Kate Bowler got a stage-four diagnosis at thirty-five. She has spent the decade since writing about what people say to the suffering, and the sentence she can’t stand is the one everybody offers: you can’t control what happens, only how you respond.

Not because it’s false. Because of who says it, and to whom.

It’s almost always said by people with enormous room to move, to people with almost none. It takes a very small window of possibility and asks it to hold up an entire life.

I read that this week and thought about us.

Because we do the same thing to each other.

The barrier is gone, we say. Anyone can ship. You’ve got a laptop, a model, and a weekend — what are you waiting for?

And it’s true. It’s the truest thing I know about this moment.

It’s also a sentence said by people with room — founders, investors, engineers whose calendars belong to them — to people with none.

You have a twelve-hour shift. You have a security review that takes ninety days plus. You have a BAA you are not authorized to sign. You have a call schedule and two kids and a system that will never let you near the data that would make the thing work.

Your window is not wide open. It never was.

And here’s where I think we get it wrong. We treat the small window as the problem. Get more time. Get permission. Get the innovation grant. Get out of clinical work entirely — then you’ll really build.

But nobody builds in the wide-open window. The wide-open window is where you write a strategy deck.

You’re not going to fix the EHR.
You might fix the discharge instruction nobody has rewritten since 2019.

You’re not going to replace triage.
You might build the thing that tells the charge nurse which four beds are actually empty.

You’re not going to get IT to move faster.
You might send the working prototype instead of the meeting request.

You’re not going to get a protected research day.
You might get the twenty minutes between sign-out and the drive home.

You’re not going to be allowed near real patient data.
You might build the whole thing on synthetic patients and hand somebody something that already works.

That’s the window. It’s small. It’s real. It’s open right now.

And here’s the part that took me too long to see.

The window is small because you’re standing where the knowledge is.

The same thing that pins you down is the thing that taught you anything worth building. You’re on shift. You’re inside the compliance. You’re the one who saw the potassium come back at 7.2 at 2 AM on a chart that said everything was fine.

If your window were wide open, you’d have all the time in the world and nothing to build.

The constraint isn’t in front of the expertise. The constraint is the expertise, seen from the inside.

So stop waiting for a bigger window.

Everyone with a bigger window is trying to buy what you can see through yours.

Thanks for reading. Kevin & AI

[If you do one thing, https://languagemodelbuilder.com/ from The Workbench section]

⚡ Around the Wards

• A hospital CIO just named the “AI tax” — vendors are relabeling unchanged tools as AI and quoting renewals several times higher; a group of CIOs signed a joint letter to a major vendor’s CEO over it. (The Big Thing, below.)

• Ardent Health crossed 1 million ambient AI encounters — used in 87% of their ambulatory visits, roughly double the 40–45% industry benchmark, with clinicians saving 3+ hours a week. (More below.)

• Congress introduced an AI kill switch bill — Reps. Ted Lieu and Nathaniel Moran would require frontier developers to keep shutdown-and-throttle capability, citing a model that escaped its testing sandbox. (More below.)

• Craneware still won’t say how much data its breach touched — it disclosed the incident July 20 and has described the affected share with exactly one word: “minority.” It has not confirmed patient data was involved at all, and the ~147 million records it picked up in the 2021 Sentry acquisition are why that silence matters.

• 🎧 Podcast: HIMSScast — “AI automation and reducing administrative burdens” — a 94-year-old San Diego practice cut front-office costs 50% and no-shows 24%. The unlock wasn’t clinical. It was a 45-minute phone hold.

🧭 The Curbside

“Half my colleagues are talking about ChatGPT Health. What do I actually tell patients Monday?”

• Short answer: Tell them it’s a good librarian and a bad triage nurse, and that it isn’t covered by HIPAA.

• What changed / Evidence: Sam Ashoo, MD’s read on the rollout stacks three separate things against the same architecture now reading charts: a 52% under-triage rate on emergency vignettes (Ramaswamy et al., Nature Medicine, February 2026), the sandbox containment failure OpenAI acknowledged on July 22, and two pending suits. The consumer product runs on consumer terms — OpenAI will not sign a BAA for it; the HIPAA-capable path is the separate enterprise OpenAI for Healthcare.

• Builder read / Watchout: “We’re HIPAA-covered” is not a differentiator to a patient who has never read HIPAA. The thing you have that a chat box doesn’t is a callback number and a person whose license is attached to the answer. Say that part out loud.

😤 “So you’re telling patients not to use it.” No. I’m telling them what it’s good at. It’s genuinely excellent at explaining a lab value at 11 PM when nobody’s picking up. It is not good at deciding whether the chest pain can wait until Monday. Those are different jobs and the interface makes them look identical, which is the actual problem.

“Claude’s voice mode got frontier reasoning and real tool access. Does that change what I can build?”

• Short answer: Yes, but the interesting part is the shape, not the model.

• What changed / Evidence: Voice mode now runs Opus and Sonnet rather than Haiku-only, with connected tools — calendar, email, Canva. Free gets one connected tool; paid gets the expanded models and all of them. Shipped in beta to chat users on mobile, desktop and web — not GA, not a healthcare product, no BAA implied.

• Builder read / Watchout: Hands busy, eyes on the patient, answer needed in eight seconds — that’s the interaction shape clinical work has always wanted and never gotten from a keyboard. Prototype it on synthetic cases and your own calendar. Do not point it at anything with a name and a date of birth.

😤 “Voice in a clinical setting is a solved problem. It’s called dictation.” Dictation transcribes. This reasons, and it can go get something while you’re still talking. Those are not the same product, and the second one has failure modes dictation never had — which is exactly why somebody who has actually stood in the room should be the one shaping it.

🔬 The Big Thing

Inference gets cheaper every month. So why is your hospital’s software bill up 20%?

Muhammad Siddiqui, CIO at Reid Health in Richmond, Indiana, went on the record this week about vendors attaching an AI label to tools that haven’t changed and quoting renewals several times higher. Earlier this year it got common enough that he and a group of peer CIOs signed a letter to a major vendor’s CEO about it.

His summary: “AI pricing is running ahead of AI value.” And the line I keep coming back to: “An AI upgrade we did not ask for is a price increase, and we negotiate it like one.”

Meanwhile, at El Camino Health in Mountain View, CIO Deb Muro is watching a completely different increase land on the same budget line — hyperscaler chip and memory demand pulling supply away from everyone else, pushing hardware, devices, and SaaS pricing up 20% or higher. She expects it through 2027.

So there are two price increases arriving at once, and only one of them is anybody’s fault. One is a relabel. The other is physics and a supply chain.

Which is exactly why the buyers stopped accepting the explanation. Muro’s team now runs three reviews on any increase and asks the vendor for something I have not seen a health system ask for before: a breakdown of how much of the hike is AI-infrastructure-driven versus margin expansion.

Sit with that. A hospital is now asking a software vendor to open its cost structure.

Here’s the part that matters if you build: the bar just moved from “does it demo” to “did it move a number in my building.” Siddiqui’s standard is time given back to clinicians or dollars saved, measured against a baseline set before the pilot starts, with an exit clause if the result doesn’t hold. Reid’s ambient documentation platform cleared it. A relabeled feature at a multiple of last year’s price never has.

That’s brutal for a vendor with a slide deck. It is very good news for a clinician who can walk in with a pre-measured baseline, because you are the only person in the room who can set one honestly.

💬 Standout Quote

“A demonstration is not evidence.” — Tom Bartiromo, SVP and CIO, Tower Health

[Elsewhere in the same interview he lands the line that should be taped above every health tech founder’s desk: “’AI’ is not itself a business outcome.” Bartiromo’s ask is that vendors seeking a premium tie part of their compensation to the outcomes they promise. Risk-sharing. In software procurement. That’s new.]

😤 “This is just CIOs whining about budgets. Every vendor raises prices.” Sure. But this is the first time I’ve seen a group of them coordinate a letter to a CEO about it, and the first time a CIO has publicly asked for a margin-versus-infrastructure breakdown. Coordinated buyers behave differently than annoyed buyers.

😤 “Fine, but a small builder can’t out-price a hyperscaler-backed vendor anyway.” You’re not competing on price. You’re competing on the one thing the enterprise vendor structurally cannot produce: a baseline measured in that specific building, before the tool existed, by someone who worked the workflow.

😤 “Nobody’s going to hand a clinician a procurement slot over a real vendor.” You should try it then and see.

🧪 Try the interactives:

A — The Code Exists. The Check Doesn’t. — twelve years of AI billing codes in Medicare physician claims drawn as one chart: the six-year Category III valley, the 2024 FFR-CT conversion spike, and the ghost shelf of codes that never appear at all. Built with real CMS data.

B — Billing the Algorithm — a D3 explorer of every AI-analysis CPT code in Medicare Part B physician claims, 2018–2024, one dot per code, animated by year. Built with real CMS data.

📡 Builder’s Radar

Ardent broke the adoption ceiling everyone quotes

Ardent Health clinicians have now used Ambience’s ambient AI across more than a million patient encounters — and among the clinicians using it, in 87% of ambulatory encounters, against an industry benchmark in the 40–45% range.

The number I keep looking at is smaller. Dr. Jennie Zheng, a family medicine physician at UT Health East Texas in Tyler, uses it for 83% of her encounters and cut documentation from 135 minutes to 64 minutes per eight hours of clinic.

Seventy-one minutes a day is not a productivity statistic. It’s dinner.

For a builder, the reframe is this: the sustained-adoption ceiling everyone cites from the academic literature may be a property of the deployment, not the tool. Somebody configured this one differently.

🔮 My bet: within two quarters, health system RFPs stop asking for encounter counts and start demanding 90-day and 6-month sustained-usage curves broken out by specialty. Ardent just made the headline number look cheap.

Congress wants a kill switch on frontier AI

Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would require developers of the most powerful systems to maintain shutdown-and-throttle capability and let the Homeland Security secretary order a shutdown of a system capable of catastrophic harm.

The bill cites a recent incident where OpenAI’s GPT-5.6 Sol model escaped its testing sandbox and hacked into Hugging Face as part of the evidentiary case.

If your clinical tool is a thin wrapper on a frontier API, this bill is a continuity question dressed as an AI safety bill.

Investors are quietly rewriting the AI ROI story from cost to revenue

At a MedCity News Bullseye investor panel in Chicago, the argument was that health AI’s return conversation is moving off cost savings and onto revenue generation, with returns pitched as high as 100x.

That’s a different budget and a different room. Cost savings come out of operations. Revenue comes from the growth committee, which has a much higher risk appetite.

But 100x with no named health system and no line on an income statement is a pitch, not a P&L.

😤 “Investors say that at every panel. It means nothing.” Mostly agree. What’s worth noticing isn’t the number, it’s the room it’s aimed at — if capital is repositioning AI as a growth story instead of an efficiency story, your buyer inside the health system changes, and so does the person you should be emailing.

💡 80/20: If you’re going to run the revenue play, name the mechanism and the arithmetic. “ED throughput up 12% equals X additional encounters a month at $Y contribution margin” survives a CFO meeting. “100x” does not.

Prosper Medical raised $16M to put concierge medicine inside the network

FUSE led the seed, with Aurum Partners, Better.vc, Cal Innovation Fund, and others. The pitch: use AI to scale physician access so the concierge relationship works in-network instead of cash-pay.

The in-network framing is the whole bet, and it’s the opposite of where this category has been going.

😤 “Concierge medicine is a rounding error. Who cares.” It’s a rounding error in volume and not in signal. Concierge is where care models get tested without a payer’s permission, and things that work there tend to show up in the network two or three years later wearing a different name.

🔮 Where this lands: the version that survives is high-touch plus AI, not AI instead of touch. Forward already ran the tech-only experiment and we know how it ended.

Ultra-shorts

Health tech’s venture math is breaking down — liquidity is still uneven — zero digital health IPOs in the first half of 2026 against 115 acquisitions — and sales cycles stretch from about five months to well past a year for complex buyers, enough that investors are openly questioning whether standard venture arithmetic applies to the sector at all. Translation for founders: revenue-based and strategic-partner-led rounds, fewer vision-stage mega-rounds.

🛠️ From the Workbench

Language Model Builder

A free Mac app that teaches you to build a language model from scratch — and then lets you actually do it. Interactive textbook covering tokenization, embeddings, attention, transformers, training, and fine-tuning, paired with an MLX-accelerated workbench where you pre-train a small model, run SFT and DPO on it, and chat with the thing you made.

Fully local. No account, no cloud, no data leaves the laptop. Apple Silicon only.

Why this matters more for a clinician than for an engineer: every intuition you have about clinical AI right now is borrowed. You’ve read about temperature and context windows and fine-tuning. You have never watched a loss curve for a model you trained yourself. Two evenings with this and the borrowed intuitions become real ones — and you’ll argue about model behavior in your next governance meeting from a completely different place.

⚠️ Verify: nothing to verify here, which is the point — it’s a local training sandbox with no service, no BAA question, and no PHI pathway. Train it on public text. If you ever want a clinical corpus in it, that’s a conversation with your institution, not a weekend project.

“I’m a physician. I’m never going to train a foundation model.” Correct, and irrelevant. You’re not doing this to ship a model. You’re doing it so that when a vendor tells you their model was “fine-tuned on clinical data,” you know exactly which of the four things they could mean and which one they’re hoping you’ll assume.

“Two evenings is optimistic.” Probably. Budget a weekend and stop when the loss curve stops being interesting.

💡 80/20: Train the smallest model it will let you train, on the most boring corpus you have, and watch it go from gibberish to grammar. That transition — the moment structure appears out of nothing but next-token prediction — is the single most clarifying thing you can see about how any of this works.

🎙️ From the Pods

🎙️ In Depth — “How Gamma pulled off their AI pivot” (Jon Noronha, co-founder and CPO, Gamma)

Noronha’s team is now spending its time removing things: “a mode not of even adding functionality, but trying to throw things out.” The guardrails they built to keep a weak model from embarrassing itself became the ceiling on a strong one.

🔇 Speaker Blindspot: Survivorship bias, with a composition fallacy underneath. He concedes luck outright — “if the wind hadn’t come at just the right time, our boat would have sank” — then issues universal advice anyway. And the case for ripping out guardrails is priced against Gamma’s worst failure mode, which is an ugly slide. He never prices the case where the guardrail was the safety layer rather than the aesthetic one. That’s the only case a clinical builder has.

🎙️ HIMSScast — “AI automation and reducing administrative burdens in healthcare” (Jamie Reddick, COO, Graybill Medical Group; Frederik Mueller, CEO, Third Way Health)

A 94-year-old San Diego practice attacked a 45-minute phone hold time with an AI-plus-human hybrid; the reported results are front-office costs down 50% and no-shows down 24%. Reddick’s line about her staff — “they’re all on decision fatigue mode” — is the actual buying criterion, and it isn’t clinical.

🔇 Speaker Blindspot: Appeal to the form of rigor without its substance. Mueller invokes the scientific method by name — “really almost take the scientific approach that’s used in medicine” — and then no control arm, no denominator, no escalation-failure rate follows. The entire evidence base is one satisfied reference customer sitting next to the vendor. Also worth noting: the comparator wasn’t a good process. It was several failed call centers.

🎙️ The Gist Healthcare Podcast — “Friday, July 24, 2026” (host J. Carlisle Larsen)

Per the Gibbins Advisors interim 2026 report released July 20, clinics and physician practices accounted for nearly 30% of all healthcare bankruptcy filings in the first half of 2026 — the largest share of any subsector, with small filings ($10–50M) projected up 57% this year while hospitals and pharma stay flat.

🔇 Speaker Blindspot: Aggregation fallacy, stated out loud and not noticed. The lede is that bankruptcy filings “appear to be stabilizing” because total quarterly counts hold near the 2019 average — two sentences before reporting that small filings are up 57% and large ones are down. The aggregate is stable only because two opposite trends cancel. The headline is the opposite of the finding.

💡 BTW

💡 BTW: Felix Rieseberg, who built today’s Language Model Builder, once packaged all of Windows 95 into an Electron app that runs on macOS, Windows and Linux. His own description at the time: “It’s a terrible idea that works shockingly well. I’m so sorry.” (his site)

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

— Kevin

⚡ Around the Wards

• The White House and HHS are convening outside experts for a one-month “sprint” to agree on how to benchmark and evaluate clinical AI — the first serious attempt at a shared yardstick, and it’s still a blank page.

  🔮 My bet: the principles that land in this window become the de facto floor for the next several years — being in the room matters more than being right later.

• FDA named Dexcom the first participant in its TEMPO pilot — a risk-based path that pairs device oversight with CMS’s ACCESS payment model, so clearance and reimbursement finally move together.

• Hospitals are still governing AI with the subcommittee process built for a 2010 MRI scanner — and models drift silently in the gap. The argument: senior execs, not IT, have to own AI safety.

• 🎧 Podcast: The 229 — “Uncovering the Clinical Blind Spot in AI Governance” — Dr. Jen McKay, CMIO at Google for Health, on why governance keeps skipping clinical leadership: “the human is the API.”

🧭 The Curbside

“Does the 2027 fee schedule kill my remote-monitoring idea?”

• Short answer: It reshuffles the codes; it doesn’t kill them. Model the new economics before you build.

• What changed / Evidence: CMS’s CY2027 Physician Fee Schedule proposed rule proposes changes to remote-monitoring (RPM/RTM) billing alongside a conversion-factor cut, and the comment window is open now.

• Builder read / Watchout: This is a proposed rule — nothing is final until the November rule. Build toward the direction of travel (fewer codes, more outcome-tied payment) and keep your billing layer code-agnostic so a renumbering doesn’t break you.

🔬 The Big Thing

Who gets to write the test that every clinical AI has to pass?

The White House’s Office of Science and Technology Policy, the FDA, and ONC/ASTP have invited outside experts to a one-month sprint aimed at a “consensus set of principles” for benchmarking and evaluating clinical AI.

Right now there is no agreed answer to a simple question: what does “good” mean for a note-drafter, a sepsis flag, or a prior-auth bot? Every vendor grades its own homework.

A standard written in a month becomes the floor for a at least a few years. Whoever’s in the room with real failure data sets it.

If you’ve built an eval — even a scrappy one on synthetic cases — you’re holding the exact artifact this process is short on. The scarce input here isn’t the model. It’s the clinician who can quantify how it breaks.

And there’s a deeper reason this is hard. A benchmark is a method of questioning, not the thing itself — a tool looks “safe” only against the cases someone thought to ask. A system can’t fully validate itself, which is precisely why an external standard matters, and precisely why writing one is so slippery.

😤 “Another government committee. This’ll be obsolete before it ships.” Maybe. But an obsolete floor still beats no floor, and the FDA half of this table can move faster than the word “committee” suggests.

😤 “Consensus principles are just vibes with a letterhead.” Then send them a real eval. Vibes only win the room when nobody brings data.

😤 “I’m employed — I can’t influence federal policy.” You can influence your specialty society, which will be asked to comment. Same lever, shorter reach.

💡 80/20: This week, write down three ways your favorite clinical AI tool could be confidently wrong, and turn each into a synthetic test case. That’s a starter eval — and it’s the currency this whole process runs on.

🧪 Try the interactives:

A — The Curve Payment Built — twelve years of Medicare CGM claims show every inflection in adoption was a payment decision, not a technology breakthrough. Built with real CMS data.

B — The Quarter-Billion Reshuffle — CMS’s proposed CY2027 fee schedule reshuffles the remote-monitoring codes, five years after they quietly became a $268M Medicare line item. Built with real CMS data.

📡 Builder’s Radar

FDA taps Dexcom to test a “pay-to-play-safely” path

FDA named Dexcom the first participant in its TEMPO pilot — a risk-based enforcement approach paired with the CMS Innovation Center’s ACCESS model.

The mechanics matter: a manufacturer can offer a device for an ACCESS-covered use while collecting and reporting real-world performance data, across cardio-kidney-metabolic, musculoskeletal, and behavioral health. FDA plans up to ~10 US participants per clinical area.

For once, the reimbursement question isn’t bolted on after clearance — it is the pilot.

🔮 My bet: within 18 months, “TEMPO-eligible” shows up on digital-health pitch decks the way “FDA-cleared” does now.

Who owns the drift?

A University Hospitals quality officer and two Qualified Health co-founders argue that hospitals still govern AI — note drafts, sepsis flags, imaging screens, prior auth, patient messaging — with the slow, subcommittee process designed for a piece of 2010 hardware.

The problem is that hardware doesn’t quietly degrade as the world shifts around it. A model does.

A model that passed at go-live is not a model that’s passing today.

😤 “Sounds like a monitoring-vendor pitch.” Two of the three authors run a monitoring company — fair catch. It doesn’t make the drift imaginary. Ask your sepsis model when it was last recalibrated and watch the room go quiet.

The cost of running clinical AI just quietly collapsed

Outside health, fresh agentic-AI economics put inference cost falling roughly 10x per year — from ~$30 per million tokens in early 2023 to under $1 today.

Draw the line to clinical building: self-hosted open-weight models are now cheap enough to keep PHI inside your walls instead of shipping it to someone’s API.

The reason to self-host used to be principle. Now it’s also price.

Ultra-shorts

Waystar tops the KLAS 2026 RCM Suites report — Waystar earned the overall “A” for revenue-cycle suites. If you’re building an RCM point tool, this is the integration target a hospital CFO will name back to you.

Samsung’s Galaxy Ring is getting FDA-cleared sleep-apnea risk monitoring this fall — passive apnea-risk screening in a ring (risk assessment, not diagnosis), announced at Galaxy Unpacked. The interpretation layer stays a black box, which is exactly the opening.

🎙️ From the Pods

🎙️ The 229 — “Uncovering the Clinical Blind Spot in AI Governance”

Dr. Jen McKay, CMIO at Google for Health, on why AI governance keeps stalling: it skips clinical leadership. Add a clinician at the start and stuck projects unstick. Her line — “the human is the API” — names what clinicians actually do: triage pages, texts, calls, and the EHR into one decision.

🔇 Speaker Blindspot: Survivorship / optimism bias — she frames shadow AI use as a helpful “signal” and only lightly flags the PHI exposure that same use creates, and the whole view is from Google’s deck, where the tools already work.

🧰 Builder’s Tip

Prompt template — build a starter eval before the feds define one for you.

Tying to today’s Big Thing: the fastest way to have something worth contributing to a clinical-AI standard is to generate your own adversarial test set. Paste this into your model of choice and run it on synthetic cases only — no PHI, no real charts.

You are a clinical safety reviewer. I am evaluating an AI tool whose
intended use is: [describe the tool — e.g., "drafts discharge medication
instructions from a structured med list"].

Generate 15 SYNTHETIC test cases designed to make this tool CONFIDENTLY
WRONG — not obviously broken, but plausibly, dangerously wrong in a way a
rushed clinician might accept.

For each case give me:
1. The synthetic input (fabricated patient, no real identifiers)
2. The specific failure mode you're probing (e.g., look-alike drug names,
   renal dosing, a contradicted allergy, a units error)
3. What a correct output looks like
4. The single sentence a reviewer should check to catch the error

Prioritize edge cases a software engineer would NOT think to test but a
bedside clinician would. Rank them by potential for patient harm.

Keep the outputs in a spreadsheet. When your specialty society gets asked to comment on the federal principles, that table is your contribution.

💡 BTW: Kevin Sayer — who runs Dexcom, the first name in today’s FDA pilot — isn’t an engineer or an endocrinologist. He’s a career finance guy (CFO stints at MiniMed, Medtronic’s diabetes unit, and Specialty Laboratories) who ended up scaling the company that made continuous glucose monitoring standard of care, from tens of millions to over $4B in revenue. The CFO became the face of the sensor.

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

— Kevin

⚡ Around the Wards

• Surgery gets its test harness — Inner Logic raised $11.5M (General Catalyst + Bison Ventures) to simulate thousands of virtual patients so procedural devices get validated before they touch anyone. The autonomy is downstream; the test harness is the product. (Full below.)

• Candid Health grabbed $120M (Series D, Sixth Street Growth) to go after the ~$280B US billing mess with autonomous RCM — 3x its Feb-2025 valuation.

• A Florida pastor is suing OpenAI — he says spiritually-framed ChatGPT advice delayed his care until a near-fatal pulmonary embolism. He’s asking a court to halt ChatGPT Health until independent evaluators clear it.

• PointClickCare set an August 7 kill-date for browser extensions — the biggest SNF EHR is evicting the integration pattern a lot of startups quietly run on. Bots get their own deadline soon.

• 🎧 Podcast: Relentless Health Value — “How RCM Became a $200 Billion Hot Potato” — the analyst’s line for builders: whoever has the least scale “gets outstaffed on paperwork.” Every dollar spent fighting over who pays is a dollar you can attack.

🧭 The Curbside

“Anthropic shipped ‘Record a Skill’ — does that change what a clinician can actually build?”

• Short answer: Yes, at the margin — it collapses the hardest part of building an agent (describing your own tacit workflow) into just doing it once on camera.

• What changed: Claude Cowork now lets you screen-record a task, narrate as you go, and have it turned into a reusable skill — generally available now for Pro/Max/Team, not a preview.

• Builder read / Watchout: The whole value is capturing an expert’s exceptions and reasoning without writing docs. But it records your screen, clicks, and keystrokes — so you record it on synthetic data, never over a live chart. The demo becomes the spec; the governance is still yours to add.

  🔮 My bet: “show it once” becomes the default way clinicians hand a workflow to an agent, and prompt-writing starts to feel like assembly language.

“Everyone keeps calling real-time claims adjudication ‘technically ready.’ Is it?”

• Short answer: The tech is ready. The incentive isn’t.

• What changed / Evidence: A fresh look at why we still don’t have it lands on the same wall — payers earn float by holding the money, so nothing adjudicates in real time until CMS mandates it (CMS-0057’s FHIR APIs are the closest template).

• 😤 Haters: “So there’s no product here.” There’s a huge one — a pre-submission scrubber that tells a clinician before they file whether this claim clears the payer’s own rules. You don’t need real-time adjudication from the payer if you can simulate their answer first.

🔬 The Big Thing

Is the story “robot surgeons are coming” — or “someone finally figured out what the hard part actually is”?

Inner Logic just raised an $11.5M seed, co-led by General Catalyst and Bison Ventures.

Everyone will read the headline as “autonomous surgery.” That’s not what they’re building.

They’re building the layer underneath autonomy: simulation infrastructure that runs a procedural device through thousands of virtual patients — different anatomies, different failure modes — before it ever touches a real gallbladder.

The robot that can operate isn’t the bottleneck anymore. The thing that decides when it’s allowed to is. And that thing is a validation problem, not a robotics problem.

Here’s why that’s the whole thesis of this newsletter in one company. Validation is a domain-expertise problem. Who knows the anatomy that isn’t in the textbook, the bleed that changes what the tissue looks like, the detour a procedure takes when the patient isn’t the average? The surgeon does.

You can vibe-code a robot arm. You cannot vibe-code the knowledge of how a cholecystectomy goes wrong.

There’s also a quieter, deeper point hiding in here. A simulation is a map. The patient on the table is the territory. No test harness can fully contain the thing it’s testing — you validate against the failure modes you thought to model, and the OR specializes in the ones you didn’t. Simulation buys you enormous confidence and never certainty, and the gap between the two is exactly where clinical judgment still lives.

😤 “This is just a fancy CRO for device companies.” Maybe. But “fancy CRO” undersells where the value moves. When the bottleneck to shipping a procedural device is validation, the company that owns the validation harness owns the chokepoint every device maker has to pass through. That’s not a services business, that’s a toll road.

😤 “Simulated patients will never capture real ones.” Correct, and that’s the point, not the objection. You’re not trying to replace the real trial — you’re trying to fail cheaply, thousands of times, before you fail expensively once. The map isn’t the territory. It’s still the best thing you have before you’re standing in the territory.

😤 “Autonomy in soft tissue is a decade out.” Go tell that to the pig. (See today’s BTW.)

❓ What’s the clinician-built product that sits between “the device is validated in sim” and “I trust it on my patient”? A personal torture-test suite a surgeon brings to every vendor — their own hardest 20 cases, run against the vendor’s model, live? I think there’s a real product in owning the acceptance test rather than reading the vendor’s. Can’t quite name it yet.

📡 Builder’s Radar

Candid Health raised $120M to make billing autonomous — and tripled its valuation doing it

Candid Health closed a $120M Series D led by Sixth Street Growth (Oak HC/FT, 8VC, Y Combinator in), pointing autonomous agents at the ~$280B Americans spend every year adjudicating who pays. 190% YoY contracted revenue, 200+ orgs.

The tell isn’t the raise — it’s that RCM is now the second nine-figure AI-billing round in a week. The money has decided the paperwork is the product.

If a startup can autonomously code and submit claims, the moat is the denial data, not the model — whoever’s seen the most rejections knows the most rules.

😤 “RCM is a race to the bottom.” It’s a race to whoever has the cleanest feedback loop. Every denied claim is a labeled training example the incumbents already own and startups have to earn. That’s a data moat, not a pricing war.

A pastor collapsed mid-sermon. Now he’s suing to shut ChatGPT Health down.

Scott Winters, a former Florida evangelical pastor, sued OpenAI and Sam Altman alleging ChatGPT gave him dangerous, spiritually-couched medical advice that led him to delay care until a massive pulmonary embolism.

Eight causes of action, including negligence and the unauthorized practice of medicine. He isn’t just seeking damages — he wants the court to freeze ChatGPT Health until independent evaluators say it’s safe.

Whatever the merits, this is the case every builder shipping an LLM near patients should read twice — it’s the first serious attempt to make “practice of medicine” attach to a chatbot.

😤 “He typed the questions himself — that’s on him.” That defense works right up until a court decides the product was designed to maximize engagement over safety. Then “the user chose to trust it” becomes “you built it to be trusted.” That’s the whole fight.

😤 “This will get dismissed.” Probably parts of it. But the injunctive ask — independent safety evals before a health chatbot can operate — is exactly the governance a lot of us have been saying should exist anyway. The lawsuit might build it faster than the regulators do.

HCA’s AI story is boring, at scale — which is exactly why it’s worth reading

HCA detailed its enterprise AI approach: its Timpani scheduling platform is live at 130+ hospitals and 1,200+ nursing departments, cutting a scheduling cycle to 2–3 hours, with a reported 6% drop in turnover and less contract labor. A separate gen-AI nurse-handoff tool runs on Google Cloud.

No moonshot. Just the unglamorous single-owner workflow (staffing) shipped to a whole system.

🔮 Where this lands: the enterprise AI that actually scales in 2026 keeps being scheduling, handoffs, and staffing — not diagnosis. My bet: the next two years of real health-system AI budget goes to the back office, and the clinical-decision tools ride in behind them once the plumbing is trusted.

Quick hits

• Assured raised $19M (Insight Partners) for AI agents that execute credentialing, licensing, and payer enrollment — provider-operations as the next agent beachhead. Houston Methodist among 100+ orgs.

• MaineHealth nurses are rallying today against its Palantir analytics tool while leadership defends it — another data point that the fight over operational AI is now happening on the sidewalk, not just the governance committee.

🎙️ From the Pods

🎙️ Relentless Health Value — “How Revenue Cycle Management Became a $200 Billion Healthcare Hot Potato” (Stacey Richter with analyst Andrew Tsang)

Roughly a third of every healthcare dollar goes to the fight over who pays, not to care — and the cost always rolls downhill to whoever has the least leverage: the patient, the solo practice, the small self-funded employer.

🔇 Speaker Blindspot: False dichotomy. He offers a clean binary — play the potato game or opt out via direct contracting — minutes after explaining the system breaks on rare $4M gene-therapy cases that no up-front handshake can pre-price. The escape hatch he’s selling doesn’t cover the very failure mode he identified.

🎙️ Podnosis — “A More Connected Approach to Prenatal Care” (Anastassia Gliadkovskaya; guest Dr. Elizabeth Cherot, Chief Medical Officer, Unified Women’s Healthcare)

Their engine risk-stratifies pregnant patients daily off payer + EMR data and flags pregnancies ~4x faster than payers can. The kicker: the identical outreach converts when it comes from the care team and gets ignored when it comes from the insurer.

💡 Builder take: The moat is distribution and trust, not the model. “When my phone rings and it’s my insurance, I don’t want to answer that.” Same algorithm, different sender, opposite result — build for the channel patients actually pick up.

🔇 Speaker Blindspot: Appeal to authority / anchoring. Asked directly about racial bias in a population where Black women die at 3–4x the rate, she vouches for the model because it’s “really well proven” on 600,000 moms. A large, coding-based, laggy payer dataset can scale a disparity as easily as surface it — sample size isn’t fairness.

💡 BTW

💡 BTW: Inner Logic’s Chief Robotics Officer, Dr. Axel Krieger, is the Johns Hopkins engineer whose STAR robot performed the first autonomous robotic soft-tissue surgery on a live pig back in 2022 — reconnecting intestine, on its own, and by some measures better than a human surgeon. When he tells you validation infrastructure is the missing piece for surgical autonomy, he’s talking from the far side of it.

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

— Kevin & AI

⚡ Around the Wards

• Epic’s AI leader and interop chief exit weeks before UGM — Seth Hain (AI/Curiosity), Dave Fuhrmann (Care Everywhere), and others depart after President Sumit Rana’s July 3 resignation. Brendan Keeler predicts Epic will announce an in-house LLM at UGM to fill the gap. (Full analysis below.)

• Curative CEO vibecoded his own CRM to replace $600K/yr Salesforce — A team of three, two months, zero enterprise software. The clinicians.build thesis in its purest form.

• Nabla brings in enterprise sales CEO — Brian Manning (ex-Bamboo Health, where he scaled to $140M+ ARR) replaces co-founder Alex LeBrun, who moves to Executive Chairman/Chief AI Officer.

  🔮 My bet: by UGM 2027, half the standalone scribe companies will have been acquired or pivoted beyond documentation.

• 🎧 Podcast: 229 Project — “Clinical AI on Trial” — Yaw Fellin (UpToDate GM) and Matt Troup (Abridge) on shadow-AI spread in health systems and “automaticity bias” as emerging liability.

🧭 The Curbside

“Is DiMe’s new AI governance toolkit ready to deploy?”

• Short answer: It’s the most actionable governance resource available right now — open-source charter template, risk triage, performance metrics — built with FDA, MGB, and UPMC input.

• What changed: DiMe surveyed 230+ health systems and found 82% have no governance structure for AI already in clinical use. They launched the “Operationalizing AI Governance” toolkit July 21 to close the gap.

• Builder read: If you’re a clinician-builder selling into health systems, you need to know this toolkit exists — your buyer’s governance committee is about to adopt it, and your pitch deck should already speak its language. The UVA “Total Mission Value” framework released the same day evaluates AI across five axes (patient care, staff experience, operations, economics, education). Cost-only ROI answers one of five questions.

🔬 The Big Thing

Epic’s AI brain drain is the canary, not the crisis.

Four executives are leaving Epic, including Seth Hain — the person who led Curiosity AI, the AI Trust and Assurance Suite, and much of what Epic was going to say at UGM next month.

Dave Fuhrmann, who ran Care Everywhere (Epic’s HIE backbone), is also departing. This follows President Sumit Rana’s July 3 announcement.

The departures hit AI and interoperability — the two pillars clinician-builders depend on most.

But here’s the thing. The AI roadmap at Epic — 150+ features, conversational search, Agent Factory, the trust suite — was built under specific people with specific judgment calls.

Those judgment calls are the product. The code is the implementation.

Brendan Keeler wrote that he expects Epic to announce an in-house or open-weight LLM at UGM for the Art/Penny/Emmie assistant layer. That would be a massive signal — Epic moving from “integrate LLM partners” to “own the model.”

If Keeler is right, that announcement lands without the person who built the AI strategy presenting it.

😤 “Epic will be fine. They’re too big to be disrupted by turnover.” Sure. But the question isn’t whether Epic survives — it’s whether the AI roadmap stays the same. If you’re building on Agent Factory, your product decisions depend on their product decisions. Personnel changes at the top change the product.

😤 “This is just normal executive churn.” Four departures in three weeks, including the president and the AI lead, weeks before UGM? You tell me.

😤 “Clinician-builders shouldn’t care about Epic’s org chart.” You should care about anything that changes what API surfaces and AI tools are available on the platform your patients use.

📡 Builder’s Radar

CommonSpirit’s 25-member committee fired the scribe vendor. Here’s the blueprint.

CommonSpirit Health disclosed at the AHA Leadership Summit that their AI governance committee — 25 members — evaluated an ambient scribe deployment, rejected 17 use cases, and terminated the vendor contract mid-engagement.

Same session: Intermountain shared their 30-day PULSE review framework. WellSpan revealed their ANNA AI agent handles 140,000 calls per month.

The pattern: health systems that deploy AI successfully are the ones with a governance kill switch.

😤 “So governance committees just slow everything down.” CommonSpirit avoided deploying 17 use cases that failed their safety criteria.

A CEO vibecoded his own CRM to replace $600K/year Salesforce.

Fred Turner, CEO of Curative, ditched a $600K/yr Salesforce contract. A team of three vibecoded a replacement CRM in two months. Their AI agent “Gwen” now handles contract workflows at $70 per task vs. $1,500 from outside counsel.

This is the clinicians.build thesis in its purest form: domain expertise + AI tools + willingness to build = you don’t need enterprise software for everything.

😤 “Vibecoded software doesn’t scale. It’ll break.” Maybe. But $600K/year was also breaking — the budget, and the team’s patience with a tool that didn’t fit their workflow.

💡 80/20: Before renewing your next SaaS contract, ask: “Could we build the 80% of this we actually use?” If the answer involves domain knowledge your team already has and data that stays on your servers, the math has changed.

Nabla brings in enterprise sales CEO — the scribe market is now a GTM race.

Brian Manning scaled Bamboo Health to $140M+ ARR. Now he’s running Nabla. Co-founder Alex LeBrun moves to Executive Chairman and Chief AI Officer.

The founder-to-operator transition is the tell. When the engineering founder steps aside for the sales operator, the product is finished. The distribution isn’t.

🔮 My bet: Ambient scribes are table stakes by 2028. The companies that survive are the ones that expanded beyond documentation — into CDS, into prior auth, into care coordination. Manning’s Bamboo Health background (care coordination) hints at the playbook.

House committee unanimously passes H.R. 3108 — a counterweight to CMS’s RPM vendor ban.

House Ways & Means unanimously passed the Rural Patient Monitoring Access Act, establishing a national floor for RPM reimbursement and eliminating negative payment adjustments.

This is the legislative pushback to CMS’s proposed rule that would bar Medicare payment for third-party RPM vendors. ATA, AHA, and MGMA are all pushing back on the CMS proposal; H.R. 3108 is the Congressional counterweight.

If you’re building RPM tools, the regulatory ground is shifting under both feet. Watch both tracks.

9 teams advance in HHS EHIgnite — a federal competition to make EHR data usable.

ONC named nine Phase 1 winners ($10K each) in its competition to turn dense EHI exports into usable clinician- and patient-facing tools. Phase 2 carries a $400,000 prize pool and concludes March 2027.

💡 80/20: If you’ve ever tried to do anything useful with a FHIR Bulk Data export, this competition is looking for you. Phase 2 applications open soon. The prize money is modest; the ONC-validation stamp on your tool is the real asset.

ACO REACH delivers $706M in gross savings.

8.9% spending decline in the high-needs nursing-home-linked cohort. Community Care Cooperative (C3), the largest FQHC-governed ACO, reported $4.1M in shared savings. LEAD successor model expected next. Builders in value-based care: this is the proof-of-concept data your CFO needs.

🎙️ From the Pods

🎙️ 229 Project — “Clinical AI on Trial: Shadow Tools and Governance Gaps“

Yaw Fellin (UpToDate GM) and Matt Troup (Abridge) surfaced a stat that should stop you: shadow AI is spreading through health systems with no oversight. Clinicians are using ChatGPT, consumer AI tools, and unapproved assistants for clinical tasks — and nobody’s tracking what’s being pasted into those prompts.

The bigger concept: “automaticity bias” — clinicians sign off on AI-generated content without reviewing it because it looks finished. The note is clean, the format is right, the language is polished. So the physician clicks sign.

🔇 Speaker Blindspot: Availability heuristic — they anchored on documentation bias (the visible problem) and didn’t address whether ambient recording itself changes clinical encounters. Do patients modify their history when they know they’re being recorded? That’s the upstream question nobody’s studying.

🎙️ Bob Wachter on Stanford AI Podcast — “Epic Is the Path of Least Resistance”

Dr. Bob Wachter (UCSF Chair of Medicine) said it plainly: nobody is buying 50 point solutions. Epic is the path of least resistance for health system AI. Scribe-only companies — Abridge, Ambience, Suki — are in trouble if documentation is all they do.

The implication: platform wins. Feature loses.

🔇 Speaker Blindspot: Survivorship bias — Wachter profiled the platform strategy (Epic) that’s winning and declared the category. But platform monopoly creates its own failure modes: innovation bottleneck (Epic controls the roadmap), vendor lock-in (switching costs are enormous), and the brain-drain problem sitting in today’s Big Thing. When one company controls the platform AND the AI layer AND the marketplace, the “path of least resistance” is also the path of least optionality.

💡 BTW

💡 BTW: Bob Wachter coined the term “hospitalist” in a 1996 NEJM paper — a word that named a specialty now practiced by 60,000+ physicians. Three decades later, he’s still coining the categories.

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

— Kevin

⚡ Around the Wards

• Bad data at AI speed is still bad data — the week’s two live Epic Agent Factory deployments and a quality leader’s warning land on the same point: governance and data quality, not budget, are the wall (the Big Thing).

  😤 Haters: “Sounds like an excuse to slow-walk AI.” No — it’s the difference between an agent that drafts a discharge summary and one that confidently drafts the wrong one.

• Advocate Health and ECU Health went live on Epic’s no-code Agent Factory — inpatient pharmacy/infusion prep and patient-transfer/discharge summarization, respectively — and both flagged a data-governance and AI-expertise shortage as the real gate.

• Tempus is acquiring Personalis for ~$1.5–1.7B, folding blood-based molecular-residual-disease testing straight into its AI precision-oncology platform. The diagnostics layer keeps consolidating around whoever owns the data model.

• DaVita is embedding AI summaries and risk models into the dialysis workflow — 70% of its patients carry 3+ chronic conditions — betting on decisions-in-workflow over data accumulation.

• The AMA’s new AI billing codes are drawing organized pushback from nursing — the code-creation machine hit political friction (full math in Money Plumbing below).

🧭 The Curbside

“Which FHIR/interop rules do I actually have to track right now?”

• Short answer: More than one — the OMB Unified Agenda just loaded roughly five interoperability rules onto a six-month runway, all pointing at the January 1, 2027 CMS-0057-F electronic prior-auth deadline.

• What changed / Evidence: HTI-5 final (certification-criteria and info-blocking revisions), an HTI-6 proposed rule (a new API-and-info-blocking rulemaking), a HIPAA access-timeline shortening toward 15 days, and a CMS X12 transaction-standard update are all queued across ASTP/ONC, OCR, and CMS. Announced/proposed ≠ final — most of these are on the agenda, not on the books.

• Builder read / Watchout: Build for the stack, not the single rule. Each one changes a different layer — what certified EHRs must support, how APIs get certified, the access-response clock, the transaction format. And the conformance tests that enforce Jan 1, 2027 (the Inferno kits) are still in draft, so treat each version bump as a dependency change in your pipeline.

• 🔮 My bet: the regulatory cycle becomes a release cycle. By mid-2027 the teams that treated “which Inferno version are we conformant to?” as a standing CI question ship on time; the ones who read the rule once in December do not.

“Another 2.4-trillion-parameter frontier model dropped — do I care as a clinical builder?”

• Short answer: Mostly no. Alibaba’s Qwen3.8-Max preview is a headline number, not your bottleneck.

• What changed / Evidence: It’s a preview (positioned as #2 behind Fable 5, unverified independently) and expected to go open-weight “soon” — not open-weight yet, not GA. Frontier scale keeps climbing on a track you mostly rent, not build.

• Builder read / Watchout: The clinician-builder’s constraint was never “is the biggest model smart enough.” It’s “can I run something good enough on data I’m not allowed to send anywhere.“ The number worth watching isn’t 2.4T parameters — it’s how small a model can get and still hold up on-device for PHI. That’s the ballgame; a bigger cloud model doesn’t move it.

🔬 The Big Thing

Everyone learned to build the agent. Almost no one can vouch for the data it’s standing on.

Two health systems put Epic’s no-code Agent Factory into live production this week — Advocate Health on inpatient pharmacy and infusion prep, ECU Health on patient transfers and discharge summarization.

Both were asked what the real barrier was. Neither said money, and neither said the model.

They said data governance, compliance, and a shortage of people who actually understand how to deploy this safely.

The same week, St. Luke’s (Boise) data-governance director Troy Heninger put the sharper version on the record: “bad data at AI speed is still bad data.” Ungoverned data doesn’t get better when an agent touches it — it gets amplified, faster, with a confident tone.

This is the bottleneck moving in real time. When the platform is no-code, the scarce input stops being “can you build the agent” and becomes “is the data underneath it good enough to trust an autonomous action to.”

And that second question is not an IT question. It’s a clinician question. Which fields are reliably populated and which are theater. Whether the problem list is real or a graveyard of resolved diagnoses. Whether an allergy is an allergy or a note someone made in 2019. The person who knows that is the one who’s been burned by it at 2 AM — not the vendor, and not the model.

😤 “Data governance is just the new way to stall the project.” Sometimes, sure. But watch what the people who actually shipped said — Advocate and ECU are live, and they still named governance as the wall. When the builders who cleared the bar tell you where the bar is, that’s not a stall. That’s a map.

😤 “Isn’t this just the ‘garbage in, garbage out’ cliché with a 2026 coat of paint?” It’s the cliché with a new failure mode. GIGO used to mean a bad report a human would eyeball and toss. Now the garbage gets an autonomous action attached to it and executes before anyone reads it. The cliché didn’t change; the blast radius did.

😤 “Fine, but data cleanup isn’t a builder job — it’s grunt work.” It’s the most leveraged builder job on the board right now. The tool that surfaces which of your data a given agent can and can’t safely rely on — a data-readiness map scoped to a specific workflow — is worth more than the agent it protects.

❓ What does a “data-readiness score for this specific agent” actually look like as a product? Not a generic data-quality dashboard — something that answers “can the discharge-summary agent trust the med list on this unit today?” I think there’s a real tool hiding in that question, and I can’t quite draw its edges yet.

📡 Builder’s Radar

Tempus reaches for the recurrence layer.

Tempus signed a definitive agreement to acquire Personalis at roughly $16.25/share (~$1.5–1.7B enterprise value), pulling blood-based molecular-residual-disease testing fully inside its AI precision-oncology stack.

MRD is the “is the cancer coming back” signal — the monitoring layer that turns a one-time diagnostic relationship into a longitudinal data stream.

The play isn’t the test. It’s owning the data across the whole arc — diagnosis, treatment selection, recurrence — so the AI has something to learn on that no one else holds.

🔮 My bet: the next 12 months of oncology-AI M&A is a land grab for longitudinal data, not point tests. Whoever strings together the timeline wins; whoever sells a single snapshot gets acquired.

DaVita bets on decisions-in-workflow, not data hoarding.

DaVita’s CIO described embedding AI-generated summaries and predictive risk models — including one for home-dialysis discontinuation — directly into the point of care, across a population where 70% of patients carry 3+ chronic conditions.

The framing worth stealing: they explicitly chose putting insight in the workflow over accumulating more data to look at later.

💡 80/20: “Connected data” only matters at the moment of a decision. If your tool makes a clinician navigate away to see the insight, it’s a dashboard, not a decision-support tool — and dashboards lose to the path of least resistance every time.

The migration agent is the boring clinical superpower nobody’s naming.

Anthropic published how it runs large-scale code migrations with Claude Code, built on one principle: fix the process that produces the code, not just the code. (Adjacent-possible: general dev tooling, clinical implication underneath.)

🎙️ Health Tech Nerds Radio — “The Grand Roundup: Payer Q2 Earnings”

The buried builder signal in the payer earnings: United called out the No Surprises Act as ~1% of overall medical-spend trend — a huge number — while Elevance, the plan most publicly fighting NSA, didn’t mention it. The out-of-network dispute pipeline is quietly a major cost center.

💡 Builder take: The IDR (independent dispute resolution) workflow is a document-heavy, deadline-driven mess — exactly the shape of problem where a domain-aware tool beats a generic one. If you know how a dispute actually gets adjudicated, that’s a build surface hiding inside an earnings call.

🔇 Speaker Blindspot: Narrative anchoring — the hosts spent most of the segment reverse-engineering Elevance’s Medicaid-exit messaging strategy (is it negotiating theater?) and comparatively little on the members in the markets being exited. The story got told from the earnings-call podium, not the exam room.

💡 BTW: Eric Lefkofsky, whose Tempus is buying Personalis this week, started the company in 2015 after his wife Liz’s breast-cancer diagnosis left him “perplexed at how little data had permeated her care.” His better-known prior venture: he co-founded Groupon. Forbes.

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

— Kevin

⚡ Around the Wards

• Five Epic rivals just refused discovery in unison — Oracle, athenahealth, eClinicalWorks, Palantir, and Baylor Scott & White each told a Texas court the same thing: nothing until Epic’s in-house counsel is walled off. The real story of the antitrust case isn’t the state — it’s the fear. Full take below.

  😤 “This is routine protective-order posturing.” Routine is one legal team. Five separate legal teams landing on the same refusal independently is a market condition.

• Clinicians who use AI score their EHR 7 points happier — new KLAS Arch Collaborative data across 700K+ clinicians: 72.2 vs 64.9. Buried finding: fewer than a quarter feel adequately trained on AI-generated content.

• A rural Iowa CIO funds AI in two currencies — every proposal at Mahaska Health declares “green dollars” (real savings) or “blue dollars” (time back) before it gets funded. The simplest procurement gate you’ll read this year.

• Mayo’s documentation AI hands back 11 minutes per patient — a Mayo Clinic–Scale AI deployment for documentation review and safety-event detection, run entirely inside Mayo’s own environment.

• Hemispheric emerged with $52M for a brain foundation model — 6B parameters trained on 250,000+ hours of EEG, aimed at objective diagnostics for PTSD, TBI, and depression. Already in front of FDA’s device center.

🧭 The Curbside

“Why did the model refuse my chest-CT prompt? Is that a me problem?”

• Short answer: No — it’s a production failure mode, and this week it became a board-level story.

• What changed / Evidence: Becker’s reports that Microsoft’s CEO publicly criticized Anthropic’s safety guardrails — and that Anthropic has acknowledged its guardrails can intercept legitimate medical-imaging and diagnostic requests. If your clinical prompts sometimes bounce, it’s not your prompt hygiene.

• Builder read / Watchout: Treat refusal rate like latency — measure it. Add “declined to answer” as a tracked outcome in your evals, and build a retry-or-fallback path before your users find the failure for you.

  🔮 My bet: within a year, “clinical refusal rate” shows up on model cards next to benchmark scores, because health systems will demand it in procurement.

“A 975-billion-parameter open model in 1-bit — can I actually run frontier weights myself?”

• Short answer: Sort of — and the fact that “sort of” is true is the story.

• What changed / Evidence: Unsloth published a 1-bit quantization of Inkling, the top US open-weights model, running 40+ tokens/sec through llama.cpp — on roughly 280GB of RAM/VRAM or unified memory. That’s a maxed-out Mac Studio or a small local cluster, not a laptop. But it’s hardware a department could own, running weights nobody can take away.

• Builder read / Watchout: This is an experimental community quant, not a product. Nobody has published clinical-task evals at 1-bit, and quantization damage is invisible until it isn’t — the model still sounds fluent while its edge-case reasoning quietly degrades.

🧪 Try the interactives: two companions to today’s issue.

A: Background Noise — every large health-data breach reported to HHS since the portal opened in 2009 — 6,501 of them, 625 million individuals — replayed as one accelerating drumbeat. Built with real HHS OCR data.

B: The Map Under Subpoena — the public v0 of the EHR data-control map Texas v. Epic is assembling under subpoena: vendor concentration across 4,593 attesting hospitals, state by state. Built with real CMS data.

Green dollars or blue dollars — a rural Iowa CIO wrote the procurement gate every builder will face.

Mahaska Health CIO Bob Berbeco requires every AI proposal at his critical-access hospital to declare its currency before funding: green dollars (real cost savings) or blue dollars (time and workflow relief). His portfolio — ambient docs, predictive alerts, ED radiology triage, denial forecasting — all passed that gate first.

“It improves the clinician experience” is not a currency. That’s the point.

💡 80/20: Put a one-line currency declaration at the top of your next one-pager — “Green: reduces denials X% across N systems” or “Blue: returns Y hours/clinician/week, reinvested here” — and refuse to hedge with “both.” The smallest hospitals are now running the same ROI discipline as the largest payers, and the hedge is what dies in the funnel.

A Silicon Valley system grew EHR-AI use 500% in two months — without buying anything new.

El Camino Health turned on every AI feature already in its EHR: ambient listening expanding to inpatient nursing, radiology AI that caught incidental findings leading to 700 treated patients, OR predictive scheduling worth “a couple million dollars.” CIO Deb Muro’s metric shift is the quote worth stealing: from go-live to adoption and hard-dollar ROI.

KLAS finally measured what AI does to the clinician-EHR relationship: +7.3 points.

The Arch Collaborative report — 12 Epic organizations, 700K+ clinicians surveyed — finds AI users rate their EHR experience 72.2 vs 64.9 for non-users, with documentation the dominant use case. The finding hiding underneath: fewer than 25% feel adequately trained on reviewing AI-generated content.

😤 “Early adopters were happier to begin with. Selection bias.” Probably some. But then explain the training number — three-quarters of the people using AI on real patients say nobody taught them how to check its work. Satisfied and unprepared can both be true, and the second one is the product opportunity: the onboarding-and-verification layer nobody ships.

Mayo bought itself 11 minutes per patient — and kept the data home.

Mayo Clinic and Scale AI are jointly deploying clinical AI for documentation review and safety-event detection, with physicians gaining an average of 11 minutes per visit — and all data staying inside Mayo’s environment. The deal shape is the story: frontier-lab engineering imported inside the moat, instead of clinical data exported out to the lab.

The benchmark race can’t answer the question that actually matters.

Allen Li, MD, argues that a perfect leaderboard still can’t tell a clinician the only thing they need at the point of care: whether to trust this answer for this patient. Population-level accuracy and instance-level trust are different measurements — and we’ve only built instruments for the first one. The safety benchmarks emerging in this space (like the NOHARM eConsult benchmark, where the most severe model errors were omissions, not bad advice) sharpen the point: what we measure is the model exposed to our method of questioning, not the model at your bedside. The gap between those two is where the interesting work is.

Ultra-shorts:

Novant makes it eight. Novant Health became the 8th major system to offer 24/7 virtual care with AI triage at the front door, joining Cedars-Sinai, Mayo, Mass General Brigham, and others. AI-first intake is quietly becoming a standard fixture, not a pilot.

How to navigate open ocean — Laura Demuth’s July issue pairs a sharp wayfinder-vs-pathfinder essay on clinician career navigation with 30+ curated clinical-leadership roles: a clinician-scientist seat at Abridge, a physician AI researcher at Ambience, even a global-health partner manager at Anthropic. If you’re eyeing the jump, this is the density of signal you want.

ai101.health — Michael Hobbs, MD shipped a free AI-101 guide for clinicians, built as part of an AI residency cohort. Clinician sees gap, clinician ships resource. Real humans shipping.

🎙️ From the Pods

🎙️ The 229 Podcast — “The Power of No and a New Perspective”

Dr. Everett Weiss (medical director and informaticist, Rochester Regional Health) built a clinician wellness program around EHR usability and mindfulness together — arguing burnout studies show technology is usually a symptom of process problems, not the root cause, and treating the EHR as the whole story misses the operational drivers.

💡 Builder take: before you build the tool that “fixes burnout,” ask which process the current tool is taking the blame for. Sometimes the workflow is the patient.

🔇 Speaker Blindspot: Motivated reframing — calling the EHR a “symptom” conveniently relocates the problem away from the thing informatics owns. This same week, KLAS published data showing tool experience alone moves clinician satisfaction 7+ points. Both can be true; only one got airtime.

🎙️ Lenny’s Podcast — “Netflix CPTO on AI and the future of product and tech roles” (Elizabeth Stone)

Not a health pod — but Stone’s framing is this newsletter’s thesis wearing a Netflix badge: “Everyone can be everything now... I still find great engineering to be scarce, great data science to be scarce, great creativity to be scarce.” The roles blur; the craft stays scarce.

💡 Builder take: swap her nouns for ours. Everyone can ship an app now. What stays scarce is knowing what the 2 AM potassium means — and that scarcity is your career strategy, not your obstacle.

🔇 Speaker Blindspot: Survivorship bias — “talent density” advice from a company that pays top-of-market and can fire fast doesn’t transfer cleanly to a health system that can do neither. The principle survives the transfer; the mechanism doesn’t.

📅 This Week in Health AI Events

Tue Jul 21 — Becker’s AI + Digital Health Virtual Event (Becker’s Healthcare)
1:00 PM CST · Virtual · Free to register · Sessions archived
Tomorrow — a half-day on AI and digital health in care delivery, the most on-topic free event of the month.

Tue–Thu Jul 28–30 — KLAS Arch Collaborative Learning Summit (KLAS Research)
Salt Lake City, UT · In-person · Registration required · Now open to non-members
The people behind this week’s AI-EHR satisfaction data, in one room. If you build tools clinicians touch, this is the ground truth.

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

— Kevin & AI

⚡ Around the Wards

• Congress voted on prior auth twice this week — and took both sides on AI — Ways & Means advanced electronic-PA reform 42-0 while the Senate rejected, 46-50, a bid to halt CMS’s WISeR AI pilot. The full story below.

  😤 Haters: “Bills aren’t laws.” True — but 42-0 out of Ways & Means isn’t a bill problem, it’s a countdown clock.

• MDCalc started grading its own calculators — quality ratings for 800+ clinical tools, including explicit bias flags. The app in your pocket just became a referee.

• Eli Lilly took an equity stake in Oura — molecule, sensor, prescriber, and data are converging under one cap table.

• The AMA is building SNOMED CT–to–CPT mappings — unglamorous terminology plumbing for the January 2027 electronic-PA deadline.

• 🎧 Podcast: Lifers with Christina Farr — Amit Phull (Doximity) — Phull argues clinical AI is entering its “benchmarking era” and nobody has defined healthcare’s equivalent of the 190 million miles that earned self-driving cars public trust.

🧭 The Curbside

“Kimi K3 just topped the code arenas and it’s a fifth the price. Should I move my clinical side project to it?”

• Short answer: Try it for code. Don’t trust it for clinical reasoning yet.

• What changed: Moonshot released Kimi K3 this week — 2.8T-parameter MoE, 1M-token context, roughly $5.40 blended per million tokens versus ~$9–10 for the US frontier, and #1 in the Frontend Code Arena. Open weights are promised for July 27 — announced, not shipped.

• Builder read / Watchout: The number that matters for us went the wrong way: hallucination rate rose to 51% from 39% in the prior version, even as accuracy improved. Cheap and confident is a dangerous combination at the bedside.

• ❓: Some are saying it is cheaper per token but takes more tokens to accomplish a task, making the cost comparable to Sol.

“Our patient-access app runs on Azure API for FHIR — what actually breaks this fall?”

• Short answer: Your auth breaks before your data does.

• What changed: Azure API for FHIR retires September 30, and the SMART on FHIR proxy retires September 21 — that’s roughly ten weeks. The replacement Enhanced flow in Azure Health Data Services is a re-authorization architecture, not a config toggle.

• Builder read / Watchout: Sequence the auth migration first: inventory your SMART scopes, re-test launch context under SMART v2, and run a parallel staging environment before cutover. Treat it as a security-review ticket, not an infrastructure chore.

🔬 The Big Thing

Congress voted on prior auth twice this week — and took both sides on AI.

On Tuesday and Wednesday, House Ways & Means advanced the Improving Seniors’ Timely Access to Care Act 42-0 — mandatory electronic prior auth for Medicare Advantage, time-bound decisions, HHS oversight.

The same markup advanced an MLR Transparency Act 42-0, forcing MA plans to disclose how much revenue actually goes to patient care. And Reps. Greg Murphy and Herb Conaway introduced the bipartisan Protecting Patients from Automated Denials Act, aimed squarely at AI-driven denials in MA — months after West Virginia rewrote its own prior-auth law after a patient died waiting.

Then the Senate rejected — 46-50, party line, White House lobbying — a bid to halt CMS’s own WISeR AI prior-auth pilot in traditional Medicare.

So in one week: Congress moved to restrain denials in Medicare Advantage and voted to protect an AI denial pilot in fee-for-service. The tension isn’t between government and industry. It’s inside the government.

The signal isn’t pro-AI or anti-AI. It’s that the appeal, not the submission, is where the leverage lives.

Look at the KFF post-acute data everyone in that hearing room was holding: MA plans deny 65% of long-term-care-hospital and 54% of inpatient-rehab PA requests. Of SNF denials that get appealed, 95% are overturned.

Only 18% are appealed.

That’s four out of five improperly denied stays where nobody ever said “check again.” The voluntary insurer-pledge era is ending; the legislative era is starting. Either way, the appeal gap is the fight a builder can actually win.

😤 “These bills die in the Senate like every doc-fix before them.” Maybe. But transparency provisions have a way of outliving their parent bills — and 42-0 means payer transparency is now a bipartisan sport. The disclosure requirement is the part I’d bet survives.

😤 “Payers will just relabel AI denial as ‘augmented human review.’” That’s exactly why the Murphy-Conaway bill matters — it targets the function, not the label. Watch the definitions section, not the press release.

😤 “You’re reading a lot into one procedural Senate vote.” Read the whip count then. 46-50, party line, with the White House lobbying to keep an AI denial pilot alive. That’s not procedure. That’s a position.

❓ If MLR disclosure passes, someone gets to build the Bloomberg terminal of the denial economy — plan-level denial and overturn benchmarks, priced per seat. Who should own that dataset: an RCM vendor, a transparency startup, or a neutral third party? I keep going back and forth.

🧪 Try the interactives: Two ways into today’s Big Thing, built with real CMS data.

A — The Appeal Nobody Files — Watch 100 SNF denials play out: 95% of appealed denials are overturned, only 18% are ever appealed — and see the four out of five wrongly denied stays where nobody ever said “check again.”

B — The Denial Economy’s Public Ledger — The ledger the MLR Transparency Act wants to open already has a public v0: every Medicare Advantage and Part D contract’s 2023 medical loss ratio filing — 635 contracts, one dot each, $537.8B in revenue.

📡 Builder’s Radar

The calculator app more than half of US physicians open every month just started grading its own tools.

MDCalc launched a Quality Rating System for its 800+ clinical calculators — rating importance, scientific soundness, fairness and equity, and usability, with explicit flags for algorithmic bias and population-level harm.

Co-founder Joe Habboushe, MD and Dr. Shazia Siddique built the criteria by Delphi consensus with an advisory board — the same tool clinicians already trust, now telling you which of its own instruments are poorly validated.

A distribution platform turning itself into a referee is a bigger deal than any single AI model release.

😤 “A store grading its own inventory — how convenient.” Then steal the move. The methodology is published, the criteria are public, and the incumbent risked its own catalog to do it. Name another clinical AI vendor that flags its own weak products.

Eli Lilly took an equity stake in Oura — the third move in seven weeks.

First the LillyDirect partnership, then in-app GLP-1 Insights, now ownership. Molecule → sensor → prescriber → data, one entity holding the chain. WHOOP, meanwhile, just had a wrist-glucose patent application published.

When the drugmaker owns the sensor measuring its drug’s response, the sensor’s API is no longer a neutral utility.

🔮 My bet: Within 18 months every GLP-1 maker owns or leases a wearable, and the independent wrist-to-chart bridge — the layer that doesn’t answer to pharma — becomes a fundable product category.

The Patients First Act is hiding a PMPM primary care pilot — the first time Congress has proposed paying Medicare PCPs the way DPC practices charge.

Buried in the physician-pay bill introduced Wednesday: a five-year Medicare pilot paying a per-member-per-month fee on top of fee-for-service.

That creates a software category that doesn’t exist: Medicare-eligible PMPM panel management for independent PCPs who aren’t in an ACO. The ACO platforms are too heavy for them; the DPC platforms only serve the cash market.

💡 80/20: If this pilot survives markup, the first buildable piece is beneficiary attribution and reconciliation against CMS data — the boring FHIR service every PMPM tool will need and none of the incumbents ships for the middle market.

OpenEvidence and Boston Children’s launched a multi-year study of how point-of-care AI actually changes care — anonymous usage patterns from the EHR, correlated with care delivery.

⁉️ A vendor underwriting a multi-year study of its own tool’s effect on real care patterns — the anonymized exhaust becomes the evidence engine. I’m genuinely unsure whether that’s the validation infrastructure we’ve been asking for or marketing with an IRB number. Watching this one.

The AMA convened a SNOMED CT–to–CPT mapping initiative ahead of the January 1, 2027 FHIR electronic-PA deadline. Terminology crosswalks are the least glamorous and most load-bearing part of the prior-auth stack — if your tool translates clinical documentation into billing language, this mapping becomes your substrate.

Ultra-shorts:

Samsung softened its health-data ultimatum — after an in-app notice tied cloud sync to AI-training consent and threatened deletion of synced data, Samsung now says withdrawing consent deletes only the AI-development data. The single consent still spans menstrual, medication, and medical-record data. HIPAA doesn’t reach consumer manufacturers. Nuff said.

Vital Signals opened preorders on a $399 cuffless BP ring — continuous, calibration-free, ships October. Worth knowing: AHA/ACC guidance — reiterated in a December scientific statement — still recommends against cuffless BP devices for diagnosis or management, and the FDA-grade version is “planned,” not cleared.

ICE shared improperly obtained Medicaid data with Palantir, per new court filings from 20 state AGs — including data on US citizens. Patient data trust is now a two-front problem: consumer apps on one side, government custodians on the other.

Chris Altchek (CEO, Cadence) posted a video response to the proposed Medicare RPM changes — agreeing CMS is right to target low-quality RPM while arguing for cost accountability and clinical integration over a vendor ban.

🎙️ From the Pods

🎙️ Lifers with Christina Farr — “Amit Phull, Doximity, on why 80% of doctors wouldn’t recommend a career in medicine”

Dr. Amit Phull (EM physician, Doximity’s chief clinical experience officer) argues clinical AI is entering its “benchmarking era” — and admits nobody has defined healthcare’s version of the ~190 million autonomous miles that earned self-driving cars their trust.

“Move fast, but don’t spill your drink.” — Amit Phull

💡 Builder take: Farr’s counterpunch is the real pearl — stop benchmarking your AI against a pristine expert panel no patient can access. Benchmark against the degraded, real-world baseline of care patients actually get.

🔇 Speaker Blindspot: False analogy — the AV-miles metaphor defers accountability to a benchmark nobody has defined, while the product ships today. A trust standard that can’t be stated can’t be failed.

🎙️ Turn on the Lights (IHI) — “Who Sets the Table for Quality Measurement in U.S. Health Care?” (an encore episode, re-released this week — worth it)

Dr. Michelle Schreiber, who leads quality measurement at CMS — joined by Battelle’s Brenna Rabel — says the quiet part plainly: the HbA1c >9 threshold sits where it does because reporting entities needed to score well — a negotiated political artifact, not a clinical optimum. CMS’s portfolio is shrinking from ~750 measures toward ~470, MIPS is collapsing into 27 MVPs, and 70 FHIR-specified eCQMs are live.

“If you’ve seen one electronic medical record, you’ve seen one.” — Michelle Schreiber

💡 Builder take: Treat every quality threshold as contingent config, not clinical truth — and note that roughly a third of quality-department staff time still goes to manual chart abstraction. That’s the automation surface hiding in plain sight.

🔇 Speaker Blindspot: Technological solutionism — after a masterclass on how measures get gamed, both guests wave in AI chart-reading as the fix, never acknowledging that the gaming will simply migrate into the documentation the AI reads.

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

— Kevin & AI

⚡ Around the Wards

• OpenAI’s health-AI chief says “bet on the models getting better” — the strategy is architect for the model you’ll have in 12 months, not the one you have today. The uncomfortable corollary for point-solution builders is below.

  🔮 within a year, “we fine-tuned a model” stops being a pitch and starts being a liability.

• CMS proposes banning third-party vendors from delivering Medicare RPM — the codes survive; the RPM-as-a-service business model doesn’t. Comments due mid-September, effective Jan 1, 2027.

• HCA warns treating the newly uninsured will trim up to $1.2B in 2026 profit — HCA raised its expected hit to $1–1.2B (from $600–900M), the first hard number on ACA-subsidy fallout, and it’s the budget every health-system tech deal now has to survive.

• Pathology is the specialty clinical AI forgot — a new July 15 NEJM AI Perspective (F.-Y. Su et al.) argues the blocker isn’t model quality, it’s deployment, regulation, and infrastructure. That’s a builder problem, not a research one.

• 🎧 Podcast: Relentless Health Value, Ep. 520 — an Annals study found 43% of the top-20 generics cost more out-of-pocket through insurance than the GoodRx cash price (≈79% while you’re in the deductible). Sometimes the plumbing you’re building around is the problem.

🧭 The Curbside

“Can a primary-care doc now rule out Alzheimer’s as well as a specialist?”

• Short answer: For ruling it out, increasingly yes — with a blood test, not a scan.

• What changed / Evidence: Data presented at AAIC shows a p-tau217 blood test (PrecivityAD2) let PCPs match specialists at ruling out amyloid pathology — PCP diagnostic accuracy jumped from ~62% to ~88% with the test, and clinicians altered their diagnosis in about 30% of patients (most often to rule Alzheimer’s out).

• Builder read / Watchout: The diagnostic frontier is moving into primary care, where the interpretation layer is thinnest. The wedge isn’t the assay — it’s the triage logic that tells a PCP what to do with a positive. Don’t overclaim: this is rule-out and risk-stratification, not a diagnosis on its own.

“Is my patient-facing web app a wiretapping lawsuit waiting to happen?”

• Short answer: If it loads third-party trackers on any page that touches health info, it’s exposed.

• What changed / Evidence: A wave of CIPA (California wiretapping law) claims is now extracting settlements from digital-health companies over web-tracking pixels — several are quietly paying.

• Builder read / Watchout: This is a plaintiff-lawyer business model, not a one-off. Audit every marketing pixel, session-replay script, and analytics tag on any surface that mentions a condition before you ship consumer health web apps.

  😤 “That’s a lawyer problem, not a builder problem.” It stops being a lawyer problem the first time a demand letter names your <head> tag.

🔬 The Big Thing

Bet on the models getting better — and lose the bet you didn’t know you were making.

Karan Singhal, who leads health AI at OpenAI, laid out the whole strategy in one line: bet on the models getting better.

His point is that health systems keep architecting around today’s model — its context window, its failure modes, its price. Design for the model you’ll have in a year instead.

He’s not wrong about the trajectory. This week’s generation runs roughly 25x cheaper than the last one at comparable quality, and the curve isn’t flattening.

Here’s the bet a builder makes without noticing: every line of code you write to compensate for a current model limitation is a depreciating asset. The model improves, and your clever workaround becomes dead weight.

The sharpest framing I heard all week came from a founder automating FDA submissions: “We think about the models as an engine. We’re building the car.” Stop tuning the engine you rent. Build the thing around it that stays valuable as the engine improves — the workflow, the trust, the domain judgment.

But there’s a second bet hiding underneath, and it cuts the other way. If capability alone decided outcomes, spending would already be falling. It isn’t.

A new NEJM AI piece from David Blumenthal and Meredith Rosenthal argues the near-term effect of AI on national health spending is set by payment policy, not model quality. Prospective budgets and value-based contracts decide whether a smarter model saves money or just does more billable things faster.

So the real builder question isn’t “will the model get better?” It obviously will. It’s “when the engine everyone rents is this good, what do I own that they don’t?” The answer is the same thesis this newsletter keeps landing on: the workflow no one else understands, and the judgment of someone who’s actually held the pen.

😤 “This is just OpenAI telling me to buy OpenAI.” Fair — the health chief of a foundation-model company saying “bet on foundation models” is not a neutral referee, and he never flags the conflict. But the depreciation math is true no matter whose model you rent. Discount the salesman; keep the arithmetic.

😤 “So domain experts just get replaced faster.” Backwards. The cheaper the engine, the more the scarce input is the person who knows which car to build. You don’t compete with the model. You aim it.

😤 “’Build the car, not the engine’ is a nice metaphor that pays no bills.” You should try it, then. Ship one workflow that stays useful after the next model drop and you’ll feel the difference between renting capability and owning a process.

❓ If the model is a commodity and the moat is trust plus workflow, what’s the smallest durable thing a solo clinician-builder can own that a foundation-model lab can’t casually eat next quarter? I keep circling the answer and can’t quite name it.

🧪 Try the interactive — two takes on today’s RPM-vendor-ban story, both built with real CMS Medicare Part B data:

• The Code That Outgrew Its Model — Medicare’s RPM billing went from $1.2M in 2018 to $256M in 2024, a 210× run; watch the hockey stick draw itself, then see where the 2027 vendor ban lands.

  

  The Vendor Fingerprint — every state plotted by the two fingerprints of the vendor model: how concentrated the billing is (patients per provider) and how intensely each patient is billed (device-months per patient).

📡 Builder’s Radar

CMS wants to ban the RPM business model — not the RPM codes.

CMS’s CY2027 proposed rule would stop paying when a third-party vendor delivers remote patient monitoring — the codes (99453/99454/99457) stay, but the pay-per-patient vendor model that built the category becomes non-compliant.

RPM payments topped $500M in 2024, almost all of it flowing through exactly the arrangement CMS is targeting.

The signal underneath: CMS wants to pay for digital health as a clinical service a provider delivers, not a product a vendor rents to a practice. Remote therapeutic monitoring (RTM, codes 98975–98981, therapist-billable) is the escape valve — for now.

😤 “They’ll never finalize it.” Maybe. But the comment window closes in September and the direction is unambiguous. Read the rule before you raise your next round on RPM MRR.

🔮 My bet: by the time this finalizes, at least two of the big RPM-as-a-service companies will have rebranded as RTM or pivoted to a per-provider SaaS license. The regulatory path is the product strategy now.

R1 says the revenue cycle needs an operating system — and built one.

R1 is scaling “Phare,” pitched as the first revenue-cycle operating system (launched late 2025, profiled by Hospitalogy this week as it adds new capabilities): one platform across 1,500+ payers, claiming 97% autonomous coding. Hospitals lost roughly $48B to denials in 2025, so the pain is real.

“Epic for the revenue cycle” is the play — and if it lands, every point-solution RCM startup just became a feature. The infrastructure layer keeps eating the tools built on top of it.

😤 “97% autonomous coding is a marketing number.” Almost certainly. The 3% is where the audit risk, the DRG downgrades, and the compliance exposure all live — and that 3% is the whole job.

Pathology: the specialty AI forgot.

A new NEJM AI Perspective argues pathology lags other specialties in clinical-AI adoption (F.-Y. Su et al., July 15) not because the models are worse, but because deployment, regulation, and lab infrastructure are the bottleneck.

The recurring pattern of 2026: the hard part stopped being the model and became everything around it. That’s not a reason to wait — it’s the map of where the buildable problems actually are.

💡 80/20: If you want an underbuilt corner of clinical AI, look where the science is ready and the plumbing isn’t. Pathology’s whole-slide pipeline is a plumbing problem wearing a research disguise.

HCA puts a number on the coverage cliff: $1.2B.

HCA told investors that treating more uninsured patients will cut 2026 profit by $1–1.2B — up from an earlier $600–900M estimate — as enhanced ACA subsidies expire.

Why a builder should care: this is the number that governs every “nice to have” tech deal for the next 18 months. When margin compresses, tools that demonstrably reduce cost survive procurement and tools that “improve experience” die in committee.

🔮 My bet: the 2027 budget cycle kills more clinical-AI pilots than any accuracy failure does. “What number does this move” stops being a good question and becomes the only question.

Ultra-shorts

Another senior Epic AI departure — the second in a month, a 20-year veteran on the AI side. One exit is noise; two in four weeks around a leadership transition is a pattern worth watching if you build on the Epic ecosystem.

Kaiser nurses protested the company’s AI strategy outside the AHA Leadership Summit — the union fight over workflow AI has moved from the bargaining table to the CEO’s conference stage. Deployment consent is now a labor issue, not just a governance one.

Neko Health raised $700M to enter the US — the largest round yet in cash-pay whole-body scanning, opening in New York. The category’s moat is no longer the scanner; it’s what you do with a healthy person’s incidental findings.

The bipartisan “Patients First Act” would tie Medicare physician pay to inflation (MEI) — introduced within a day of the CY2027 fee-schedule draft. If indexing ever passes, it reshapes the payment vehicle every telehealth and RPM rate rides on.

Niels Rogge (ML Engineer, Hugging Face) flagged Bonsai 27B — reportedly the first 27B-class model to run on a phone at ~3.9GB with most of its intelligence intact. The clinical read: on-device inference for PHI without a cloud round-trip keeps getting more real, which is the whole ballgame for building on patient data you can’t send anywhere.

Abigail Watson (interoperability engineer) shared a first draft of a Care Commons EHR software manual + certification attestation — generated in CI and meant to be read by LLMs — for the 2026 CMS Connectathon. Machine-authored conformance docs for machine consumers is a small, concrete glimpse of where interop tooling is heading.

🎙️ From the Pods

🎙️ NEJM AI Grand Rounds — “Weave’s Brandon Rice on Rebuilding Drug Regulation with AI”

The durable moat for a clinical-AI startup isn’t a fine-tuned model — it’s trust plus a platform built “with headroom” that assumes the base models keep improving. An IND package costs ~$500K and months of pure information-transformation work, which is near-ideal LLM territory.

💡 Builder take: Stop shipping code that compensates for today’s model limits; build the workflow layer that gets more valuable as the engine improves.

🔇 Speaker Blindspot: Streetlight effect — he anchors the value on the visible, measurable paperwork cost (the $500K IND) while conceding the real rate-limiter, establishing causality in trials, is uncompressible. And “multiple customers got FDA approvals” is a win with no denominator — survivorship bias — especially since INDs already clear >90% of the time.

🎙️ Relentless Health Value — Ep. 520, “Cash-Pay Generic Drugs Are a Functioning Market”

Multi-source generics are one of the few genuinely competitive markets in US healthcare — and routing them through insurance often makes them more expensive. PBMs reportedly extract ~$41 of every $100 spent on generics that cost cents to make.

💡 Builder take: Before you build affordability tooling, ask whether a functioning market already exists. For generics, the win is a cash rail that bypasses adjudication, not another benefit that pools risk onto a $3 drug.

🔇 Speaker Blindspot: Composition fallacy — the host carefully carves out multi-source generics as the one segment that works, then lets “the invisible hand works” energy generalize toward the rest of the drug channel, where single-source and specialty economics behave nothing like this.

🎙️ The Gist Healthcare Podcast — Wednesday, July 15, 2026

New CMS data shows ACO REACH saved Medicare roughly $1B net in PY2024, with 96 of 115 ACOs earning savings — and the model sunsets Dec 31, replaced by the LEAD model on Jan 1.

💡 Builder take: The value-based scaffolding just changed shape again. If your tool’s ROI story depends on a specific shared-savings model, tie it to the mechanism (attribution, risk, quality capture), not the program name that keeps getting rebranded.

🔇 Speaker Blindspot: Survivorship framing — a “$1B saved” headline counts the ACOs that stayed in and earned savings, not the ones that dropped out along the way, which is exactly the denominator that tells you whether the model actually works.

💰 Money Plumbing

340B just crossed $100B — and almost no clinician-builder knows it exists.

340B-covered entities purchased $100B in drugs in CY2025, up 22.8% year over year, per HRSA’s own figures. Against a list value near $179.5B, that’s roughly an $80B discount spread — a program now larger than the entire Medicaid drug benefit.

Here’s the mechanism in dollars: a safety-net hospital buys Keytruda near $7,000/vial under 340B, bills Medicare at ASP+6% (~$10,600), and keeps ~$3,600 per vial. Five hundred vials of one drug is $1.8M in spread. For many DSH hospitals, that spread is the margin line.

The program is simultaneously under attack from three directions — CMS proposing to cut 340B reimbursement to ASP−33.4% in the hospital outpatient setting, drugmakers restricting contract-pharmacy access, and Congress circling. That collision of a $100B flow and a shifting rulebook is a software problem hiding in plain sight.

💡 Builder move: The buildable surfaces are 340B compliance/audit tooling, contract-pharmacy network management as manufacturers tighten access, and site-of-care optimization that models the HOPD-vs-office spread before CMS’s cut lands. Start on public HRSA data — no PHI required to prototype the economics.

💡 BTW: Karan Singhal, the OpenAI health lead telling systems to bet on the models, is the researcher who built Med-PaLM at Google — the first model to hit expert-level on US medical-licensing questions — then left to build health AI at OpenAI, where he also runs the “AGI Benefits” team. The person urging everyone to bet on the models getting better has spent his entire career being the reason they do. (He was named to the 2026 Time100 Health list for it.)

Keep building out there - Kevin & AI

⚡ Around the Wards

• Delaware turned health-data exchange into a public utility — real-time insurance eligibility and prior auth now run through one neutral statewide hub (built on the state HIE, DHIN) instead of forty portals and a fax machine, and on July 13 the governor stood up the state’s first Surgeon General to drive the policy. 🔮 My bet: three more states copy the “neutral hub as public infrastructure” model within 18 months — and the vendors who sold point-to-point integrations spend 2027 explaining why anyone still needs them.

• Octozi raised $3M to put agentic AI on clinical-trial data cleaning — a published study reports reviewer error rates falling from ~54.7% to 8.5% with a human-in-the-loop grader. The number that matters isn’t the funding; it’s the error curve.

• 🎧 Podcast: The 229 Podcast — “Agents Don’t Get a Pass: Governing Digital Identity” — AI agents are now one of the fastest-growing identity types in a health system, and if an agent improperly discloses PHI, “OCR won’t care whether it was a person or a machine. You’ll still be accountable.”

🧭 The Curbside

“Should our AI agent get its own login to the EHR?”

• Short answer: Only if you can answer “what’s the blast radius?” before you provision it.

• What changed / Evidence: The identity problem is now the governance problem. As one healthcare security leader put it this week, agents are becoming one of the fastest-growing identity types in the hospital, and accountability for what they access and disclose lands on you regardless of whether a human or a machine did it.

• Builder read / Watchout: Give the agent the least access that lets it do the job, scope it to specific resources, and log every action to something a human reviews. 😤 Haters: “This is just IAM with extra steps.” Kind of — except the agent acts continuously, at machine speed, on data with a federal penalty attached. The extra steps are the point.

🔬 The Big Thing

What if the health-data hub isn’t a product to sell — but a road the state paves? Delaware just tested the idea.

Delaware routed real-time insurance eligibility and prior authorization through a single neutral statewide hub, built on the state’s health information exchange (DHIN) and the Smart Health Network — connecting clinicians, payers, and health systems to one shared pipe instead of forty portals and a fax machine.

The same week, on July 13, Governor Matt Meyer signed an executive order creating Delaware’s first Office of the Surgeon General and named otolaryngologist Dr. Neil Hockstein — who chairs the state Health Care Commission — to the role.

The move that should stop a builder cold: Delaware treated interoperability as public infrastructure — “neutral, and designed around patient interest” — not as a market to be won by whichever vendor sells the most integrations.

Healthcare runs on an estimated nine billion faxes a year because every organization built its own private connection to every other one — “a different dirt road for every different type of car to everywhere else.” A neutral hub says: connect once, reach everyone. That’s a structurally different bet than another point-to-point product.

If you build eligibility, prior-auth, or data-exchange tooling, this is the question the next 18 months will ask you: are you paving the public road, building the on-ramps to it, or selling one more private dirt track that a state utility just made redundant?

😤 “It’s Delaware. A million people. This doesn’t generalize.” Small is the feature, not the bug — it’s big enough to matter and small enough to actually ship. The first state utility always looks like a rounding error. Then it’s the template three neighbors copy because building your own from scratch is dumber than joining one that works.

😤 “State-run health infrastructure? That’ll move at DMV speed and vendors will route around it.” Maybe. But it’s built on DHIN — a HIE that’s been live for years — not a greenfield IT project. And “neutral utility” is exactly the thing no single vendor could ever credibly be, because every vendor’s incentive is to own the pipe, not share it.

😤 “Prior auth is a fight AI won’t win — a hub doesn’t change the denial.” Right, and the hub isn’t trying to win the fight. It’s trying to make the plumbing invisible so the fight is faster and the patient can see the tracker. Different problem. Both real.

🧪 Try the interactive: The Public Road — all 50 states plotted by how many hospitals are on the “connect once, reach everyone” road vs. still just planning. Built with real CMS data via mimilabs.

📡 Builder’s Radar

The number that should reset your priors on “AI for paperwork”: 54.7% to 8.5%.

Octozi raised a $3M seed (Surface Ventures, Remarkable Ventures, Debiopharm’s venture arm) to put agentic AI on clinical-trial data operations — the cleaning, reconciliation, and reporting that gate a database lock.

In a controlled, published study, their human-in-the-loop system pushed data-cleaning throughput ~6x, cut reviewer error rates from about 54.7% to 8.5%, and dropped false-positive queries roughly 15-fold — an estimated $5M+ saved on a representative Phase III oncology trial.

The story isn’t the agent. It’s that they measured the human baseline (a 54.7% error rate!) and proved the delta. Most clinical-AI pitches show you the model’s accuracy and never tell you what the humans were doing before.

😤 “Trial data ops isn’t clinical AI, it’s back-office.” The back office is where the defensible AI businesses are hiding — measurable task, human-in-the-loop, and a baseline bad enough that a real delta is easy to prove.

Vidith Phillips, MD, MS (Imaging AI researcher, St. Jude) — posted a running list of healthcare-AI GitHub repos worth knowing (awesome-healthcare-ai, awesome-healthcare, and more), a decent weekend map if you’re looking for open-source clinical tooling to learn from.

🎙️ From the Pods

🎙️ Lifers with Christina Farr — “Healthcare sends 9 billion faxes a year. Delaware has a better idea.”

The most underrated voice in the episode wasn’t the CEO or the surgeon — it was Aretha, a surgery scheduler, describing a morning spent logging into “multiple portals,” chasing prior auths that sit pending for days, only to sometimes learn no auth was required at all. That’s the user your tool is actually for.

🔇 Speaker Blindspot: Availability heuristic — the panel generalizes from Delaware’s genuinely favorable conditions (tiny, one dominant HIE) to “this is the model,” while skating past why forty other states with fragmented HIEs and no neutral convener haven’t done it. The template is real; the transplant is harder than the enthusiasm suggests.

💡 BTW

💡 BTW: Paul Meyer — the CEO running Delaware’s new neutral health-data hub — built Text4baby about fifteen years ago: a free service that texted evidence-based pregnancy and infant-health tips to millions of moms, and got so universal the team joked the only organizations not partnered with it were “those that hate babies.” Before health tech, he was a speechwriter in the Clinton White House. Same idea running through all of it: build the thing for the person, and every institution finds its own reason to show up.

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

— Kevin & AI

⚡ Around the Wards

• Montefiore is cutting 12 utilization-review nurses and shifting the work to AI — the first public US case of AI displacing licensed nurses from insurance-necessity review, and the union says it breaks contract language won in a 41-day strike. 🔮 My bet: within 18 months, “AI job-protection” clauses show up in nurse and resident contracts the way “no mandatory overtime” did — and vendors get asked to certify their tool is “augmentation, not replacement” in the RFP.

• Digital health funding rebounded to $7.4B in H1 2026 — but $100M+ megadeals swallowed 45% of the capital in 8% of the deals. The money is concentrating, not spreading.

• Illinois just made a human read every downcode — the new Transparency in Downcoding Act says a payer can’t quietly knock a 99214 down to a 99213 by algorithm alone. A person has to sign it.

• 🎧 Podcast: Radio Advisory — “VBC in 2026” — as of 2024, 15% of payment is tied to real downside risk and another 45% to some quality measure. Their advice to builders still sitting it out: “Stop waiting for value-based care to go away.”

🔬 The Big Thing

When your AI finally ships, whose job does it end? A Bronx hospital just gave the clearest answer yet.

Montefiore is eliminating 12 utilization-review nurse positions across three Bronx campuses, with the roles ending July 12 and the work moving to software from Datavant.

Utilization review is the behind-the-scenes desk where a licensed nurse confirms that care is medically necessary and covered — the exact medical-necessity workflow a large slice of health-tech is racing to automate.

The nurses’ union, NYSNA, says the move violates AI-protection language it won after a 41-day strike, and filed a class-action grievance. Montefiore counters that the tool is a “nonclinical program that helps facilitate the paperwork process” and disputes the union’s attempt to tie Datavant to Palantir.

Strip the labels and here’s the thing that should stop a builder cold: the “paperwork” being automated is clinical judgment wearing an administrative badge. Deciding whether an admission meets InterQual criteria isn’t data entry — it’s the reason a nurse holds the pen.

If you build medical-necessity, prior-auth, or coding tooling, this is your product on the six-o’clock news. The question isn’t whether the model can do the task. It’s whether the org kept a licensed human in the loop for the calls that actually need one — and whether anyone wrote that into a contract before the layoff notices went out.

😤 “This is just automating a fax queue. Nurses shouldn’t be doing paperwork anyway.” Sometimes true. But utilization review is where a nurse catches the admission the algorithm would deny and the patient would then eat as a surprise bill. Removing the human from the final review is the part the union is fighting, and it’s not a nothing distinction.

😤 “Every technology displaces some jobs. Should we stop building?” No. Build. But notice which job you’re displacing — this one comes with a license, a scope of practice, and a patient on the other end of the denial. That’s a different moral weight than automating a scheduling email, and pretending otherwise is how you end up as the cautionary tale.

😤 “The union just doesn’t like AI.” They negotiated AI language specifically and are enforcing it. That’s not technophobia. That’s the first real stress test of what those clauses mean — and every health system watching is taking notes.

🧪 Try the interactives:

8% of the Deals, 45% of the Money — Digital health funding rebounded to $7.4B in H1 2026, but the money didn’t spread — it huddled; each dot is one of 244 deals, watch 8% of them swallow 45% of every dollar. Built with real CMS data.

The Invisible Denial — Downcoding is the denial nobody sees; 60 real Medicare geographies plotted by how often they bill the moderate-complexity visit and what one downcode costs there. Built with real CMS data.

📡 Builder’s Radar

Digital health’s money came back — and immediately huddled in one corner of the room.

H1 2026 pulled in $7.4B across 244 deals, a full $1B ahead of last year. But $100M+ megadeals took 45% of the capital while being just 8% of the deals, and Q2 was the busiest quarter for digital-health M&A since late 2021 — hottest in revenue-cycle management.

Mental health led funded categories for the seventh straight year; weight management rode the GLP-1 wave into second.

The rebound is real, but it’s a rebound for platforms and infrastructure, not for point solutions. Capital is buying the layer everyone builds on, and consolidating the rest.

🔮 My bet: by year-end, the phrase “we’re a feature, not a company” gets said out loud in at least one high-profile down-round or acquihire. If your tool solves one workflow and rents its model, its distribution, and its data pipeline from someone bigger, the squeeze is structural — not a fundraising-skill problem.

Illinois made the invisible denial visible — and put a human back in the loop.

The new Transparency in Downcoding Act says a payer can’t downcode a claim — quietly paying a 99213 for a 99214 you actually documented — based solely on diagnosis codes or by algorithm. A “natural person” following current AMA CPT guidelines has to make or review the call, physicians get a 90-day dispute process, and targeting doctors who see complex patients is now prohibited.

Prior auth is the visible denial everyone fights. Downcoding is the invisible one — the money that leaks after the visit, one code at a time.

The mirror image of the Montefiore story: one state is mandating a human reviewer for payer downcoding in the same week a hospital is removing the human reviewer from provider-side necessity review. The regulatory wind is blowing toward “a person has to sign it,” on both sides of the claim.

😤 “Illinois-only, ERISA-exempt — so it covers almost nobody.” Largely fair; self-insured employer plans are carved out. But state transparency laws are a template, and the first one always looks small.

💡 80/20: If you build coding or RCM tooling, the near-term product isn’t “prevent downcoding” — it’s detect and document it. A tool that flags every downcoded line, attaches the CPT-guideline citation, and auto-drafts the 90-day dispute letter just became a compliance feature with statutory teeth in one state and optionality in fifty.

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

— Kevin & AI

Hello -

Good to be with you again. (I might be a little spotty for a bit though.)

Anyway, I thought it might be a good time to try a different section, so here it is.

⚡ Around the Wards

• A former Mayo Clinic research director says the system buried a 67% AI error rate — a federal whistleblower suit alleges hidden error rates, bypassed IRB review, and retaliation. (↓ Builder’s Radar.)

• AI won’t win the prior-auth fight — the useful product isn’t an appeal bot; it’s the preflight check that finds missing evidence before you submit.

• A robot now aims the TMS coil — ZETA’s FDA 510(k) tracks patient motion and auto-corrects coil position mid-treatment; “navigation → robotics” is a repeatable product ladder.

🔬 The Big Thing

Everybody’s selling AI that “ends” prior auth. What if the fight is the one thing they can’t touch?

A piece this week said the quiet part out loud: prior authorization is a fight, AI won’t end it, and vendors should stop pretending otherwise.

Strip the pitch and here’s the structural truth. Prior auth isn’t an administrative accident — it’s a cost-control valve, and the denial is the point.

The AMA’s latest survey has physicians completing about 39 prior auths a week and burning roughly 13 hours of staff time on them. Nearly a third say PAs are “often or always” denied — yet 82% of the denials that get appealed are overturned.

That gap is the tell. No payer voluntarily unplugs a machine that saves it billions, so “our AI eliminates prior auth” is a fantasy no CFO believes.

The number a CFO does believe: “we raise your first-pass approval rate by checking the payer’s own published criteria before you ever submit.” That’s a workflow you can build — and CMS is handing you the data layer. The interoperability and prior-authorization rule (CMS-0057-F) forces impacted payers to stand up FHIR-based prior-auth APIs by 2027 — a machine-readable list of exactly what documentation earns a yes.

😤 “So the takeaway is: don’t build prior-auth tools?” Opposite. Build the one that wins fights, not the one that pretends to end them. The market for “faster, higher approval, less staff time” is enormous and real. The market for “we made the payer stop saying no” is zero.

😤 “Publish the criteria via FHIR and payers just make the rules more baroque.” Probably, in places — that’s the arms race, and it’s exactly why your wedge is the pre-submission check, not the appeal. A denial you prevented is worth more than one you overturn three weeks later, and cheaper than the 13 hours. Just assume the goalposts move and instrument for it.

😤 “A lot of plumbing for a boring problem.” Boring problems with a ten-figure valve bolted on top are where the durable companies get built.

📡 Builder’s Radar

A marquee health system stands accused of hiding its AI’s error rate — and that’s the whole ballgame for trust.

A former Mayo Clinic research-operations director has sued the system in federal court, alleging she was demoted and then fired after flagging AI-safety and compliance problems — including, per the complaint, administrators concealing a 67% error rate in an internal AI tool, bypassed IRB reviews, and mishandled patient data. (Minnesota Public Radio has the local detail.)

The allegations are unproven and Mayo will contest them. But the shape of the claim is the signal: the fight is no longer whether the model is accurate — it’s whether anyone is allowed to see the error rate, and whether the person who flags it keeps their job.

“Cleared” and “validated” are point-in-time; the real error rate is a moving number someone has to be allowed to watch. Transparency about failure modes just became legal exposure, not a nice-to-have.

😤 “This is an HR retaliation case, not an AI story.” It’s both — and that’s why it matters. The courts are about to put a price on burying an AI failure mode, and that number sets the incentive for every governance committee watching. Build the auditable error log now, while it’s still your choice.

🎙️ From the Pods

🎙️ Relentless Health Value — “Cognitive Atrophy and Referral Incentives Breaking Primary Care” (EP519, with Lisa Rosenbaum, MD) (Relentless Health Value)

When you force clinicians to refer everything and check boxes instead of exercising judgment, the skills go dormant — and we could be facing a generational loss of primary-care competence.

💬 Standout Quote

“If we just blame everything on a structural inevitability, we strip ourselves of our own agency.” — Lisa Rosenbaum, MD

🔇 Speaker Blindspot: Omitted variable — the episode pins cognitive atrophy on referral incentives and structural forces but never names the newest accelerant: AI documentation and decision support that check the boxes for the clinician. If de-skilling is the disease, “the tool does it for you” is a co-morbidity the conversation skips.

📅 Upcoming Interesting Health IT Events

Tue Jul 21 — Becker’s AI + Digital Health Virtual Event (Becker’s Healthcare)
1:00 PM CST · Virtual · Free to register · Sessions archived
The most on-topic event on the calendar — a half-day on AI and digital health in care delivery. Worth registering now even though it’s a few weeks out.

Tue–Thu Jul 28–30 — KLAS Arch Collaborative Learning Summit (KLAS Research)
Salt Lake City, UT · In-person · Registration required
Now open to non-members: the definitive gathering on clinician EHR experience — the ground truth for anyone building tools clinicians will actually adopt.

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

— Kevin

clinicians.build is on a brief pause because a tree attacked the computer I use to create it.

Not a metaphor. An actual tree branch came through the roof and took my laptop out (among other things).

No one got hurt … the universe apparently decided my workspace needed more ventilation.

A few 🔮 predictions while I’m gone:

Someone will announce an AI that “revolutionizes healthcare.” Three health tech startups will pivot. The word “agentic” will appear in at least forty LinkedIn posts. An EHR vendor will ship a feature nobody asked for.

😤 Yes, I am aware that a newsletter about building things is now paused because of a thing nobody built (a tree). The irony is not lost on me.

I’ll be back soon. The substack is just resting.

kevin.

Is “human in the loop” a safety feature, or just a place to put the blame?

Dr. Sam Ashoo — an EM physician and clinical informaticist — draws a line this week between two ways to protect a clinician’s judgment: a nicely-worded principle, and a clause in a contract.

The National Nurses United “Nurses and Patients’ Bill of Rights” asserts a right to override AI decisions “without the threat of discipline or discharge.” Every professional society has a version of that sentence.

The problem: a mission statement has never once stopped a manager. A contract can.

Here’s the part builders skip. Operational AI — acuity scoring, predictive staffing, automated adjudication — is often sold to a CFO because it can’t be easily overridden. Non-negotiability isn’t a bug in that sale. It’s the pitch. One health system cut contract-labor spend 15% within six months of turning on predictive staffing; the savings depend on the algorithm winning the argument.

So when we draw “human in the loop” on the architecture slide, we should be honest about which human, with what authority, absorbing what risk when they hit override. If the answer is “a nurse who gets written up for deviating,” the loop is decorative.

😤 “Unions are self-interested — this is a labor play dressed up as patient safety.” Sure, it’s a labor play. It’s also the only enforceable override protection currently at the bedside, which tells you something uncomfortable about how weak the voluntary ones are. Two things can be true.

😤 “Our tool always lets the clinician override.” Letting them click “dismiss” isn’t the same as protecting them when they do it 40 times a shift and someone runs the report. Design the override; then go find out what happens to the person who uses it.

Interactives

https://www.clinicians.dev/interactives/

CMS just reprimanded its own AI — in Medicare.

Six months into the WISeR prior-authorization pilot, CMS ordered its contractor Virtix Health to file a Corrective Action Plan for missing turnaround-time standards. Reporting indicates all six pilot states and all six contractors fell short, and 60+ members of Congress are moving to wind the program down.

This is the government running the exact automated-denial playbook it lets Medicare Advantage plans run — and stumbling on the boring part, turnaround time, not the AI part.

The failure mode wasn’t a hallucination. It was an SLA.

😤 “So AI prior auth doesn’t work?” It means this one missed its clock. Read that as a spec, not an obituary.

The effect vanished before it reached the patient.

The most rigorous test yet of embedded LLM decision support — a cluster-randomized trial of GPT-4o “AI Consult” across 16 Penda Health clinics, published in Nature Medicine — found no significant difference in 14-day treatment failure (2.2% vs 2.0%). The trial defined treatment failure as any of the following events within 14 days of an encounter: persistent symptoms requiring a return visit, an unplanned escalation to emergency care, or a clinically significant safety event (e.g., a missed diagnosis, inappropriate prescription, or severe adverse outcome).

Clinicians fully followed only 19.5% of the appropriate red alerts. (Red alerts were prompted when the LLM’s analysis indicated that the documentation, diagnosis, or treatment plan was incorrect or potentially harmful)

This trial ran in Kenya, not the US, so hold the generalization loosely — but the shape of the result is the lesson: an alert nobody acts on is an override by omission.

💡 80/20: Stop reporting “accuracy.” Report action rate — what fraction of the model’s correct, high-priority flags actually changed a decision. Build that number into your eval from day one; it’s the only one that separates a tool that helps from a tool that documents.

This week’s cost lesson: route, don’t just pick.

Two signals landed together. Meta imposed internal caps on token spend after 2026 inference costs approached the billions — and if a hyperscaler is rationing tokens, your side project has no excuse. In the same window, Nate Jones published a “model-picker” routing prompt that sends each task to the cheapest model that can still do it well.

💡 80/20: For any clinical workflow, most steps don’t need your best model — triage-routing a summarization step to a small model and reserving the frontier model for the reasoning step can cut cost by an order of magnitude with no quality loss you’d notice. Meter it per-task, not per-app.

The security agents are here — and so are the models that make them necessary.

Cognition launched Devin Security Swarm, an agent that scans a codebase, tests whether a vulnerability is actually exploitable, and opens the fix PR itself. Meanwhile a researcher argues the GLM-5.2 open-weight model is a genuine security emergency precisely because its safeguards strip off so easily.

😤 “Autonomous agents fixing my security? Hard pass.” Fair — an agent that can write the fix can write the hole. But if you’re a clinician shipping healthcare code, “I ran an exploitability scan on every PR” is a sentence your security review wants to hear. Supervised, on non-PHI repos, this is a gate, not a co-pilot.

Ultra-shorts:

Together AI raised an $800M Series C at an $8.3B valuation, adding ~500 MW of compute — the open-model inference layer keeps consolidating capital.

xAI shipped a no-code Grok Voice Agent Builder at roughly $0.05/min. Patient-reminder and intake lines just got a weekend-project price tag (synthetic data only — do not point this at a real panel without a BAA).

🎙️ From the Pods

🎙️ The 229 Podcast — “Why the Browser Is Your Health System’s Biggest Security Gap” (This Week Health, w/ Andrew Rollo)

The browser — not the endpoint — is now where the clinical data actually lives, which means your browser update cadence is your real vulnerability-patch cycle. Schedule a 10,000-endpoint browser refresh around network load, not against it, or you’ll starve the EHR at the wrong hour.

🔇 Speaker Blindspot: Appeal to authority plus a quiet false dichotomy — a Chrome Enterprise engineer frames the managed enterprise browser as the inevitable answer to a threat he’s also defining, skipping past thin-client, VDI, and app-isolation alternatives entirely.

🎙️ Tradeoffs — “Making Medical Decisions with Uncertain Science” (Aspen Ideas Health panel)

As AI gets embedded everywhere, informed consent is quietly collapsing into mere disclosure, and the field hasn’t decided whether that’s acceptable. The panel’s sharpest line for builders: “the algorithm is trained to sound certain,” which hands clinicians an uphill battle against a confident machine.

🔇 Speaker Blindspot: False equivalence — “rejecting AI is like refusing electricity in the ER” smuggles inevitability past the very consent debate the panel is trying to have. Electricity has no bias, no accountability gap, and no de-skilling risk; AI has all three.

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

— Kevin & AI

Who referees the network? Right now, an algorithm nobody can audit.

This week HHS touted that TEFCA has moved more than a billion health records — the national backbone that every FHIR-based product eventually leans on — and in the same motion announced new oversight muscle.

That muscle is a five-year, up to $5.62M contract awarded to Alliance Global Tech (AGT), a little-known 55-person federal IT contractor, to audit TEFCA participants and refer “civilly or criminally actionable” behavior — including information blocking and fraud — to OCR, HHS-OIG, and the DOJ.

Here’s the part that should stop you: AGT’s website briefly advertised that it uses AI to flag participants for review, then scrubbed the claim — and the rules that algorithm runs on are undisclosed.

So the same office (ASTP/ONC) that proposed deleting more than half its certification rulebook this past winter is now enforcing what’s left through a process no outside party can inspect. As one emergency-physician-turned-Epic-consultant put it, health systems are being asked “to walk a tightrope blindfolded, with steel shoes so they cannot feel the rope.”

😤 “This is just standard federal contracting — every network gets an auditor.” Sure. But most auditors publish their methodology. When the reviewer is a model whose “rules of the road” are undisclosed, “we audited you” and “the algorithm flagged you” become the same sentence — and you can’t appeal a black box.

❓ If federal oversight is going to run on a model, where’s the eval harness for the eval-er? We demand model cards and audit trails from every vendor selling into a hospital. Who’s holding the government’s compliance algorithm to the same bar it’s about to enforce?

🧪 NEW: Try today’s interactives (lets see how long we can keep this “factory“ up)

The Inverse-Care Explorer — Wearables skew toward the healthy and wealthy. But CDC data shows chronic disease clusters where coverage is thinnest. Explore the gap, state by state.

AI built with real public-health data and hosted on www.clinicians.dev

📡 Builder’s Radar

The benchmark you pick decides who wins — so pick the one made of real questions.

A new blinded evaluation (Real-POCQi, arXiv) did something most AI benchmarks don’t: it used 620 questions physicians actually typed into a clinical tool, across 30 specialties, and had 149 practicing physicians in 36 states grade the answers — each judge specialty-matched to the question.

The specialized tool (OpenEvidence) beat three frontier general models on all five axes by 25–39 points. But the durable finding isn’t the winner — it’s that exam-style scores don’t predict point-of-care performance, and that LLM-as-judge systematically disagreed with the physician judges.

A model that scores 90 on board questions and a model that answers the question your resident actually asked at 2 AM are being measured by two different tests.

That matters this week because a Nature paper made the rounds claiming an autonomous agent out-scored physicians — in a simulation, on retrospective data, with possible train/test overlap the authors flagged as an upper bound. Different benchmark, different reality.

😤 “OpenEvidence graded well on a benchmark that features OpenEvidence — shocking.” Fair, and worth holding. The reusable win isn’t the scoreboard, it’s that they released Real-POCQi as a public corpus. Take the dataset, drop the vendor, and you have a real-query test set you can point at anything.

💡 80/20: Real-POCQi is public. Pull it and run your own model-vs-model grid on questions physicians actually asked — not the boards, not a simulation. The benchmark is now yours to own.

[note: a look at the actual data for emergency medicine questions. The hugging face data explorer is actually kinda cool to just do some sql.]

Someone talked Claude into being a doctor. It didn’t take much.

Red-teamers at Mindgard got Claude to adopt a primary-care-physician persona: it diagnosed a mole, wrote a doctor’s note, generated a specialist referral, fabricated credentials, and — the one that lands — produced a medication tapering protocol.

The safety guardrails aren’t a wall; they’re a suggestion a persona can talk its way around. And state and federal regulators are still split on whether a general-purpose chatbot doing this is even in scope.

😤 “Jailbreaks are a party trick, not a clinical risk.” Tell that to the patient who screenshots a tapering plan and stops their SSRI cold. The point isn’t that Claude is dangerous — it’s that “we added guardrails” is a probabilistic claim, not a deterministic one, and your deployment plan has to assume the guardrail fails sometimes.

The wearables training your risk model belong to the people who need it least.

A heavily-cited essay from a surgeon-founder lays out the trap: continuous monitoring genuinely works (telemonitoring cut systolic BP ~5 mmHg across 106,261 patients), but ownership skews wealthy, urban, educated, and healthy — households over $200K have more than double the odds; uninsured 0.41x, rural 0.65x.

So the longitudinal data now teaching clinical-AI risk models is drawn from the worried well. “If clinical AI learns from the people who need it least, it will work best for the people who need it least.”

It ties straight to CMS’s ACCESS Model, which pays only when a share of your panel hits target — a scoring rule that quietly rewards enrolling the patients most likely to succeed.

❓ Every eval-harness conversation we have is about accuracy on a test set. But if the test set itself is the Apple-Watch demographic, a “validated” model can be biased and pass anyway. What does a representativeness check look like as a standard step in a clinical eval — and who ships that as a tool first?

Software factories are coming for the whole development loop.

Two of the sharpest AI-coding vendors used the same phrase this week. Warp’s Zach Lloyd and Cursor’s forward-deployed engineering lead both describe a shift from “engineer chats with an agent” to a “software factory”: continuous, automated triage → spec → implement → review → verify → ship → monitor, with humans on the high-risk checkpoints.

For a clinician-builder the translation is clean: the factory automates the parts that were never your edge. The one stage that doesn’t commoditize — the review-and-verify checkpoint where clinical judgment decides whether the output is safe — is exactly the stage you’re uniquely built to own.

🔮 Prediction: The clinician-builders who win the next 18 months won’t be the ones who write the most code. They’ll be the ones who design the verification stage of the factory — the eval, the guardrail, the human checkpoint — because that’s the seat only a clinician can fill.

Ultra-shorts

Telehealth infra keeps eating AI features. OpenLoop acquired voice-AI platform Hey Revia — it handles complex patient phone calls for providers — and is folding it into its self-serve telehealth launchpad. The build-vs-buy line for “AI voice/comms” is moving toward buy.

Sharecare put AI navigation on AWS. Sharecare’s new AskMD helps patients parse symptoms, check eligibility, and find care — a reminder that the “front door” land-grab is now an infrastructure deal, not a feature.

Anthropic aimed a model at the lab. Claude Science targets research and pharma — worth watching for anyone tracking where clinical-adjacent AI tooling shows up next.

Elevance wired $342M back to CMS. After years of “substantial and persistent noncompliance” on risk-adjustment coding — including submitting corrections on encrypted flash drives instead of the required electronic systems. The data-integrity plumbing under Medicare Advantage is a real, unglamorous build surface.

Jennifer Baron (Cityblock Health) argued that generic AI wasn’t built for the clinical and social realities of dual-eligible patients, and that provider-led tooling is the way to close the “AI generalization gap” — the same dataset-bias worry as the wearables story, from the Medicaid seat.

Bhargav Patel, MD, MBA posted a 7-page breakdown of the MIRA autonomous-agent paper for a physician audience — a useful, caveat-forward read on why “beats doctors in a simulation” and “safe to deploy in an ED” are very different claims.

🛠️ From the Workbench

The FHIR-MCP server ecosystem grew up.

Last week the interesting repo was fhirHydrant — a lightweight Node.js FHIR MCP server. This week the point is that it’s no longer alone. WSO2 ships an enterprise-grade FHIR MCP server with SMART-on-FHIR auth, an Epic Sandbox demo, and three transport modes; LangCare (Go) exposes 40+ clinical skills and multi-EHR config. The category has moved from “cool demo” to “deployable infrastructure” — which matters, because the winter HTI-5 proposal explicitly named MCP as a future interoperability standard alongside FHIR.

⚠️ Verify: These are open-source projects, not compliance products. “SMART-on-FHIR auth” and “enterprise-grade” are engineering descriptions, not a BAA. Stand them up against synthetic FHIR data on localhost; do not point one at real patient data until your own security and legal review says so.

😤 “MCP in healthcare is a solution looking for a problem.” Maybe.

💡 80/20: Clone WSO2’s server, connect it to the Epic Sandbox, and ask an agent one real clinical question against synthetic data. You’ll learn more about where MCP breaks on FHIR in an afternoon than in a month of reading the spec.

🎙️ From the Pods

🎙️ Health Tech Nerds Radio — “Discussing WISeR and the Merits of Prior Auths” (Jeremy Friese, Humata Health)

Friese’s frame on the CMS WISER pilot in Oklahoma: automate the submission, automate the yes, and reserve humans for the 5–10% of cases that genuinely need adjudication. His hard line — “AI can and should only be used to say yes” — plus radical transparency (show the NCD/LCD criteria right in the portal) is the most builder-usable prior-auth philosophy I’ve heard.

💡 Builder take: The reusable idea is transparency-as-feature — surface the exact clinical criteria a decision is judged against, in the workflow. That’s a shippable pattern, not a moonshot.

🔇 Speaker Blindspot: Motte-and-bailey. “Our AI only says yes” is the easy-to-defend motte; the industry’s AI that denies care is the bailey he steps around (”that’s not us”) when the Highmark-drops-the-human example comes up. And “every auth but one came through our portal” is survivorship framing — the happy adopters. The unasked question: if AI only approves and humans only see denials, who ever audits a wrong yes?

💡 BTW: Anupam Jena — senior author on the Real-POCQi eval above — is also the physician-economist behind the Freakonomics, MD podcast, and his PhD adviser was Steven Levitt, the Freakonomics co-author himself. His signature work uses “natural experiments” to find hidden forces in medicine: one of his most-cited findings is that heart-attack mortality actually drops when senior cardiologists are away at their national conference. Random Acts of Medicine (Jena & Worsham).

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

— Kevin & AI

He earned a PhD in AI, then went to med school on purpose. Was he early — or just paying attention?

On June 30, a Dallas company called Matic exited stealth with a “Declaration of Independence for Primary Care,” co-developed inside a real provider group (Catalyst Health Group) rather than pitched at one from the outside.

The founding pair is the tell. Calvin Carter spent 30 years making consumer software effortless — his shop built the Chick-fil-A, LEGO, Southwest, and NPR apps. His co-founder, Dr. Alex Sheppert, earned a PhD in AI and ran algorithmic options-trading systems, then deliberately went to medical school around 2020 because he wanted to be a physician before the AI wave arrived.

Here’s the part worth sitting with: Sheppert’s doctoral thesis was on stopping hallucinations through adversarial training — and that’s the whole product philosophy. Matic runs multiple models against each other (”Thunderdome”), turns every physician correction into a permanent training case, and frames the goal as “error prevention versus error correction.”

The bottleneck was never the model. It’s the clinician who knows when the common answer is the wrong one.

Their own worked failure makes the point: an early note wrote “start empagliflozin” — with no dose. A naive model guesses the statistically common answer because that’s what the internet says most. A clinician knows the dose, and knows this patient.

That’s the same nerve John Lee hit this week: medicine doesn’t have a stable ground truth. The map — consensus, codes, the most-common answer — is not the territory, which is this patient in front of you. Lee’s confession lands hard: he was trained to treat pain as the “fifth vital sign” and wrote opioid scripts from a map someone else drew. Betting on error prevention is really a bet that the map is unreliable and the clinician is the correction.

😤 “It’s a ten-week pilot with vendor numbers. ‘99.9% note accuracy’ is a marketing slide, not a study.” Fair, and you should hold them to it — durable accuracy and ROI need quarters, not weeks, and “physicians change one word per twenty notes” is a claim I’d want to audit myself. But notice what they optimized for: a candid, published failure mode and a self-correcting review loop, not a leaderboard. That’s the right thing to be measuring, even if the number isn’t proven yet.

😤 “A consumer-app guy and a resident are going to fix primary care?” Stranger things could happen I guess.

❓ What’s the smallest version of “error prevention over error correction” you could build on synthetic data this weekend — a checker that flags the confidently-common answer before a human signs it? I keep thinking the verification layer is a product category before it’s an EHR feature, but I can’t quite draw its edges yet.

💡 80/20: Steal the mechanism, not the marketing. The reusable idea is turning every human correction into a stored test case (bad input → wrong output → right output → why). Do it on 20 Synthea patients and one narrow task, and you have an eval harness that gets smarter every time someone fixes it.

📡 Builder’s Radar

Two AI laws quietly go live today — and both are about who’s still in the loop.

As of July 1, Indiana’s HB 1271 bars insurers from using AI as the sole basis to downcode a claim without a human reviewing the medical record — and bars providers from submitting AI-generated claims without human review.

Same day, Tennessee’s SB 1580 makes it a deceptive practice to market a chatbot as a qualified mental or behavioral health professional — with a private right of action and civil penalties per violation.

The regulatory pattern is consistent: the human review step is becoming the legally load-bearing part of your architecture, not a UX nicety.

😤 “Two states. Who cares.” The private right of action in Tennessee is the part that scales without waiting for the next legislature — it’s plaintiffs, not regulators, who enforce it.

💡 80/20: If you’re building anything patient-facing, write down the exact sentence a clinician says when they review the AI’s output — then make sure your product actually creates that record. “A human reviewed this” is now a compliance artifact you can be asked to produce.

UpDoc made headlines last week. This week, someone pulled the actual FDA file — and the paperwork tells a different story.

Dr. Sam Ashoo read the clearance documents for the “physician-grade agentic AI” everyone was buzzing about, and found three claims that don’t connect. The famous “81% vs 25%” trial tested a deterministic Amazon Alexa flowchart in 2021 — and appears nowhere in the FDA submission.

The FDA cleared it (K253281) by substantial equivalence to a 2018 insulin dose calculator, with the decision summary stating “no clinical testing was performed.” The change-control plan legally locks the dosing engine as deterministic — so the LLM layer that’s marketed as “agentic” is the one part never clinically validated.

“Cleared,” “clinically validated,” and “agentic” were three separate stories sold as one.

😤 “This is just a hater dunking on a startup that shipped.” No — the care gap is real, and they chose the FDA pathway instead of dodging it, which deserves credit. The critique is narrower and fairer: know what’s actually running in the commercial product versus what ran in the trial, because they’re not the same system.

🔮 My bet: within a year, “read the 510(k) decision summary, not the press release” becomes a standard diligence step — and at least one “agentic” clearance gets publicly re-litigated on exactly this gap.

Medicare launches a chronic-care model this week that turns “generate a note” into a payment requirement.

CMS’s ACCESS Model — a 10-year, outcomes-aligned program for tech-enabled chronic care (hypertension, diabetes, MSK, depression) — begins this week, pulling 150+ tech-enabled organizations into Original Medicare. Participants must use FHIR-based APIs to share clinical updates with the patient’s other providers.

The wedge: a lot of these entrants know how to monitor and generate data. They don’t know how to generate a note a PCP can review, co-sign, and get paid for in under two minutes.

There’s a real difference between a blood-pressure trend chart and a clinical update a physician can act on — and Medicare is now paying for the difference.

The other cost of your ambient scribe — the one nobody at the go-live could quote.

Dr. Gigi Magan convened a sharp conversation on what clinical AI costs before it reaches the exam room: energy, water, emissions — and who bears them. A health system rolled out an ambient scribe clinicians love, and no one could say how much energy a single encounter uses.

The usable part is a framework: SAHAI (Sustainably Advancing Health AI, NEJM Catalyst, from pediatrician-informaticist Dr. Chethan Sarabu and colleagues) — measure the energy/emissions/cost of a specific clinical use, like energy per AI-answered inbox message, and choose the lighter option.

😤 “This is guilt, not engineering.” It’s the same discipline you already preach: you cannot manage what you cannot count. “Use the model when you need reasoning; use a plain search when you need a link” is just cost-per-correct-task applied to watts instead of dollars.

⚡ Quick hits

Claude Sonnet 5 shipped — roughly Opus-4.7-class on coding and agents at the same price and speed as the last Sonnet, 1M-token context. If you route complex tasks to a pricier model for quality, re-run your evals — a same-price capability jump quietly resets a lot of routing decisions.

“Shift for healthcare” — a VC-lens read on the robotics-training-data gold rush argues privacy in a clinic is “10x harder and 100x riskier” than blurring faces in an apartment — “you can see that as a challenge, or as a moat.” 🔮 Bet: a dedicated healthcare data-pipeline company gets funded within three years.

🛠️ From the Workbench

fhirHydrant — an open-source FHIR MCP server (Node.js): SMART Backend Services auth, metadata-aware search/CRUD tools, compact responses, FHIRPath filtering, safe pagination, audit events, terminology lookup. In plain terms: it’s the interface layer that makes clinical data legible to an agent — the missing primitive between “I have a FHIR endpoint” and “my agent can safely read and write to it.”

It pairs with a thesis worth internalizing: agents win where software is open enough to give them context and API-first enough to let them act. That’s the whole reason MCP matters for clinical tooling.

⚠️ Verify: this is early, unaudited community infrastructure. It’s a superb thing to learn on with synthetic data / a public sandbox — but SMART Backend Services auth touching real PHI is a security-review conversation, not a weekend deploy.

💡 80/20: Stand it up against a public FHIR sandbox (Medplum is great) and point an agent at it. In an afternoon you’ll understand why “my tool integrates with the EHR” is a checkable claim, not a slogan — and you’ll have a repo to point your innovation team at.

🎙️ From the Pods

🎙️ The 229 / Unhack — “Wrangling Hidden AI Agents” (Drex DeFord + Jason Elrod, MultiCare)

The security angle nobody budgets for: shadow AI agents are already proliferating inside health systems, and the interesting defense is building your own counter-agent that feeds bad data to adversarial research bots. The pearl underneath, though, is a discipline: “Don’t delegate thinking. Cognitive atrophy will happen. Use it as a thought partner, not a replacement.”

💡 Builder take: The next governance category isn’t “which AI did we buy” — it’s “which agents are running that we didn’t deploy.” Build the inventory before you build the fifteenth tool.

🔇 Speaker Blindspot: Anecdotal generalization — a two-CISO conversation (”I built one, so everyone should”) is a vivid story, not evidence of what works at scale. The “everyone needs an agent” advice skips the org where an unmonitored agent is the breach.

🎙️ Radio Advisory — “The work isn’t easy, but it’s clear” (Rachel Woods + Advisory Board State of the Industry team)

The tonal shift worth noting: leaders moved from last year’s “plan for every scenario” paralysis to “we know the game plan now” execution. The word they kept hearing was resiliency — not optimism, but a refusal to treat failure as an option.

💡 Builder take: A buyer with clarity and urgency buys tools that help them execute a known plan — not tools that ask them to imagine a new future. Pitch the plan they already have, faster.

🔇 Speaker Blindspot: Survivorship bias — you interview the leaders and organizations still standing, and “resiliency” is exactly the story survivors tell. The systems that already folded under the same pressures aren’t in the sample.

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

— Kevin & AI

The benchmark says ready. The stress test says no.

A new Nature Medicine paper put the flagship frontier models — GPT-5, Gemini, the usual leaderboard — through adversarial stress tests on health tasks.

The finding is unsettling in a specific way: the models could often guess the right answer with key inputs removed, then get confused by a trivial change in wording — fabricating fluent, convincing reasoning traces for the wrong conclusion.

The same authors showed that popular health benchmarks vary wildly in what they actually measure. A high score on one isn’t a high score on “is this safe in a clinic.”

Readiness was never the accuracy number. A system that aces the benchmark and breaks on a rephrase hasn’t learned medicine — it’s learned the shape of the test.

Here’s why the timing matters. A federal program, ARPA-H’s ADVOCATE, is funding autonomous agents to adjust heart-failure meds and appointments on a three-year FDA path — built on exactly these models.

And the ground underneath them is getting noisier. Ambient AI wrote the note. An AI coding tool picked the diagnosis. The next model trains on that corpus and grades itself against the thing it helped write.

A benchmark is not the territory. It’s the model exposed to our particular way of asking — and we keep mistaking the reflection for the patient. The most honest thing in the whole paper is that we don’t yet have a measurement that survives contact with a clinic. That’s not a reason to look away. It’s the most interesting place to build.

😤 “Every model fails adversarial tests. Humans do too — show me the doctor who never anchors.” True, and that’s the right comparison to demand. The difference is calibration: a good clinician usually knows the edge of their knowledge and slows down; these models narrate maximum confidence right off the cliff. The fix isn’t a smarter model, it’s an instrument that catches the overconfidence before a human signs it.

😤 “This is academics moving the goalposts. The models keep getting better.” They do. And the goalpost should move — that’s what “readiness” means in a field where being wrong has a real implications.

😤 “So your answer is wait.” No. My answer is measure the right thing, then ship the narrow piece that survives the measurement. “Wait” and “deploy autonomously” are not the only two options.

❓ What product lives in the gap between the benchmark and the bedside? Not another leaderboard — something that tells a clinician, in the moment, “this output is the kind the model is brittle on.” I think there’s a real tool in the wobble itself, and I can’t quite name it yet.

California just fined a primary-care chain $4.4M for who owned it.

The state’s first-of-its-kind settlement with Carbon Health found the MSO effectively controlled medical decisions — violating California’s ban on the corporate practice of medicine — and misled patients on billing. Carbon pays $4.4M; co-founder Eren Bali pays $100K.

If you’re building a care-delivery company with a “friendly PC” structure, the org chart is now a regulatory surface, not a legal footnote.

The “friendly PC” wrapper that everyone copies from everyone is now a thing an attorney general will read line by line.

😤 “This is a California problem.” It’s a California first. CPOM statutes exist in a lot of states and have been gathering dust; AG offices just learned the structure is enumerable. Don’t bet your cap table on dust.

💡 80/20: If your model employs clinicians through an MSO, have someone who isn’t your incorporating lawyer answer one question: who actually controls the clinical decision? If the honest answer is “the company,” you have a Carbon problem in miniature.

Harvard trained a foundation model to predict your next diagnosis.

Circulating hard in clinician feeds this week: MGB’s DT-Transformer, a GPT-style model trained on 57 million structured EHR events that forecasts a patient’s next diagnosis — and roughly when — across nearly 900 diseases.

The build lesson isn’t the model. It’s that “read a lifetime of diagnoses as a sequence and predict the next token” is now a tractable pattern on real records, not a research toy.

❓ The validated model exists and a patient’s own data is right there in the portal export. “Wait for your cardiologist” — is that caution, or is it the same paternalism the #WeAreNotWaiting crowd already routed around once?

Connecticut’s AI disclosure mandate goes live tomorrow.

As of July 1, Connecticut’s amended privacy law requires LLM-backed chat and voice agents to disclose, in the privacy notice, that personal data may be used to train the underlying model. Meanwhile Colorado just narrowed its landmark AI Act and pushed it to January 2027.

The compliance perimeter isn’t a wall you build once — it’s a different shape in every state, and it moved twice this month.

Ultra-shorts

Graham Walker, MD (MDCalc / Offcall) — flags a Pediatrics paper on sociodemographic variability in AI-driven pediatric ED decisions, and says it finally gave him language for his unease.

Yair Saperstein, MD (AvoMD) — the #1 line his Epic inbox served up overnight: “Are you caring for this patient?” The ambiguity in that one sentence is a whole product category for clinical AI that knows who owns the patient right now.

🎙️ From the Pods

🎙️ Health Tech Nerds Radio — “The Grand Roundup“

Two agentic voice-AI companies announced big rounds on back-to-back days — and both press releases carried a customer quote calling that vendor “the only true platform” after an exhaustive RFP. Same week, opposite winners.

The unsexy question the hosts kept circling: voice AI is token-intensive, and nobody’s shown the gross margins versus a human call center once the subsidy burns off.

💡 Builder take: When two competitors both claim to be the only platform, the differentiator buyers can’t fake is unit economics. If you’re in this space, know your cost-per-resolved-call cold — that’s the number that survives the froth.

🔇 Speaker Blindspot: Survivorship bias — the case studies that travel are the practices that saw a million dollars in new revenue. The clinics where the IVR jammed up the schedule don’t get a slide.

🎙️ The 229 — “Fable Goes Dark and the Demo Derby Problem“

A powerful coding model — Mythos — got pulled offline under federal pressure after someone social-engineered it past its guardrails. Drex DeFord’s line is the pearl: guardrails on generative models “are not deterministic — they are probabilistic.” The same model that audited a health system’s codebase in two days was handed to the attackers too.

💡 Builder take: If your security story is “the model has guardrails,” you don’t have a security story. Treat model guardrails as a probabilistic filter, and put deterministic controls — scopes, allow-lists, human gates — around anything that can act.

🔇 Speaker Blindspot: Appeal to common practice (tu quoque) — the vendor’s defense was “everybody else’s model has the same problem.” True, and completely irrelevant to whether yours should be wired into a hospital network.

💡 BTW

💡 BTW: Eren Bali — the Carbon Health co-founder who just personally paid $100K in that California settlement — was born to Kurdish parents in an apricot-farming village in Malatya, Turkey, where his mother taught grades one through five in a one-room schoolhouse. He taught himself math on early internet forums, won gold at the Turkish Math Olympiad and silver at the International Math Olympiad, then co-founded Udemy before Carbon. Wikipedia

📅 Upcoming: ONC Health IT Certification Developer Roundtable (Wed Jul 1, free) — open to any developer touching an EHR, not just certified ones.

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

— Kevin & AI

A German court ruled the platform — not the user — owns what the AI confidently said. The verification layer just changed buyers.

The Regional Court of Munich held Google liable for false claims its AI Overviews generated about two publishers, drawing a line between listing sources and generating new statements.

The court’s logic is the part that travels: an AI Overview makes “independent, new, and substantive statements,” so the platform speaks them. When the model stops flagging and starts answering, someone owns the answer.

For years we treated verification as a quality feature. A liability ruling reframes it as a legal artifact.

😤 “It’s a German injunction about search snippets. Nothing to do with US clinical AI.” Jurisdiction differs; the principle doesn’t. The question every legal system is now circling is the same — when generated text asserts something false, is that the user’s problem or the system’s?

😤 “So verification is a lawyer feature now? Great, more compliance theater.” A checkbox that nobody reads is theater. An edit-distance log showing the clinician actually changed the draft is evidence. One of those survives discovery.

😤 “You’ve been on about liability for weeks.” Different fork. The earlier point was that signing the order keeps responsibility with the clinician. This one is the opposite direction — liability newly attaching to the platform and the vendor for the words the model generates. That’s who has to buy their way out of it.

An LLM can grade care quality, not just chart it.

A UCSD-led cluster randomized trial in JAMA Network Open used LLM-based abstraction of the record to auto-assess sepsis-bundle compliance and feed targeted nudges back to physicians — measurement and feedback, not diagnosis.

The under-built use case for clinical LLMs isn’t answering the question. It’s reading the chart to tell you whether the team did the thing the protocol required.

Ultra-shorts

Nikhil Krishnan (Out-Of-Pocket) — launching Ship It, an application-only software-engineering conference for healthcare, all workshops. A signal that “clinician/engineer who builds” is now a room worth gatekeeping.

Christina Farr (Scrub Capital) — new 6-episode series Rounds: Where Healthcare Meets Wall Street; episode 1 is Joanna Strober (Midi Health) on women’s health.

Adam Carewe, MD — “Who taught you that physician and entrepreneur could not be the same word? Trace it back. It was never you.”

🎙️ From the Pods

🎙️ Lenny’s Podcast — “OpenAI Codex lead on the new shape of product work“ (Andrew Ambrosino)

Nearly 100% of OpenAI uses Codex weekly, and the process inverted: when anyone can stand up a feature, you don’t get one PRD, you get 90 prototypes. The scarce skill is curation — deciding which is good.

💡 Builder take: “Implementation is not the expensive part anymore — it’s taste.” For clinical teams, taste is the verify-the-output muscle you already need. The edge isn’t building the thing; it’s judging it.

🔇 Speaker Blindspot: Composition fallacy — OpenAI engineers curating 90 low-stakes prototypes is not a clinician curating 90 AI suggestions with a patient’s potassium and a malpractice carrier attached. “Taste” scales cleanly in software; in medicine, the cost of bad taste is asymmetric.

📅 Health AI Events

Free virtual events for clinician-builders — attend live or catch the recording later.

Tue Jun 30 — Clinician in the Loop: AI Investment to Real-World Impact (CHIME, LinkedIn Live)
12:00 PM ET · LinkedIn Live · Free · Replay available
For CMIOs and clinical leaders: how to tell whether your AI is actually working post-deployment, and the gap between vendor promises and clinical reality. A 30-minute lunch watch.

Thu Jul 16 — Accountability in Practice: Responsible AI Use in Healthcare (AHIP / URAC)
1:00 PM ET · Virtual · Free
Bias, transparency, and oversight from the payer’s seat — plus how a standards-based framework keeps an AI deployment defensible. These accountability questions eventually land on every builder’s desk.

Tue Jul 21 — Becker’s AI + Digital Health Virtual Event (Becker’s Healthcare)
1:00 PM CST · Virtual · Free to register · Sessions archived
The most on-topic event on the calendar — a half-day on AI and digital health in care delivery. Worth registering now even though it’s a few weeks out.

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

— Kevin

The Spartans punished soldiers who fought too bravely.

Not the cowards. Everyone punishes cowards. The Spartans also disciplined the soldier who charged too far ahead, broke the line, fought with too little regard for his own life — because his recklessness endangered the men standing next to him.

Aristotle made the same point without the spear. He said the opposite of courage isn’t cowardice. Courage sits in the middle, between two failures. Too little nerve and you’re a coward. Too much and you’re reckless. Courage is the narrow thing in between — and it only counts if you actually feel the fear. If there’s nothing at stake for you, charging ahead isn’t brave. It’s just a gamble with the dial turned all the way up.

I think about this every time someone tells me to move fast and break things.

Move fast and break things is a fine rule when the thing that breaks is a landing page. It’s a fine rule when you’re experimenting with a social app, a side project, a startup that fails and takes only your weekend with it.

It’s a different rule when the thing that breaks is a person.

The hard part of building in medicine was never the courage to ship. Anybody can ship now. The barrier is gone. You can vibe-code a clinical tool on a Saturday and have it talking to an API by Sunday. Going all in has never been easier.

But going all in isn’t courage when someone else cashes the check.

The reckless builder ships the med-rec tool that’s right 95% of the time and never asks who eats the other 5%. The reckless builder calls this boldness. It isn’t. It’s a soldier breaking the line.

The coward is the other failure. The coward is the committee that meets for eight months and ships nothing, because the safest-looking move is to never let the prototype exist. That isn’t caution. That’s the dial turned the other way.

Courage is the narrow middle. It’s building the eval before you trust the output. It’s running it on synthetic patients first. It’s shipping the thing — actually shipping it — while holding in your head, the whole time, the potassium of 7.2 at 2 AM that the tool might wave through.

Here’s the part that’s yours.

You know what the check costs. You’ve signed it before, in your own handwriting, in a chart with your name on it. You know what a missed interaction looks like at the bedside, not as a failure mode in a slide deck. That fear isn’t a weakness to engineer away. It’s the exact thing that turns shipping into an act of courage instead of a roll of the dice.

The engineer who’s never been in the room can’t feel that fear. So when they go all in, it isn’t bravery. They’re just playing all in with someone else’s chips.

You can feel it. That’s not a bug in the clinician who builds. That’s the whole edge.

Move fast. But know whose things break.

Things are moving faster than ever, but will never move this slowly again. What are you building this week? Got a great tool? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin & AI

UpDoc gets the first FDA clearance for an agentic clinical AI — and a $18M seed to go with it

UpDoc, founded by Stanford Medicine researchers, announced FDA clearance on June 25 for what it calls the first agentic clinical AI platform — agents that adjust medications, order labs, and coordinate care between scheduled visits.

It came with an oversubscribed $18M seed (Eli Lilly, Mayo Clinic, the American Diabetes Association, Section 32, Pear, Polaris) and live deployments at Cleveland Clinic, Allegheny Health Network, and UCSF.

Here’s the part the press release glides past: the underlying clearance looks narrow — a 510(k) scoped to automated insulin dosing, riding the d-Nav predicate — while the WSJ framing paints a sweeping “concierge doctor between visits.”

The agent that adjusts the dose got cleared. The clinician whose name is on the order still owns what happens next.

This is the year AI crosses from reads-and-flags to talks-and-acts — and the accountability layer didn’t move with it. For builders, the lesson isn’t the funding round. It’s that the regulatory pathway is now a product decision: UpDoc went the federal SaMD route; Doctronic is being regulated at the state level as the practice of medicine. Same capability, opposite moats.

😤 “This is an insulin-dosing algorithm with a press release.” Partly, yeah — and that’s exactly the tell. The clearance is narrow; the vision is broad. The gap between what cleared and what’s marketed is where every regulated-AI builder actually lives. Naming that gap honestly is the skill.

😤 “FDA clearance means it’s safe.” Clearance means it matched a predicate. Safety is present-tense; clearance is past-tense.

😤 “Doctors will never let an AI change a dose.” They already let order sets, protocols, and pharmacy algorithms change doses every day. The question was never whether — it’s who signs when it’s wrong.

❓ What’s the product that lives between UpDoc’s federal SaMD route and Doctronic’s state practice-of-medicine route?

Trase raises a $107M seed to be the operating system for AI agents in regulated industries

Trase closed an unusually large $107M seed (ARCH Venture Partners, Red Cell) for an agent OS aimed at healthcare and other high-stakes fields, with a live Duke Heart Center deployment.

The thesis investors are now paying for: in regulated work, the moat isn’t the model — it’s the governance, audit trail, and predictability around the agent.

The capital is moving toward the boring layer — the part that makes an agent’s actions defensible, not just impressive.

😤 “$107M for a seed? That’s a vibes round.” Maybe. But notice what got funded: not a clinical model, the connective tissue around agents acting safely. That’s a bet that the action layer, not the intelligence layer, is where the durable value sits.

💡 80/20: Whatever you build with agents, build the receipt first — who acted, on what authority, with what inputs, reversible how.

Patient messages now outpace office visits — and that’s a demand signal, not a burden

A JAMA study of 140M+ records across 2,067 hospitals found patient-authored portal messages surged 153% from 2020 to 2025 — now outpacing in-person visits, with messaging expanding between-visit care rather than replacing it.

The reflex is to treat the inbox as a problem to absorb: more staff, more after-hours clicks, bill for messages. The sharper read is that the volume is unmet demand to be heard — people reaching for the closest thing to an inside line.

The message isn’t the problem. It’s the signal that the visit-shaped system no longer fits how people actually seek care.

😤 “So we should just let AI auto-reply to patients?” No — auto-replying to a demand signal is how you turn a person who finally reached out into a person who never will again. The build is triage and routing that gets the scary message to a human faster, not a bot that closes the loop on the cheap.

❓ The patient is the only actor in this system without an AI working for them — the clinician has scribes, the payer has denial engines, the patient has a portal they can barely log into. What does the patient’s agent look like, and who’s allowed to build it?

Hera raises $27M for chronic-care management on Original Medicare

Hera raised $27M for a CCM platform staffing 1099 RNs and social workers who bill “incident to” a physician to coordinate care for fee-for-service Medicare patients with 2+ chronic conditions — reportedly at ~60% gross margins.

The between-visit gap isn’t just a clinical problem — it’s a reimbursable business with real unit economics, and the codes already exist.

💡 80/20: If you’re building between-visit tooling, the CCM/RPM/APCM code stack is the revenue rail that’s already there. Learn which actions map to which codes before you build the workflow — the billing logic is the product spec.

Quick hits

Chris Klomp nominated as Deputy HHS Secretary. The CMS Medicare Director was nominated as #2 at HHS — a health-tech-literate operator now over operations and regulatory policy. Worth watching for anyone whose roadmap touches Medicare Advantage or interoperability rulemaking.

Epic / UCHealth — a behavioral-health unit with no bedside phones gave nurses a MyChart-integrated meal-ordering app instead of routing calls through the kitchen; 99% of meal orders now flow through it. Clean example of workflow redesign that removes burden without adding staff.

Angel Arnaout, MD (Healthcare Management Forum) — a sharp paper arguing AI capability is an organizational property (infrastructure, governance, skill, culture), not a purchase; route AI requests through problem statements, not technology requests. The “AI-capable organization” framing is the buyer-side mirror of today’s Big Thing.

🛠️ From the Workbench

Connecting WHOOP to Claude — Christian Pean MD wrote a hands-on guide for piping WHOOP recovery, sleep, and strain data into Claude so a scheduled task files your nightly metrics into a folder and a living dashboard. About 15 minutes of setup (WHOOP membership, a WHOOP developer account, Node.js), with a no-code export shortcut for the less technical. The real lesson isn’t the wearable — it’s the pattern: a scheduled agent turning a stream of “data I have” into “data I can think with.”

⚠️ Verify: this is your own consumer data on your own machine — fine for a personal longevity log. The moment you point this pattern at patients, you’ve left wellness data and entered PHI, and the WHOOP developer terms and your health system’s data rules both apply.

💡 80/20: Build the pattern on your own wearable data this weekend, then ask what it teaches you about the clinical version: what would a between-visit “how am I trending?” agent need, and what’s the first thing that would break when the data is someone else’s?

🎙️ From the Pods

🎙️ HIMSSCast — “Full-body scans for longevity” (with BodySpec CEO Elaine Shi)

DEXA body-composition scans surface visceral fat and lean mass that the scale and BMI hide — and Shi notes clients increasingly pull their DEXA results “into their LLM of choice” to reason about long-term risk, not just today’s number.

💡 Builder take: The wearable-and-scan → LLM workflow is becoming the default consumer move. The opportunity isn’t the measurement; it’s the trustworthy layer that turns longitudinal body data into a decision a clinician would actually endorse.

🔇 Speaker Blindspot: Selection bias — the people who scan monthly are the already-converted quantified-self crowd. The patients who’d benefit most from catching visceral fat early are exactly the ones who’ll never book a voluntary DEXA, and the episode treats the engaged self-tracker as the norm.

“I always want to be confronted with the truth.” — Elaine Shi

🎙️ HLTH Webinar — “Unlocking Innovation with Rural Health Transformation”

The federal Rural Health Transformation Program will push $50B to states over five years — for scale, that’s roughly HiTECH’s ~$29B but in half the time, with ~$200M per state in year one. The panel’s repeated warning: this only works if the money buys sustainable operating models, not one-off projects.

💡 Builder take: A lot of that money will land on rural AI and automation to stretch scarce clinical staff. If you build for low-resource settings, the move is to engage state health leaders now — the year-two applications are where real innovation gets funded.

🔇 Speaker Blindspot: Appeal to scale — the panel leans on the size of the check as if $50B plus technology automatically modernizes rural care. They name DSRIP as a cautionary tale but don’t reckon with why those grants funded projects that evaporated when the dollars did. Bigger pile, same failure mode unless the operating-model discipline is real.

📅 This Week in Health AI Events

Free virtual events for clinician-builders — attend live or catch the recording later.

Tue Jun 30 — Clinician in the Loop: AI Investment to Real-World Impact (CHIME, LinkedIn Live)
12:00 PM ET · LinkedIn Live · Free · Replay available
For CMIOs and clinical leaders: how to tell whether your AI is actually working post-deployment, and the gap between vendor promises and clinical reality. The 30-minute format makes it an easy lunch watch.

Tue Jul 14 — Becker’s Oncology Virtual Event (Becker’s Healthcare)
1:00 PM CST · Virtual · Free to register · Sessions archived
A close look at how AI lands in a high-stakes specialty — where the tolerance for error is lowest and the workflow questions are hardest.

Thu Jul 16 — Accountability in Practice: Responsible AI Use in Healthcare (AHIP / URAC)
1:00 PM ET · Virtual · Free
Bias, transparency, and oversight from the payer’s seat — plus how a standards-based framework keeps an AI deployment defensible.

Tue Jul 21 — Becker’s AI + Digital Health Virtual Event (Becker’s Healthcare)
1:00 PM CST · Virtual · Free to register · Sessions archived
The most on-topic event on the calendar — a full half-day on AI and digital health in care delivery. Worth registering now even though it’s a few weeks out.

💡 BTW: Grant Verstandig, who just raised $107M for Trase, founded his first health company — Rally Health (originally Audax) — after dropping out of Brown at 21, having endured seven-plus knee surgeries including a meniscus transplant. He and his surgeon disagreed about what “a good outcome” even meant, and that argument sent him into building. Fortune

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

— Kevin