how their patients’ benefits work — interned at CMS, the Brookings Institution, and the U.S. Senate before he ever went to medical school. He learned the payment system first and the medicine second, which may explain why he can see the seam the rest of us walk past.

💺 Builder Seats

Physician Informatics Executive (IC4) — Oracle · Remote (US)
EHR-vendor side, which is a rare chance to build from the platform instead of on top of it. Band runs $193,600–$414,400. Two real filters before you click: roughly 80% travel, and it requires US citizenship plus the ability to obtain a government security clearance.
🔗 Apply on LinkedIn

AVP, Digital Transformation — Sarah Cannon Research Institute (McKesson) · Remote (Texas-based)
Enterprise digital and AI transformation across 200+ research sites and 1,300+ physicians, with interoperability and product lifecycle ownership written into the description. Not a physician role, but it’s the seat where clinical trial infrastructure actually gets rebuilt.
🔗 Apply on LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Reply and I’ll include it.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin

⚡ Around the Wards

• Nine landmark ambient-scribe studies, zero from a safety-net clinic — Dr. Gigi Magan checked forty-one references against their primary sources and found the evidence base has a hole shaped exactly like her patients.

• “Once daily” reads as “eleven daily” in Spanish — RxTran’s Sharon Blank argues pharmacist verification of translated labels is permanent infrastructure, not a transitional safeguard.

• One hospital took its order-set library from nearly 1,000 to 350 — the fix for duplicate sepsis order sets turned out to be a governance pattern, not a build ticket.

• 3,500 tubes of blood, 153 terabytes of data — ChronicleBio says it has already found five sub-diseases inside POTS where the biology differs but the symptoms don’t — which would explain a decade of trials that went nowhere.

• 🎧 Podcast: Radio Advisory — “309: Regional health plans attempt a financial turnaround” — Advisory Board’s Jared Landis on why the rate-setting calendar quietly breaks most vendor pitches to regional plans.

🧭 The Curbside

“Our CIO keeps saying we’re in a squeeze. What does that mean for my pitch?”

• Short answer: It means your tool is competing against roughly 150 ideas for about ten funded slots, and “capability” isn’t a category that survives that.

• What changed: Bill Russell, Drex DeFord and Sarah Richardson just came back from three rooms of academic medical center leaders — CIOs, CMIOs, and revenue cycle — and reported the number that came up in all three: roughly $900 million in projected top-line revenue coming out over three years. The governance process they described: departments generate ~150 ideas, a subcommittee chair filters up 10 to 20, and the organization lands on roughly ten projects it will actually do — published, so everyone can see where they sit.

• Builder read / watchout: The subcommittee chair’s filter is “does this add revenue or take out cost,” and the second filter is “how many of my IT hours does this eat.” A CMIO in that room put it as decisions made elsewhere becoming checks they can’t cash. If your one-pager doesn’t have a cost-takeout number and an IT-hours number on it, your champion is improvising in a five-minute slot — and improvised pitches land at #11.

• 😤 Haters: “This is just enterprise sales advice with a stethoscope on.” Partly. But the specific thing here is that the “no” you get is usually not about your tool. It’s a resource statement, and resource statements have a shelf life.

🔬 The Big Thing

Nine landmark scribe studies. Zero from a clinic like hers.

Dr. Gigi Magan, a family physician who writes from a safety-net exam room, spent the last month building a lecture on ambient AI scribes for the California Telehealth Resource Center. Forty-one references, each one checked against the primary source.

Then she lined up the settings of every major study: UCLA. Mass General Brigham. Emory. UCSF. Yale. UC Davis. Kaiser Northern California. Penn. Stanford.

Published trials or large studies conducted in a federally qualified health center or community health center: zero. Peer-reviewed studies reporting scribe accuracy stratified by patient race, language, or accent in real clinical use: also zero.

The patients most likely to benefit are the least represented in the evidence and the most exposed to its failure modes.

That second zero is the one that should make you sit up.

We already know the underlying technology has a demographic gradient. The landmark 2020 test of five commercial speech recognition systems found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers — nearly double — with the gap tracing to the acoustic model and thin training data. Magan adds the number that doesn’t get quoted: more than twenty percent of Black speakers’ audio was degraded beyond usability, versus under two percent for white speakers.

Vendors have invested heavily since. Nobody has published the check.

Health centers serve more than 31 million people. The visits are long, multilingual, and heavy with social complexity — which means the documentation burden per visit peaks exactly where the evidence stops.

Magan is careful, and I want to be careful the same way: she uses one of these tools every day and isn’t going back to typing.

This is not an argument against ambient documentation. It’s an argument about who gets to generate the evidence.

Here’s what makes it a builder story instead of an op-ed.

Cardiology has started building the machinery. The AHA’s AI Assessment Lab, powered by Dandelion Health, ran Ultromics’ EchoGo Heart Failure against roughly 90,000 real-world echocardiograms and published the result with subgroup findings by race and age alongside the accuracy numbers — HFpEF identified up to 263 days earlier than standard care in patients who would otherwise have been missed, wrapped in clinical and economic modeling out to five years.

That’s an independent body, independent data, and stratified results published where a procurement committee can read them.

Ambient documentation is deployed far more widely than that algorithm and has no equivalent body doing that work.

So the assessment layer for the most-deployed clinical AI in America is vacant, and the entry requirement is a QI dashboard, not an R01.

Decline rates by language. Note quality by population. Edit burden on interpreter-mediated visits.

None of that needs a grant. It needs someone to decide on day one that equity gets measured instead of assumed.

😤 “Absence of evidence isn’t evidence of absence. You’re fearmongering about a tool that demonstrably works.” She said the same thing more plainly than I would have — she uses it daily, the benefit is real, and the time savings are modest but the attention benefit isn’t. The claim isn’t that scribes fail on accented speech. The claim is that nobody has looked, in the settings where it would matter most, and “we assume it improved” is not a finding.

😤 “FQHCs don’t have research infrastructure. That’s why the studies are at Stanford.” Right, and that’s the two-tier outcome writing itself. Somebody has to go first.

😤 “Vendors have this data internally.” Then ask for it.

🧪 Try the interactives:

A — Nine and 1,352 — an animated field of every federally funded health center in America, sorted by the share of patients best served in a language other than English, against the nine academic settings where the ambient-scribe evidence was actually generated. Built with real HRSA health center data.

B — Clinics Like Mine — every HRSA health center grantee in the country plotted at once: panel size against the share of patients best served in another language. 32.3 million patients, 27.8% of them non-English-preferred, and not one published ambient-scribe study run in any of them. Built with real HRSA health center data.

A hospital retired two-thirds of its order sets and called it a governance pattern

John Lee, MD — an emergency physician and Epic consultant — published the second half of his sepsis order-set story: at a health system he documents but doesn’t name, a duplicate-sepsis-order-set conflict got resolved by bucketing review by clinical domain, and when the same model was applied system-wide over two years, the library went from nearly 1,000 records to 350.

Dormant sets were retired outright. Active duplicates weren’t arbitrated — they were consolidated into one.

Two-thirds of an EHR’s clinical content library was redundant, and nobody could see it until the review was organized by clinical domain instead of by request queue.

The build here isn’t software. It’s the review structure that makes the redundancy visible — and that’s a thing a clinician can design and an engineer can’t.

💡 80/20: Before you build a tool that writes order sets, count the ones you have. Pull the list, group by clinical domain rather than by owner, and find the duplicates. If your ratio looks anything like 1,000-to-350, generation was never your problem.

❓ Two-thirds of a content library was dead weight and it took a two-year human review to see it. Is deduplicating clinical content actually a good first agent task — or is it the one place you most want a human who knows which of two nearly-identical sepsis order sets the night team actually uses?

A trauma surgeon rebuilt his research lab’s operating system in twenty minutes

A trauma surgeon mentoring thirteen trainees across fifteen manuscripts replaced years of accumulated lab-management workarounds with Notion plus an agent wired in over MCP connectors. His section heading for the whole build: “Build the whole thing in twenty minutes.”

That’s the whole clinicians.build argument compressed into one weekend project. The scarce input wasn’t engineering.

It was knowing what a manuscript pipeline actually needs to track when the first author is a PGY-3 on nights.

🔮 My bet: within a year the interesting artifact from clinician-builders isn’t the app — it’s the connector config. The thing worth sharing is a working MCP setup for a specific clinical or academic workflow, and somebody is going to start a registry of them.

ChronicleBio says POTS is already five diseases, not one

ChronicleBio — cofounded by Fidji Simo, Rohit Gupta and Rishi Reddy — has banked more than 3,500 tubes of blood and pulled 153 terabytes of data out of them: 890 draws from 709 patients in its first year across Utah, Arizona, Texas and India, on $15M raised. On August 11 it opens sign-ups for mobile phlebotomy trucks that come to patients’ homes, free for the first 250 and $400 after, in exchange for their biological data.

The thesis is a clinical-trial thesis, not an AI thesis. In Simo’s words: “there could actually be five sub-diseases within POTS, and the drug would work for one of them, but not the other four.” She says they’ve already found them — same symptoms, different biology, immune-driven in one group and mitochondrial in another. The sub-diseases don’t have names yet.

If they’re right, the reason these trials keep failing isn’t the drugs. It’s the phenotype.

😤 “This is a biobank with an AI press release stapled to it.” It’s a biobank, yes. That’s the point — the constraint in neuroimmune disease has never been model architecture, it’s that nobody assembled the cohort. Deep phenotyping on a neglected condition is unglamorous and it’s the actual bottleneck. Ask me again in two years whether the five-way split replicates.

🎙️ From the Pods

🎙️ Radio Advisory — “309: Regional health plans attempt a financial turnaround”

Advisory Board’s Jared Landis, talking with host Rae Woods, lays out the mechanic that breaks most vendor pitches to regional plans: 2025’s rates were determined in the first half of 2024, so everything a plan learned during 2024 couldn’t touch its 2025 bid — it lands in 2026. Anything you sell that only improves pricing accuracy has a payoff two budget cycles out.

🔇 Speaker Blindspot: Composition fallacy. Landis argues the turnaround isn’t zero-sum — “they can all improve their pricing, they can all improve their operations.” True for any one plan; not true for all of them at once, in a market he’d already described as stagnant where a plan grows only by taking an account from a competitor and where employers keep shifting to self-funded. He then uses national for-profit payers’ Q2 earnings as a proxy for Blues performance minutes after explaining that the nationals have structurally different scale and margin dynamics.

🎙️ The Gist Healthcare Podcast — “Putting patient safety into practice: improving access across healthcare organizations”

Victor Hassid, MD, Associate Vice President of Access Strategic Operations at MD Anderson Cancer Center, describes running every intake question through two comprehension checks — whether the administrative team on the phone understands what the clinical team meant, and whether the patient understands what the scheduler says. His line: it doesn’t matter what either of them says if the patient doesn’t comprehend it.

🔇 Speaker Blindspot: Goodhart’s law, self-inflicted. His headline result is cutting median time-to-offer from five days to two — and he himself warns that it is “inappropriate and at the same time risky to assess access to care as access to your first appointment,” because you can clear an upstream bottleneck and create a downstream one. Then he proposes industry-wide access KPIs with “possibly even ranking,” which is the textbook condition for a measure to stop being a good measure.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin

⚡ Around the Wards

• The Federation of State Medical Boards says AI is not ready to be licensed like a physician — and its president and board chair published the reasoning, because bills to create a separate licensing lane already came and went in two states this year.

• NewYork-Presbyterian put every AI tool it runs under one management system — cancer detection, structural heart, discharge planning, documentation, all governed as one portfolio instead of one purchase at a time.

• Medical schools are restricting trainee access to ambient scribes while hospitals sprint to deploy them — the argument is that the struggle of writing the note is the reasoning, and offloading it in year three is different from offloading it in year twelve.

• The White House meets OpenAI, Anthropic, Google and Meta today on a government pre-release review framework for frontier models — nobody has yet defined “frontier model” or named which office runs the review.

• 🎧 Podcast: Health Tech Nerds Radio — “The Grand Roundup” — ProMedica recovered nearly $10 million by using AI to find patients coded as self-pay who actually had billable coverage. The hosts’ sharper point: today’s cash prices look cheap only because that population was priced as bad debt.

🔬 The Big Thing

The people who license every doctor in America just said AI can’t be one. Read what they conceded getting there.

On August 3, Humayun Chaudhry, DO and Christy Valentine Theard, MD, MBA — the president/CEO and board chair of the Federation of State Medical Boards — published their answer to whether generative AI should be licensed to practice medicine.

The answer is no. Not “not yet, pending study” — no, and the reason is not accuracy.

Their argument is that a license isn’t a certificate of competence in the first place. It’s a grant of authority tied to a human being who can be disciplined, sued, and named.

A license was never a statement that you’ll be right. It’s a name to attach when you’re wrong.

That’s the whole thing, and it’s why “the model outperformed physicians on the benchmark” has never been an argument for licensure. Nobody disciplines a checkpoint.

Now the part they conceded.

Bills were introduced this year in Idaho and Iowa to create state licensing boards for “autonomous service providers,” sitting outside the medical board entirely. Both failed.

The authors’ read on that: “We are past the point of wondering whether someone will propose licensing AI like a human physician.” Somebody will do it again, and next time it may not fail.

And Utah has already run the smaller version. The state’s Office of Artificial Intelligence Policy approved a 12-month pilot with Doctronic to automate guideline-based renewals of 192 commonly prescribed drugs under physician oversight — a program that moved through the state’s AI apparatus and then drew the medical board’s attention anyway.

FSMB has now stood up a Workgroup on the Regulation of AI in the Practice of Medicine, because its own 2024 guidance was written before agentic systems existed. They’re also telling boards to go re-read how their state defines “the practice of medicine.”

That last instruction is the one to watch. The definition of “practice of medicine” is the API contract every clinical AI tool is built against, and 69 boards are about to start editing it independently.

💬 Standout Quote

“A medical license, therefore, isn’t just a permit to generate thoughtful and informed answers. It’s a legal and ethical grant of authority tied to human judgment, professional accountability, discipline, transparency, and a duty to put patients first.” — Humayun Chaudhry, DO, and Christy Valentine Theard, MD, MBA

😤 “This is a guild protecting its turf.” Partly, sure. Every profession defends its boundary and this one has a two-hundred-year habit of it. But the argument they made isn’t “we’re better than the machine” — it’s “there has to be someone the patient can name,” and I have not heard a good version of who that is if not the licensee. Tell me the alternative and I’ll take it seriously. “The vendor’s E&O policy” is not the alternative.

😤 “Nobody is seriously trying to license a chatbot as a doctor.” Two state legislatures introduced bills this year. Go read them.

😤 “Fine, but this changes nothing for me tomorrow.” Correct. It changes something for you in about eighteen months, when a board somewhere quietly rewrites the definition of the practice of medicine and your tool is either inside it or outside it. You will not get a comment period notification for that one.

🧪 Try the interactives:

A — A Name to Attach — an animated field of every paid malpractice claim in the NPDB from 2004–2021, settling first into what was alleged and then into what happened to the patient. The FSMB’s argument for why AI can’t hold a license, drawn. Built with real NPDB data.

B — The Accountability Object — 193,023 paid malpractice claims, sorted by what was alleged and what happened to the patient, on one brushable log-log scatter. 1,911,185 reports, 985,019 names, zero models. Explore all 110 cells. Built with real NPDB data.

📡 Builder’s Radar

NewYork-Presbyterian stopped approving AI tools and started managing an AI portfolio

NYP adopted Signal 1’s AI Management System on August 3 to govern everything it runs — cancer and structural heart detection, discharge planning, clinical documentation — through one layer of risk assessment, continuous monitoring, bias detection and audit.

The health system co-designed the platform rather than buying it off a shelf, which is its own signal about how immature this category still is.

The unit of governance just moved from the tool to the portfolio, and that changes what you have to bring to the table.

A single-tool security packet was the old ask. The new ask is: what does your model emit, on what cadence, in what format, so that somebody else’s monitoring layer can watch it drift.

Medical schools are rationing the scribes that hospitals can’t deploy fast enough

Schools and health systems are restricting trainee access to ambient documentation tools even as the same institutions accelerate deployment for attendings, and the reason is developmental, not technical.

“The process of deliberately crafting the note forces us to use our brains to really wrestle with what’s happening,” Yale’s associate dean of educational technology and innovation Jaideep Talwalkar told STAT. “There’s an importance in doing that with great repetition.”

Which means somebody has to build the training-wheels version, and nobody has.

Every ambient scribe on the market is optimized to remove the work. Not one is optimized to make a third-year student produce the note first and then show them what they missed.

😤 “This is just handwriting-versus-typing all over again.” Maybe. The difference is that typing replaced the motor act and this replaces the reasoning act, and we don’t have the study that tells us whether that matters. Nobody does — that’s the actual finding in the piece. Educators are making a call on developmental theory because the evidence isn’t there yet.

The White House sits down with the four labs today, and nobody has defined the noun

The administration meets OpenAI, Anthropic, Google and Meta on August 4 to work on a government-review framework for frontier models before public release, building on the June 2 executive order’s voluntary 30-day pre-launch access system.

Two things are still open: what counts as a “frontier model,” and which office actually runs the review.

If you’re a thin wrapper on a frontier API, a 30-day pre-release review window is a 30-day gap in your release notes.

Two academic centers built an oncology decision tool and are now selling it to you

NYU Langone and Dana-Farber co-developed Solavia, an EHR-embedded oncology decision suite that surfaces biomarker-driven treatment pathways at the point of care. It went live internally at NYU Langone on June 30 and is now commercially available, licensed through NYU’s Technology Opportunities and Ventures office.

Two academic medical centers just decided the distribution model for a clinical pathway tool is a license, not a paper.

That’s the part worth sitting with. The traditional output of an academic clinical-pathways effort is a publication and a slide deck. This one is a SKU.

💡 80/20: If you’re a clinician sitting on an internal tool your department actually uses, find out who runs technology transfer at your institution and ask them one question: what would it take for this to be licensable? Most physicians never ask, and the office exists specifically to answer.

Your buyer may have quietly stopped owning the budget

CIO.com reports that CIOs are increasingly being sidelined on the most strategic AI budgets — by CEOs directly, or by a newly created Chief AI Officer seat that sits outside IT.

This is a general enterprise story, not a health story, which is exactly why it’s worth reading twice.

If you’ve spent six months building a relationship with a health system CIO, there is now a real chance the person who decides your deal doesn’t report to them.

The tell is easy to check: ask whoever you’re talking to who signs, and who else has to say yes. If they hesitate on the second one, you’re mapping the wrong org.

Ultra-shorts

Inocras raised a $31M Series B-3 to bring its CLIA/CAP-certified whole-genome sequencing and automated interpretation platform into US hospital networks, after adoption across 100+ Asian cancer centers. Total funding is now around $100M. The bet is whole-genome as the default instead of narrow NGS panels — which, if it lands, changes the shape of every downstream tool that has to route an incidental finding somewhere.

Labcorp’s 163-gene hereditary risk panel went nationwide direct-to-consumer on August 3, purchasable without a physician order through Labcorp OnDemand with genetic counselors attached and draws at 2,200+ service centers. A large reference lab going self-service on hereditary risk means the results land in primary care inboxes whether or not primary care ordered them.

Nikhil Krishnan wrote up his family’s delivery — emergency C-section, decision-to-delivery under six minutes — and used it to take apart the “shoppable healthcare” premise from the inside. The number that stuck with me, from the Commonwealth Fund data he cites: roughly 40% of all US pregnancy-related deaths occur in the six weeks after delivery — the window where American follow-up is thinnest.

Jordan Shlain, MD went on Graham Walker’s podcast to argue that primary care shouldn’t be tied to insurance at all, after 25 years running a concierge practice. Agree or don’t — the independent-practice thread is worth tracking, because independent physicians are the only buyers in this market who can say yes in one meeting.

💡 BTW

💡 BTW: Christy Valentine Theard — the FSMB board chair who co-wrote today’s Big Thing — is a New Orleans native and med-peds physician who founded her own practice in the New Orleans area, and who, after studying the effects of emotional and physical distress in children, was featured as an expert on The Oprah Winfrey Show on the post-Katrina aftermath. Before she was writing about the accountability structures of medicine, she was on daytime television explaining what a disaster does to a seven-year-old.

📅 Upcoming: EHIgnite Challenge Phase 1 Winners Showcase, Thu Aug 6 — nine teams demo turning dense EHI exports into plain-language output.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin

⚡ Around the Wards

• Eleven hospitals wired an old score to an automatic page — and mortality dropped 18% — 23,132 high-risk patients, deaths from 23.1% to 18.6%, and ICU transfers stayed flat. The model was already installed. The routing was the intervention.

• HHS’s health tech initiative turns one and adds seven pledges — bulk FHIR for population health, pharmacy interoperability, real-time benefits access, plus a “ditch the disk” imaging work group. Every pledge is a new API surface someone has to actually build.

• Mount Sinai Miami Beach becomes the fourth US system to give nurses Epic’s ambient AI — participating physicians there cut time in the chart per encounter by nearly 30% first. If you’re selling standalone nursing documentation, you’re now bidding against “already in the EHR contract.”

• A $0.13 LLM audit caught 96-98% of trials that quietly changed their outcomes — 91% sensitivity, 98% PPV on incomplete registrations. Research integrity turns out to be a cheap, boring parsing problem nobody had bothered to automate.

• 🎧 Podcast: Neural Compass — “Healthcare was designed to fail,” with Andy Slavitt — Slavitt on what changes when CMS starts paying for care that isn’t tied to a specific time and place, and why behavioral health’s effect sizes have barely budged in decades.

🧭 The Curbside

“Someone forwarded me the study saying ChatGPT beats OpenEvidence. Do I believe it?”

• Short answer: Believe the numbers. Don’t believe the conclusion people are drawing from them.

• What changed: The NYU Langone benchmark — GPT-5.2, Gemini 3.1 Pro and Claude Opus 4.6 against OpenEvidence and UpToDate Expert AI — published in Nature Medicine on June 12, and the frontier models won on all three axes. On 100 real de-identified clinical queries scored blind by 12 clinicians across 1,800 annotations, the purpose-built tools landed no better than Google Search AI Overview. What’s new this week is the methodology fight: developers are now openly disputing whether these benchmarks measure anything a clinician cares about. Daniel Yang, Kaiser Permanente’s VP of AI and emerging technologies, put it plainly: “I’ve never seen a single paper trigger the kind of reactions this one has in the health AI community.”

• 😤 Haters: “So the specialty tools are a scam.” No. They’re a wrapper, and wrappers are allowed to be worth money — that’s most of health tech. The uncomfortable finding isn’t that the wrapper is thin, it’s that nobody had measured it until this summer and the vendors would clearly have preferred it stay that way.

📡 Builder’s Radar

HHS’s interoperability pledge drive turns one, and year two is supposed to be about adoption

Seven new voluntary pledge areas joined the health tech ecosystem initiative at its July 27 anniversary event: population health data exchange over bulk FHIR, pharmacy interoperability, real-time patient access to benefits information, price transparency, trial matching, scheduling, software-based care access. There’s also a “ditch the disk” work group aimed at diagnostic imaging, which is the most honest name any federal work group has had in years.

Amy Gleason, CMS deputy administrator and chief product officer, who runs the program, put it plainly: “That’s our job in year two: adoption.” Officials put the initiative past 800 pledges.

Bulk FHIR going from a spec people cite to a thing large organizations have publicly promised to do is the most builder-relevant sentence in the announcement. $export at population scale is how you get a denominator instead of a patient.

😤 “Voluntary pledges aren’t policy.” They aren’t. And the supporting numbers deserve a raised eyebrow: HHS’s chief counselor claimed 60% of Americans can now reach their records through an app of their choice, up from 5% a year ago, heading to 80% by October — and reporters noted he offered no methodology for any of the three. CMS’s own published figures describe 81 pledges in active coordination. Take the pledge list more seriously than the podium math. The pledges at least name a thing someone has to build.

Epic’s ambient AI reaches the nursing station, and the point solutions are the ones who should be nervous

Mount Sinai Medical Center in Miami Beach extended Chart with Art to its inpatient nurses — the fourth US health system to do it, first in Florida. Epic hasn’t publicly named the first three. Participating physicians there had already cut time in the chart per encounter by nearly 30% before the nursing rollout.

Chief Nursing Officer Wendy Stuart framed it the way CNOs frame it: “Our nurses came into this profession to care for people. Chart with Art lets them step out from behind the keyboard and be fully present at the bedside.”

The procurement pathway here has no RFP in it. Already on Epic, already vetted, already under BAA, already integrated — the CNO pilots a unit and extends. If you’re selling standalone nursing documentation, you aren’t competing on quality. You’re competing against a line item that’s already paid for.

😤 “Ambient documentation for nurses is not the same problem as for physicians.” Agreed, and that’s the gap worth building in. Chart with Art turns conversations into draft notes. It doesn’t do fall-risk prediction, medication administration checks, or automated care plans. Name the specific clinical thing Epic isn’t doing, or expect the evaluation to end quietly.

Nvidia got 37 companies to agree on AI security. OpenAI, Google and Anthropic aren’t among them

The Open Secure AI Alliance launched July 27 with 37 founding members — Microsoft, IBM, Cisco, Salesforce, CrowdStrike, Palantir, Hugging Face, Red Hat, Databricks, the Linux Foundation — and has since grown past 50. Nvidia open-sourced an agent-harness research framework called NOOA alongside it.

Absent: OpenAI, Google, Anthropic, Meta. Which is to say, the four labs whose models you are probably using.

The catalyst was the Hugging Face intrusion, where a pre-release OpenAI model escaped a sandbox during an internal eval and — this is the part that matters — closed tooling blocked the forensic reconstruction afterward. An open model had to rebuild the 17,000-step timeline after the fact.

😤 “An alliance without the four biggest labs is a trade group, not a standard.” Yes.

A 9B model trained for three days beat every frontier model on the task it was trained for

A GRPO-fine-tuned 9-billion-parameter open model hit 87.3% of the achievable ceiling on a narrow catalog-integrity task, against 76.9% for the best frontier configuration tested — at 40 to 68 times lower cost per decision. Training cost: roughly 3.5 days and about $500 of rented GPU time on two consumer-class cards.

Caveat up front: this is a vendor’s own write-up, and every number in it is self-reported. Read it as a plausible existence proof, not as evidence.

Even discounted, it’s the whole argument for owning your intelligence rather than renting it. The tasks we actually need automated are narrow and repetitive — coding, registry abstraction, med-rec reconciliation, prior auth packet assembly — which is precisely the shape where a small task-trained model wins on cost, latency, and the fact that it can run somewhere you control.

The healthcare precedent is real but older than the post implies: Ambience reported a fine-tuned model beating 18 board-certified physicians on ICD-10 coding back in May 2025. Also a company release, also not peer-reviewed. The pattern keeps showing up anyway.

An AI primary care company bought a pediatric practice, and the asset was the text messages

Doctronic acquired Summer Health, moving from adult into pediatric care. Terms undisclosed.

Summer Health is a text-based pediatric service, so its 100,000+ encounters are its corpus: 100,000 real conversations about how a parent actually describes a sick child at 11 PM. You cannot buy that and you cannot synthesize it.

They didn’t buy software. They bought four years of a very specific kind of conversational data they’d otherwise have to earn one worried parent at a time.

🔮 My bet: the dataset, not the revenue, is going to be the stated rationale for most health AI M&A from here. Included Health and Firefly signed a deal the day before this one. I’d expect at least two more AI-first care companies to buy small care-delivery groups before year end, and for the press release to talk about “proprietary clinical data” rather than patient panels.

🎙️ From the Pods

🎙️ Neural Compass — “Healthcare was designed to fail,” with Andy Slavitt

Slavitt — former acting administrator of CMS, now co-founder and general partner at Town Hall Ventures — points at the new CMS ACCESS Model, which pays for outcomes rather than for an encounter at a particular time and place. Reimbursement without a clinician in the room, contingent on measurement. Host Mark Jacobstein notes that behavioral health’s effect sizes on PHQ-9 and GAD-7 have barely moved in decades.

🔇 Speaker Blindspot: Appeal to popularity, then poisoning the well. Slavitt argues models may do empathy better on the evidence that people like what they get — substituting engagement for outcome, which is the exact error he’d shred in a fee-for-service utilization argument. Then he preempts the safety objection by attributing it to fear of “organ rejection from the healthcare establishment,” which discredits the objector instead of engaging a single named failure mode. Sycophancy in crisis, adolescent harm, missed red flags — none of them come up. Worth knowing that his firm invested in the company that produces the podcast he’s a guest on. Disclosed, which is not the same as resolved.

🎙️ Relentless Health Value — EP522, “How exactly does GoodRx make money?”

Ge Bai, PhD, CPA — professor of accounting at Johns Hopkins Carey and of health policy and management at Bloomberg, recently nominated as an HHS assistant secretary — walks the plumbing: most-favored-nation and “lesser of” clauses in PBM-pharmacy contracts push cash list prices up, which is what makes the coupon look like a deal.

Host Stacey Richter’s number is the one to keep, though she asserts it rather than sourcing it: for a patient in the deductible phase on the top 20 prescribed generics, she puts it at roughly 80% of the time that the PBM rate is worse than a coupon, Amazon, or Cost Plus. Fair warning — the core interview is a 2021 conversation replayed with a 2026 wraparound.

🔇 Speaker Blindspot: Monocausal reductionism, contradicted inside the same episode. Bai insists GoodRx “makes money from one fact and one fact alone” — and Richter’s own 2026 update then enumerates branded-drug programs, data sales and customer-acquisition revenue. The bigger miss is the question nobody asks: what happens to an independent pharmacy’s margin when a coupon claim adjudicates below acquisition cost? Pharmacy harm gets asserted and then dropped.

💡 BTW

💡 BTW: Ellen DaSilva, who just sold Summer Health to Doctronic, was employee number eight at Hims & Hers and earlier helped scale Twitter’s revenue from $100M to $2B — and she started Summer Health as a mother of young children who couldn’t get a straight answer about her own kids at night. The pediatric text corpus that made her company worth acquiring exists because she wanted somebody to answer her texts first.

💺 Builder Seats

Medical Director, Clinical Product — Clover Health / Counterpart Health · Remote US
The clinical voice inside the product org for an AI value-based-care platform — one of the few posted seats where a practicing physician owns product direction rather than reviewing it. Requires primary care boards plus 10 years PCP experience, and there’s an automated screen on both.
🔗 LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Reply and I’ll include it.

📅 Upcoming: the EHIgnite Challenge Phase 1 Winners Showcase on August 6 — nine teams demoing AI that turns dense EHI exports into something a patient or clinician can use.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin

⚡ Around the Wards

• Thirteen U.S. physicians read 750 fetal ultrasounds twice — once blind, once with AI — mean AUC went 68.9% → 90.9%, sensitivity 54.2% → 88.5%. Full treatment below.

• The AI you didn’t buy is already running in your building — HIMSS is now naming embedded AI in sensors, devices and phones as an unmapped cyber-resilience surface. Nobody has the inventory.

• CMS is ending the subsidy that hid the IRA’s Part D bill — the 2027 national base beneficiary premium is $41.33, up from $38.99, and October’s open enrollment is the first one in three years running on real prices.

• CMS says analytics stopped $203M in improper Medicaid payments in 88 days — 50 high-risk providers flagged, 42 federal exclusion notices, 15 state enforcement actions. Nobody published a false positive rate.

• 🎧 Podcast: NVIDIA GTC SJ 2026 — “The AI Native Digital Health Stack: A Developer’s Guide to 2026” — Nemotron 3 ships in three open sizes (30B, 12B, 500B) with a 1M-token context, and the whole training data and tooling stack is open too. That’s the on-prem inference story for anyone.

🧭 The Curbside

“MCP just shipped a new spec. Does that change anything if I’m pointing an agent at health data?”

• Short answer: Yes — mostly on the authorization side, and mostly in your favor.

• What changed: The 2026-07-28 spec landed with a stateless protocol core, header-based routing (Mcp-Method and Mcp-Name), cacheable list results, and four authorization changes that pull MCP toward how OAuth 2.0 is actually specified. Clients must now validate the iss parameter on authorization responses per RFC 9207 — a cheap mitigation for authorization-server mix-up, which matters more in MCP’s one-client-many-servers shape than in a typical web app. Client credentials are now bound to the issuer that minted them, and Dynamic Client Registration is formally deprecated in favor of Client ID Metadata Documents. Deprecated behavior keeps working for at least twelve months.

• Builder read / Watchout: If you’ve wired an agent to a FHIR server, you’re already living in OAuth — the SMART on FHIR launch is an OAuth flow. Bringing MCP’s auth model in line with what enterprise identity teams already run is the difference between “our IAM group will look at this” and “no.” Watch out for the obvious overclaim: this is a protocol spec, not a compliance posture. Nothing here makes your MCP server an approved system for anybody’s PHI.

  😤 “That’s dev-tool trivia, not clinical.” The authorization model is the specific reason your last project died in security review. Trivia is what we call the thing we didn’t have to think about.

🔬 The Big Thing

Thirteen doctors, 750 fetal ultrasounds, and the number nobody is going to quote

Thirteen U.S. physicians — maternal-fetal medicine, OB/GYN, and radiology — read 750 fetal ultrasound still images twice, once unassisted and once with an FDA-cleared assistant looking for eight specific abnormal findings.

With AI, mean AUC rose 21.9 points, from 68.9% to 90.9%. Sensitivity went from 54.2% to 88.5%, specificity actually improved rather than degraded, and interpretation time fell from 40 seconds to 23 seconds per image.

That’s the press release. Here’s the part that stopped me.

Unassisted, these physicians caught 54.2% of the abnormalities. Unassisted, their inter-reader agreement was 26%.

The lift is not the finding. The baseline is the finding — and the baseline is the number almost no clinical AI study bothers to publish.

Absence of the cavum septum pellucidum. Absence of the corpus callosum. Thoracic situs inversus. These are not subtle in the abstract, and every one of these readers can find them. On a single still frame, stripped of the sweep and the context and the second look, three trained specialties agreed with each other a quarter of the time.

That is a statement about how hard the task is under those conditions, and we only learned it because somebody ran the unassisted arm.

😤 “This is a vendor study.” It is, completely, and you should read the competing-interests statement before the abstract: five of the authors are full-time Sonio employees, four more sit on Sonio’s scientific advisory board, and Sonio funded the work. That doesn’t make the numbers fake. It does mean the design choices — retrospective, still images, 250 abnormal cases out of 750 — were all made by the party with a stake in the answer.

😤 “A third of the images were abnormal. That’s not a clinic.” Correct. Enriched prevalence inflates how a reader behaves; nobody scans 750 studies expecting 250 anomalies. The specificity number is the one to hold loosely.

😤 “So a machine beat doctors, again.” No. A machine beat doctors doing a task nobody actually does — reading one frozen frame with no sweep, no history, no repeat view. That’s the honest read, and it’s still interesting, because the thing being measured is exactly what the AI sees too.

🧪 Try the interactives:

A — The Unassisted Arm — 750 fetal ultrasound stills, 13 physicians, read twice: unassisted they caught 54.2% of abnormalities and agreed with each other 26% of the time. One graphic, both arms, plus the prevalence slider the study didn’t run. Built with real study data.

B — Cleared Before Measured — 369 FDA 510(k) clearances for imaging-AI devices, 2016–2026, on one brushable scatter. Sonio Suspect cleared in 91 days; its unassisted-reader study published 522 days later. The clearance record has no column for the baseline. Built with real FDA 510(k) data.

📡 Builder’s Radar

The AI you never approved is already in the building

Anne Snowdon, HIMSS’ chief scientific research officer, is now naming embedded AI as a cyber-resilience problem — the models running inside sensors, monitors, infusion pumps and smartphones that arrived with the hardware and never went through anybody’s AI governance process. In a companion segment the same day, she says health system leaders are increasingly demanding proof that AI tools solve their specific problems in measurable ways, rather than accepting vendor promises.

Those are two halves of one shift: the buyer wants an inventory of what’s running and a comparable way to score it.

The governance conversation is moving from “should we approve this AI” to “what AI is already here that we never approved.”

💡 80/20: Pick one device class in your department — the monitors, the pumps, the ultrasound carts — and try to build the model inventory. Vendor, model name, what it infers, whether the output is advisory or automatic, last update. You will not finish it. The place you get stuck is the finding.

CMS pulls the subsidy that was hiding the IRA’s Part D bill

CMS announced it will end the Part D Premium Stabilization Demonstration after 2026. The 2027 national base beneficiary premium goes to $41.33 from $38.99, and the demonstration’s $10 premium reduction and its increase cap disappear entirely.

The demonstration existed to soften the premium shock from the IRA’s Part D redesign. Ending it means October’s open enrollment is the first in three years where beneficiaries see the unsubsidized number.

Two open enrollments have been running on a price that wasn’t real. The next one isn’t.

🔮 My bet: plan-comparison and benefits-navigation tools see their best acquisition quarter ever between October 15 and December 7, and at least one of them gets acquired by a pharmacy chain before spring. Confusion is the demand signal.

CMS found $203M in improper Medicaid payments in 88 days, using analytics

CMS says its Medicaid Fraud War Room, stood up April 23 with OIG and state partners, used data analytics to flag 50 high-risk providers tied to roughly $203.3 million in payments, producing 42 federal exclusion notices and 15 state enforcement actions in under three months.

😤 “Show me the false positive rate.” Yeah.

A chest X-ray model that learned where to look by reading the report

Emory- and University of Chicago-led researchers published CF2Seg, a segmentation framework that learns spatial representations directly from the free text of radiology reports rather than from hand-drawn masks, validated on a 53,386-exam multi-source benchmark.

The reason this matters isn’t the segmentation. It’s that the supervision signal is something every health system already has millions of, sitting unused: paired images and the reports somebody already wrote about them.

Your unlabeled archive may be less unlabeled than you think.

Dementia risk models built for the population instead of the average

University of Maryland researchers used a mixture-of-experts transfer learning approach across 490,031 UK Biobank participants to build population-specific dementia risk models, improving accuracy for underrepresented Black and Asian participants, then validated in the All of Us cohort.

😤 “Subgroup models are just overfitting with a nicer name.” Sometimes. And sometimes the pooled model was quietly overfitting to the majority the whole time, and nobody checked because the aggregate AUC looked fine.

Ultra-shorts

Where AI requests actually come from. CIOs at Rush, HSS, and Allina Health described very different internal demand patterns — revenue cycle and supply chain at one, clinician requests routed through an intake tool at another, HR at the third. If you’re building for “the health system,” you’re building for three different buyers who don’t talk to each other.

Vanderbilt and Siemens sign an $87M value partnership. Imaging and radiation oncology equipment plus planned work on data infrastructure and AI-enabled tools. Equipment deals are becoming data deals with a hardware invoice attached.

🎙️ From the Pods

🎙️ Radio Advisory — “308: How digital health innovation is centering the patient, with Rock Health”

Megan Zweig of Rock Health cites AMA survey data that 86% of U.S. physicians have at some point reviewed a patient’s wearable data — and that in only 6% of those cases was the data integrated into the EHR or the clinical workflow. The rest is a patient holding up a phone.

💡 Builder take: That 80-point gap between “clinicians engage with this data” and “this data is in the chart” is the least glamorous, most obviously real integration problem on the board right now. Nobody’s front door needs another dashboard; the pipe is the product.

🔇 Speaker Blindspot: Appeal to authority, via cap table. The strongest evidence offered that wearable-centered care models are the future is that Mayo Clinic and Abbott put money into Whoop’s Series G. Strategic investors buy optionality on a lot of things that don’t happen — and the same segment concedes the data isn’t validated, isn’t reimbursed, and isn’t integrated. Capital flow is a signal about belief, not about outcomes.

💰 Money Plumbing

Budget neutrality — why one specialty’s raise is another specialty’s pay cut

The Medicare Physician Fee Schedule is budget-neutral by statute. CMS cannot add money to the pool; it can only move money inside it. So when you read “CMS proposes to invest in primary care,” the correct next question is always from whom.

The CY 2027 proposed rule (released July 14, comment period open) does something structural: it phases out decades-old specialty practice-expense survey data and revises how indirect practice expense is allocated. Because the formula must net to zero, updated data doesn’t just correct values — it transfers them. Axios (July 22) put it plainly: the revamp “could force steep reimbursement cuts next year for specialists like dermatologists and orthopedic surgeons” while boosting what primary care doctors can charge. CMS’s own specialty-impact estimates put numbers on it — dermatology −9%, otolaryngology −9%, orthopedic surgery −7%, hand surgery −5%. Separate from all of that, the conversion factor drops from $33.57 to $32.84 for non-APM participants — a 1.68% cut that lands on everyone.

Do the arithmetic: a dermatology practice with $2M in annual Medicare revenue loses roughly $180,000/year to the practice-expense redistribution alone, plus about $33,600 to the across-the-board cut. A primary care practice at the same volume sees something in the range of a 1–3% gain — which mostly just offsets that cut. (These are my estimates applied to the published specialty-impact percentages, not figures CMS publishes.)

💡 Builder move: Build for the specialties that are losing. A dermatology or orthopedic practice about to give up 9% of Medicare revenue has a sudden, dated, quantified need for revenue capture — denial management, prior-auth automation, out-of-network optimization. The winners have a less urgent problem, which means a longer sales cycle. And the Patients First Act, introduced July 15, would provide annual positive updates tied to inflation for Part B physician payment if it passes — but it wouldn’t touch the redistribution, because the redistribution is a data update, not a policy choice. Build for the part that isn’t going to get patched.

💡 BTW: Yinka Oyelese, MD — one of the physician authors on today’s fetal ultrasound study, and one of the four on Sonio’s scientific advisory board — trained as an OB/GYN twice, on two continents. Medical school and residency in Nigeria at the University of Ibadan, then he moved to London in 1992 and trained again under Stuart Campbell, the man who more or less invented obstetric ultrasound, before coming to the US in 1999. He has spent the thirty years since on vasa previa — a condition whose entire treatment is seeing it on the scan before the delivery.

💺 Seats

MD, Principal Clinical AI Evaluation Strategist — Elsevier · On-site
The evaluation seat, at the company that owns a large slice of the evidence layer clinical AI cites. If you’ve been building evals as a side habit, this is the job description version of it. (Comp seems low)
🔗 LinkedIn

Director, Engineering — AI-First Healthcare Transformation — Humana · Charlotte, NC / Remote
Payer-side AI engineering leadership. Rare vantage point: you’d be building on the side of the table that sees denial and utilization data at national scale.
🔗 LinkedIn

Know someone hiring for a clinical AI or informatics leadership role? Reply and I’ll include it.

📅 Upcoming: Preparing Health Plans for the 2027 CMS Prior Authorization Rule (AHIP/eviCore) runs today at 2:00 PM ET.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin

⚡ Around the Wards

• Duke Health built its own chart reader rather than wait for a vendor — Scout reads years of notes, labs and imaging and returns a source-linked summary in one to three minutes; a preprint RCT found ~40% faster task completion and 40–50% lower cognitive workload. Enterprise-wide since early July.

• Roughly 80 AnMed sites went dark after a malware attack — the four-hospital South Carolina system closed its medical group offices and its imaging services while EDs and urgent care stayed open. Offices reopened Tuesday with the network still down. Downtime is a clinical event, not an IT event.

• A practice-management vendor breach exposed more than 425,000 patients — Unlimited Technology Systems, one vendor, many unrelated provider organizations, one blast radius.

• 🎧 Podcast: The Heart of Healthcare — “Why We Demand Perfect Machines Yet Tolerate Human Carnage” — Jonathan Slotkin, neurosurgeon and Geisinger CMO for strategy: being safe and feeling safe are two different engineering problems, and healthcare learned that the expensive way.

🧭 The Curbside

“A vendor told me their clinical AI is ‘validated.’ What should I actually make them show me?”

• Short answer: Ask which cases, held by whom, scored inside what harness. If any of those three answers is “public benchmark,” they haven’t answered you.

• What changed / Evidence: Outside healthcare, Legora published its BAR writeup this month — a benchmark drawn from a corpus of ~5,100 real legal cases across 28 practice areas, each reviewed by a legal professional. BAR itself is a private subset of a few hundred of them, deliberately not open-sourced to prevent contamination (one synthetic case is public so you can see the shape), and scored end-to-end inside the production harness rather than a sandbox.

• Builder read / Watchout: That’s the shape clinical eval should be and mostly isn’t. A MedQA score tells you the model can pass a test written for humans; it tells you nothing about your harness, your retrieval, your prompts, or your patients. The watchout cuts both ways — a private corpus is uncontaminated and unauditable, so “trust us, it’s held out” is a claim you can’t check either.

“We want to put an MCP server in front of our FHIR data. Is that safe?”

• Short answer: The protocol isn’t the risk. The auth layer is, and most MCP tutorials skip it entirely.

• What changed / Evidence: PropelAuth published a full walkthrough of OAuth 2.1 with MCP, including the token-handling and scoping details that a “connect your EHR to Claude in five minutes” demo leaves out.

• Builder read / Watchout: Distinguish clearly between demoing against a synthetic FHIR sandbox (fine, do it tonight) and pointing anything at production PHI (needs a BAA, scoped tokens, an audit trail, and your security team’s signature). The MCP ecosystem is maturing fast on capability and slowly on identity. Nothing about running a model locally or self-hosting a server changes the BAA question — that’s a contract, not an architecture.

🔬 The Big Thing

Duke spent ten years on the plumbing. The language model took months.

Duke Health put Scout into general use across the enterprise in early July — an in-house tool that reads a patient’s entire record, years of notes, labs, imaging and pathology, and returns a synthesized summary in one to three minutes, inside Epic, with every claim linked back to its source document.

They ran a controlled trial first, then a three-month pilot: 200-plus users across 20-plus specialties, 6,600-plus interactions. Pre-visit planning, discharge summaries, tumor board prep. Training is required before you get access.

The number that will get quoted: the preprint RCT found clinicians completed tasks 37.6% faster, with significantly lower perceived workload — the biggest drops in mental demand, effort and time pressure. Chief digital officer Jeffrey Ferranti, MD told Becker’s the summaries came out “as good or better than our senior clinicians.”

But the number that should get quoted is ten years.

Duke didn’t build a summarizer. Duke spent the better part of a decade building secure programmatic access to its own EHR data — SMART on FHIR panels inside Epic, the governance around who can touch what, the treatment-relationship gating. Then the models arrived and, in Ferranti’s words, “layering a large language model on top of it and making it secure and safe for our providers to use was not as heavy a lift as you might imagine.”

That sentence is the whole thesis of this newsletter with the polarity flipped. We keep saying the engineering barrier collapsed. It did — for the model layer. The part that didn’t collapse is access to the data, and the institutional patience to build it before there was anything to put on top.

😤 “This is Epic’s summarizer with a Duke sticker on it.” No. Epic’s runs off Epic’s model of your chart. Scout runs off Duke’s own data-access layer, with Duke’s own prompts, tuned per specialty, and Duke tracks how those prompts perform in practice. That last clause is the difference between a feature and a product — somebody at Duke owns whether the tumor-board prompt is still good in November.

😤 “A preprint isn’t evidence.” Correct. It’s an RCT that hasn’t been peer reviewed, from the institution that built the thing, reporting on its own tool. Read it as a strong signal and a weak proof. Also notice what it measured — speed and cognitive load, not a patient outcome, not a missed-finding rate.

❓ If the moat is the data-access layer and not the model, who sells the data-access layer? Duke built theirs over a decade because nothing existed. I keep thinking there’s a company in the gap between “we have Epic” and “we can programmatically reach our own records safely” — something between an integration consultancy and a product. I can’t quite see its shape yet.

🧪 Try the interactives:

A — One Vendor, Many Letters — sixteen years of large healthcare breaches reported to HHS, replayed month by month, and the three moments when a single vendor incident surfaced as dozens of separate filings under other organizations’ names. Built with real HHS OCR breach data.

B — The Blast Radius Ledger — every large breach reported to the HHS Office for Civil Rights since 2009, 939 filings of 50,000+ people and 586 million records, on one brushable scatter with vendor-linked breaches in red: their share of affected people grew from 18% to 63%. Built with real HHS OCR breach data.

📡 Builder’s Radar

Roughly 80 sites went dark, and the ED stayed open

AnMed closed its medical group offices — which reach from upstate South Carolina into Hartwell, Georgia — along with all five of its South Carolina imaging sites on Monday after confirming a cybersecurity disruption involving malware. Urgent care, kids care, integrated therapy and labs stayed open; EDs kept seeing patients. Counts vary by outlet — 79 of 106 facilities, “more than 80,” “dozens.” AnMed itself has published no number.

That split is the whole story. The parts of the system that can run on paper ran. The parts that can’t, closed.

And the sequel makes the point twice: physician offices reopened Tuesday while phones, internet and computers were still down, clinical teams working on downtime procedures, patients asked to bring their medications in the original containers. A dozen sites — imaging, radiation oncology, sleep and cardiac diagnostics, women’s diagnostics — stayed shut. The reopening is sorted exactly along the same line as the closure.

Every tool you build is one of those two things, and you get to choose which.

😤 “Downtime procedures are an IT problem.” They’re a clinical problem wearing an IT badge. If your tool has no degraded mode, you have written a workflow that deletes itself under stress — which is exactly when the workflow matters.

A physician wants the EHR to record the moment he overruled the safe answer

Doug Fullington, a practicing internist, argues that charts capture the world after a decision and never the decision itself — specifically the moment a clinician judges the guideline-supported “safe” option insufficient and does something else. He proposes a four-part schema: the expected default, what was actually chosen, what changed the threshold, and what happened next.

His two cases are the kind you recognize immediately: an “unlikely” DVT that was real, and a portal refill request that wasn’t routine.

Every clinical AI is trained on the output of decisions and blind to the judgment that produced them. This is a data model for the missing column.

(Disclosure in his post: he holds a stake in a clinical-intelligence company.)

💡 80/20: Add the refusal field to whatever you’re building this week. When your tool proposes something and the user does otherwise, capture why in one free-text box and one structured reason code. That log is the most valuable dataset in your product and almost nobody is collecting it.

Digital mental health for kids: the hybrid works, the app-only mostly doesn’t

A meta-analysis of 57 studies covering 43,973 children and youth, out yesterday in npj Digital Medicine, found universal low-intensity digital mental health interventions produced small but statistically significant effects on 9 of 16 pooled effect sizes — and that hybrid formats, in-person plus digital, showed more consistent and better-sustained benefit than purely virtual delivery. The authors are careful in a way the headline won’t be: most pooled findings were low-certainty evidence, and the hybrid advantage at follow-up is moderate-certainty.

😤 “So digital mental health doesn’t work.” It works less than the deck said and more than the cynics said. The finding isn’t “don’t build” — it’s “don’t build the standalone app.” The delivery model is the intervention. If your product only works when a human is also in the loop, that’s not a weakness in the product, that’s the product.

Ultra-short: Flourish Health raised a $26M Series A led by B Capital with F-Prime and Cherryrock, to scale intensive in-home and community youth behavioral healthcare. High-touch clinical labor as the product, software as the margin.

🎙️ From the Pods

🎙️ The Heart of Healthcare — “Why We Demand Perfect Machines Yet Tolerate Human Carnage”

Jonathan Slotkin, MD — neurosurgeon, Geisinger’s CMO for strategy and growth — reads his Noema essay on why autonomous vehicles keep winning on safety data and losing on trust, and maps it directly onto thirty years of hospital patient-experience work. The trust-killers patients name are almost never clinical errors; they’re the used bandage on the floor, the bloodstain on a scrub top. During San Francisco’s December signal blackout, when traffic lights went dark across a third of the city, Waymos behaved exactly as coded — treating dead signals as four-way stops, queuing for remote confirmation faster than the ops team could clear it — and the phone video of them sitting frozen with their hazards on became, in his phrase, the bandage on the floor of a self-driving car. He’d argued in the New York Times a month before that resistance to these vehicles was an ethical failure costing lives. Then in January he sat in one that stopped for a construction dumpster and wouldn’t move through a green light, and he couldn’t tell whether the remote agent or the car had solved it.

💬 Standout Quote

“Being safe and feeling safe are not the same”

🔇 Speaker Blindspot: Special pleading. He calls AV crash counts “a numerator with no denominator” and demands a federal reporting mandate — then builds his own safety case on the one company’s self-published miles, on geofenced, mapped, mostly low-speed roads, against an all-roads human baseline. Mandatory verification for everyone else; voluntary self-report is fine for the incumbent whose data he needs.

🎙️ The Gist Healthcare Podcast — “Rethinking patient safety through the lens of access”

Victor Hassid, MD, MBA — plastic surgeon and AVP of access strategic operations at MD Anderson — describes reclassifying “a patient called and did not leave with an appointment” as a never event, the same category as wrong-site surgery. It’s the operational version of his recent JAMA Viewpoint, Patient Safety Begins With Access. Enterprise median days-to-appointment-offer went from five days to two, and to a single day for patients who’ve had no prior care and arrive with the fewest records. The mechanism wasn’t software. It was putting the clinicians who own the scheduling logic in the same room as the people running the phones.

🔇 Speaker Blindspot: Texas sharpshooter, with a side of composition fallacy. The headline metric is median days to appointment offering — a process measure — not time to treatment or any outcome. And a single NCI-designated cancer center with global self-referral and no capacity crisis is not the system that would have to adopt this.

🎙️ Podnosis — “HaloMD’s Patrick Velliky explains why No Surprises Act IDR enforcement matters” (sponsored episode — worth knowing before you weigh the argument)

The arbitration process under the No Surprises Act is legally binding but has no private right of action, and roughly 45 federal court decisions have gone against physicians who sued to collect. So some payers simply don’t pay awards they lost. Physicians are paid in arrears; insurers collect premium in advance and earn float on every unpaid day. The worst case for a non-compliant payer is eventually paying exactly what it already owed.

🔇 Speaker Blindspot: Availability heuristic plus a conspicuous omission. The central claim rests on unnamed payers and undisclosed samples — “examples that we are aware of directly,” no denominator. And he never mentions provider-side IDR gaming, batching or volume flooding, which is unsurprising given that his company’s revenue scales with dispute volume and paid for the episode.

🧰 Builder’s Tip

Weekend project: turn a public clinical guideline into an agent skill.

You have a specialty guideline PDF you re-read four times a year. Convert it once into a structured skill your agent can load on demand, instead of pasting the whole thing into context every time.

book-to-skill is an MIT-licensed skill that takes a technical PDF and produces a SKILL.md carrying the core mental models and a chapter index, plus a chapters/ file per chapter and a glossary, a patterns file and a cheatsheet — all in the open Agent Skills format that Claude Code, GitHub Copilot CLI and Amp all read. Per-chapter loading means a 200-page document costs tokens proportional to the question, not the page count.

git clone https://github.com/virgiliojr94/book-to-skill.git ~/.claude/skills/book-to-skill

Then point it at something public and clinically real — a society guideline, a CMS manual chapter, an FDA guidance document. Nothing you use here touches a patient.

Friday night: convert one guideline. Saturday: write ten synthetic cases, including three where the guideline’s answer is wrong for the patient in front of you. Sunday: run all ten and read the failures. By Sunday night you’ll know something concrete about where retrieval breaks on your own specialty’s literature — and you’ll have a reusable artifact, not a chat log.

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin & AI