The Four Boxes

“I have two kinds of problems, the urgent and the important. The urgent are not important, and the important are never urgent.”

Dwight Eisenhower - 1954

He’d just finished running D-Day and was in the middle of running the country.

The insight wasn’t theoretical.

It was operational.

The man who coordinated the largest amphibious invasion in history was telling a room full of people that the hardest part of leadership isn’t the big decisions — it’s not letting the small loud ones eat the big quiet ones alive.

Stephen Covey later turned the quote into the 2x2 grid most people know: urgent/important on two axes, four quadrants, four actions.

Quadrant 1: Urgent + Important. Do it now. The crashing patient. The server that’s down. The kid with a fever at school. You don’t choose these — they choose you.

Quadrant 2: Important + Not Urgent. Schedule it. The FHIR app you’ve been speccing in your head for six months. The Georgetown course prep. The Scrub Capital scout thesis. The conversation with your spouse about what the next year looks like. This is where all the compounding work lives — and it’s the quadrant that gets eaten first, because nothing in it is on fire today.

Quadrant 3: Urgent + Not Important. Delegate it or decline. The Slack thread that feels like it needs a response right now but doesn’t. The meeting that could have been an email. The prior auth that someone else on the team can handle. These feel like Quadrant 1 because they’re loud, but they’re not yours.

Quadrant 4: Not Urgent + Not Important. Delete it. The doom scroll. The third redesign of a logo for a product that doesn’t have users yet. The LinkedIn rabbit hole that started as networking and ended as comparison. You know this quadrant by how you feel afterward: empty.

Here’s why this matters for clinician-builders specifically.

Clinical practice is almost entirely Quadrant 1. Every shift is a series of urgent, important decisions made under time pressure. You’re trained for Q1. You’re exceptional at it. Your entire residency was a Q1 factory. The muscle is overdeveloped.

Building is almost entirely Quadrant 2. Claude Coding on a weekend. Designing a workflow. Thinking about what the product should actually do before you build it. Reading the FHIR spec. Having the conversation with a potential user. None of this is on fire. All of it compounds.

The failure mode for clinician-builders isn’t laziness — it’s that Q1 muscle memory makes Q2 feel wrong. When you sit down to build on a Saturday morning and nothing is urgent, the absence of urgency feels like the absence of importance. So you check Slack. You answer an email. You find a Q1 task to do instead, because Q1 feels productive in a way that Q2 doesn’t — until you look up and realize the FHIR app is still in your head six months later.

Eisenhower’s insight — the one the unnamed university president actually said — is that urgency and importance are independent variables. They feel correlated. They’re not. The thing that will matter most in your life a year from now is almost certainly not urgent today. And the thing screaming for your attention right now will almost certainly not matter in a year.

Three things to try this week.

Name your Q2 list. Write down the three most important things you’re not doing because nothing is forcing you to. Not the tasks — the outcomes. “Launch the beta” is a task. “Have 10 clinicians using the tool” is an outcome. The outcome is what belongs in Q2.

Block Q2 time like Q1 time. Your shifts are scheduled. Your Q2 time isn’t. That’s the problem. Put two hours on your calendar this week with the same non-negotiability as a clinical shift. If someone asks you to cover, you’re already booked. You are. You’re building.

Audit your last Saturday/Sunday. How many hours did you spend in each quadrant? Be honest. If you spent your building time in Q3 (answering Slack, clearing inbox, “quick” tasks that felt productive but weren’t yours), that’s the diagnosis. The treatment is boundaries, not more hours.

Eisenhower ran a world war and a presidency with this grid. You’re running a clinical practice, a family, and a side project. The grid is simpler than your life, and that’s the point.

Complexity is the excuse for not prioritizing.

The grid removes the excuse.

The important is never urgent.

Build anyway.

What’s in your Quadrant 2 right now? Reply and tell me — I read every one.

— Kevin

Utah’s medical board called for immediate suspension of the state’s AI-doctor pilot. The line on autonomous prescribing just got drawn [or redrawn?].

Reporting on April 24 confirmed that the Utah Physicians Licensing Board has demanded the immediate suspension of a state pilot using an LLM-based bot to renew prescriptions for Utah patients. The same day, the Federation of State Medical Boards posted the board’s letter publicly. The letter’s argument is structural: the pilot was authorized through the executive branch in a way the board considers a circumvention of the medical practice act and of the board’s authority to define what constitutes the practice of medicine in the state.

That letter sits inside a tight one-week stack. Health Affairs Forefront published a critique of the same pilot, framing it as a state operating ahead of federal oversight in a way that exposes patients to safety risks. And JAMA Network Open published a study — Mass General Brigham, 21 leading LLMs, 29 clinical vignettes — that found all 21 models (GPT-5, Claude, Gemini, Grok, all of them) failed 80%+ of differential diagnoses when patient data was incomplete. Final-diagnosis accuracy hit 90%+ when the data was there. The hard part is the iterative reasoning under uncertainty that pulls more data into the picture before settling on a diagnosis. That is what clinicians do. It is also what the Utah pilot was using AI to do, unsupervised.

The framing matters. This was not a vendor deploying a clinical tool that a state board found out about through a complaint. This was a state government, through executive action, declaring that an LLM could practice medicine. The board letter is not about the company; it is about the principle that defining the practice of medicine sits with the medical board, not the executive branch. The FDA has already signaled it will not regulate general-purpose LLMs even when they give health advice. State boards just established that they will fill the regulatory gap. For a clinician-builder, the line is now visible: shipping a tool that helps a clinician prescribe more efficiently is a real product. Shipping a tool that prescribes without a clinician in the loop has a regulatory letter on FSMB letterhead arguing the state cannot — by fiat — declare an LLM a practitioner.

😤 Haters

“This is one state board defending its turf, not a precedent.” Maybe — and the procedural challenge is real. But the FSMB hosting the letter publicly is the move that converts a single-state action into a national template. Other state boards facing similar pilots now have language to copy. The first letter is the precedent; the next five make it a movement.

“AI prescribing is happening anyway through Doximity, OpenEvidence, and ChatGPT for Clinicians — this is just labeling.” That confuses two different things. Decision support tools with a verified human in the loop — even free clinician-facing ones — are not what the Utah pilot was. Watch which states draw that distinction explicitly. The vendors with a human-in-the-loop posture should be glad this happened; it strengthens their position by clarifying it.

“The JAMA paper is just academics being academics — clinicians use Doximity AI all day and it works.” It works because they are the safety layer. The 80%+ failure rate on incomplete data is exactly what gets caught when a clinician reads the model output, notices what is missing, and asks the next question. Take the clinician out, and the failure rate is the actual failure rate. The paper is not arguing against AI in clinical practice; it is mapping where the human still has to be.

💡 80/20: Autonomous prescribing without a clinician in the loop is a product that just got a regulatory ceiling drawn under it, in writing, on FSMB letterhead. Reframe: stop asking “can this AI prescribe by itself?” and start asking “what does the clinician need from this AI in the moment they are deciding to prescribe?” The first question dies; the second one is where the next decade of defensible product gets built.

→ Full write-up

📡 Builder’s Radar

The “good enough” Epic AI debate became a real argument this week. The lesson cuts in both directions.

Bob Wachter published “Damn it, they’re good”, arguing Epic’s increasingly capable AI features are going to slow innovation by giving hospitals a $100K bundled option that undercuts the $1M standalone vendors. Halle Tecco echoed it on LinkedIn from a CMIO conversation. The pushback in the community channels was sharp: hospitals revealing a preference for cheaper-and-good-enough is the system working. If standalone vendors cannot charge $1M for the marginal upgrade, the marginal upgrade is not worth $1M.

😤 Haters

“This is just incumbents being incumbents — Epic always bundles, the market always survives.” It does survive, but the shape of survival changes. Ambient scribes already pivoted out of pure scribe positioning earlier this month because the bundle made the standalone scribe a feature. The same pattern is coming for decision support, image triage, prior auth automation. The vendor who owned the best version in 2024 is now selling a feature.

“The Wachter framing is anti-patient — bundled good enough is the social-welfare optimum.” Cleanly correct on the macro. The error in Wachter’s piece is treating the standalone vendor’s existence as the thing worth protecting. It is not. The patient who gets the diagnosis right is the thing worth protecting. The standalone vendor that cannot show outcomes commensurate with its pricing premium is pricing itself out, not being squeezed out.

💡 80/20: The defensible position for a 2026 clinical AI product is “thing Epic structurally cannot do” — specialty depth, multi-EHR portability, or a workflow that crosses the EHR boundary into payer/employer/pharma data. Try: write a one-sentence answer to “what can Epic not do here?” If the answer is a feature, the bundle is your competitor. If the answer is a structure, the bundle is your distribution clearer.

→ Full write-up

Providence stood up 12 Epic-native AI tools at once. Project Pixel is the Wachter argument made concrete.

Providence — 51 hospitals, headquartered in Renton, Washington — now has 12 AI use cases live inside Epic following an April upgrade, under an internal initiative called Project Pixel. Use cases span ambient documentation, clinical decision support, and operational tooling — categories that, until very recently, had standalone vendors competing for them as separate sales motions. The integration tax is being amortized across 12 features instead of being paid 12 times.

😤 Haters

“12 launched is not 12 actually used. Adoption is the metric.” Right, and Providence has not published outcomes data yet — that is the next thing to watch. But the buyer just declared 12 AI use cases solved by the existing vendor. A 13th vendor knocking is selling against an installed base, not into a procurement window.

“This is one IDN. Not every system is on Epic and not every Epic system has Providence’s internal capacity.” Both true. The constraint is the Epic upgrade cadence and an internal program lead. Both are spreading. The next 18 months are likely to look like a wave: HCA, Ascension, CommonSpirit, Sutter, Banner, UPMC each running their own version with whatever lag their upgrade schedule allows.

Nebraska Medicine’s CTDO says academic systems should build their own AI in-house. The buyer class is changing.

Michael Hasselberg, Chief Transformation and Digital Officer at Nebraska Medicine, argues academic systems now have the talent, the data, and the AI tooling to build internally with better workflow fit than they can buy. His harder claim: the bottleneck is executive alignment, not technology. He also flagged a specific empirical surprise — ambient AI is producing more lift in nursing than in physician documentation, a finding that should rewire any builder still chasing physician-only ambient capture.

😤 Haters

“Build-internal sounds great until the team turns over and the system inherits an unmaintained homegrown stack.” Real risk. The mitigation is a build-with partner who transfers the model, the eval suite, and the maintenance pattern — not a SaaS vendor. The unit economics of services-and-platform are different from SaaS, but the failure mode the hater names is what kills the SaaS-only model in this buyer profile.

“Nebraska Medicine isn’t Mass General. Most academic systems can’t actually build.” Two things. The capacity bar is dropping every quarter as the model + tooling stack matures. And the operator class — CTDO, Chief Transformation Officer, VP AI Roadmap — is being hired in systems that did not have it 18 months ago. The job titles are the leading indicator that the buyer profile is shifting.

💡 80/20: The CMIO who picks one of three vendors after a six-month evaluation is being augmented — and in some systems, replaced — by a CTDO-class operator with an internal platform team that can ship in weeks. Reframe: if your sales motion was built around the long evaluation cycle, you are selling to a buyer who is being routed around. If it was built around being the implementation partner that helps the internal team ship faster, you are selling to the buyer who is being elevated.

→ Full write-up

What are you building this week? Reply and tell me — I read every one.

— Kevin

The RAG-for-medicine author just raised $10M to build evidence-grounded clinical AI.

Almanac Health launched this week with a $10M seed led by F-Prime, with General Catalyst and Lightspeed participating. Total raised is close to $12M. The founder is Cyril Zakka, MD, a physician-researcher whose Stanford work produced what NEJM AI called one of the most-cited papers on retrieval-augmented generation for clinical medicine. The product is described as “safe, evidence-grounded AI for clinicians and health systems” — grounded in peer-reviewed evidence, governed by institutional controls, integrated directly into existing EHR systems, and explicitly free of pharmaceutical advertising.

Read this next to yesterday’s ChatGPT for Clinicians launch and the shape of the category becomes clear. OpenAI owns the distribution play — free to every verified US clinician, 99.6% physician-rated-safe across 7,000 pre-release conversations, a benchmark built on 700,000 model responses. Almanac is not competing on distribution. It is competing on the only thing a generalist model cannot fake: the claim that every answer is tethered to a peer-reviewed source, in an EHR integration a health system can govern, built by someone whose paper is already in the citations.

That second position is where clinical domain expertise becomes the moat. You cannot clone “the author of the RAG-for-medicine paper” with more compute. You can clone access, scale, and polish — but the retrieval index, the evaluation against ground truth, the decision about which specialty pathway gets pre-verified first, the negotiation with an AMC IRB over what counts as an acceptable citation — those are all clinician judgment, applied at every layer of the stack.

😤 Haters

“Every clinical AI company says they’re ‘evidence-grounded.’” Most are. Almanac’s differentiator is provenance: the founder authored the mechanism being commercialized. That is a harder claim to pattern-match around than “we use RAG.” Whether the product is actually better is an empirical question — the company says academic medical centers are doing the validation. Until those studies read out, the pitch is the pitch. But the starting credibility is not nothing.

“A $10M seed in 2026 is noise. OpenAI’s launch is the real story.” One is distribution, one is an evidence-first thesis. They co-exist. The more interesting question is whether the free-ChatGPT clinical tier raises or lowers the ceiling for a specialist product like Almanac. My read: it raises it. Free generalist tools calibrate clinicians to expect citations, BAAs, and audited outputs. Almanac sells into a market OpenAI just primed.

“You’re burying the lede — what’s the actual product?” Fair. Public-facing detail is thin; almanac.chat is a marketing page. What we know: EHR-integrated clinical decision support, peer-reviewed retrieval, currently undergoing validation in AMC settings, no pharma ad surface. What we don’t know: workflow integration granularity, BAA posture, pricing, specialty coverage on day one, and which RAG index they’re actually building against. Track these when the product reaches general availability.

💡 80/20: The funded thesis is that clinician-authored evidence retrieval is the defensible layer, not the model. Try: pick one clinical question you answer 20 times a week and write down the three citations you would want a tool to retrieve before it answers. That list is the spec for what a specialist-grade AI has to hit, and it’s a better evaluation rubric than any public benchmark for your actual workflow.

→ Full write-up

📡 Builder’s Radar

Two no-code healthcare AI agent builders launched the same day. Domain operators — not software engineers — are the named user.

Infinitus Systems launched Studio, a natural-language agent builder that the company calls the first healthcare-specific no-code AI agent platform. Reported metrics: 90% faster deployment, 40% higher accuracy than manually built agents, and a patent-pending Agent Response Control layer that claims to route sensitive clinical or privacy queries through pre-verified compliant response paths. Infinitus says 44% of Fortune 50 healthcare companies already run its platform, with over 100 million minutes of clinical and administrative conversations behind it. The same day, Gravity Rail launched with a $2.75M seed led by Redesign Health: model-agnostic, HIPAA-compliant with a BAA, zero data retention, natural-language SOP-to-agent translation across voice, SMS, email, and web.

😤 Haters

“No-code always leaves the safety layer as someone else’s problem.” Usually yes. Both launches are explicitly selling the safety layer as a feature — Infinitus with ARC, Gravity Rail with a model-agnostic BAA posture and SOP-as-spec. That pitch has to survive red-team pressure before it’s true. But the framing is correct: the reason ops teams cannot ship agents today is not the LLM, it is the compliance wrapper, and both companies are targeting that wrapper as the product.

“44% of Fortune 50 is a vendor vanity number.” It’s a logo count, not a contract value. Still, if even a fraction of those accounts flip from Infinitus’s existing voice product to Studio-built custom agents, the number of health-plan and provider workflows running no-code agents goes up by orders of magnitude this year.

CMS and FDA launched a new breakthrough-device coverage pathway and paused TCET. Medicare coverage can now land 2 months after market authorization, not a year.

CMS and FDA jointly announced the RAPID pathway — Regulatory Alignment for Predictable and Immediate Device coverage — for certain FDA-designated Class II and Class III Breakthrough Devices. The mechanism: CMS issues a proposed National Coverage Determination the same day an eligible device gets FDA market authorization, triggering the 30-day public comment window. Total target: Medicare national coverage and payment as soon as two months after authorization, compared to a year or more today. Eligibility requires an IDE study enrolling Medicare beneficiaries with clinical outcomes agreed on by both agencies. The existing TCET pathway is paused for new candidates while RAPID gets stood up.

😤 Haters

“This is a medtech story, not a software story.” Devices increasingly are software — software-as-a-medical-device, AI/ML-enabled devices, closed-loop systems. The coverage-before-evidence problem has been the single biggest reason clinician-built algorithms don’t get reimbursed. RAPID does not fix that problem, but it shows CMS is willing to accept a tighter evidence loop alongside FDA authorization — a precedent the AI/ML device lane will draft off of.

“IDE enrollment of Medicare beneficiaries is an expensive gate.” It is, and it probably keeps RAPID out of reach for early-stage builders. But the structural move is what to watch — same-day NCD proposal, 30-day public comment, simultaneous FDA/CMS review. If RAPID works, a future lane for algorithm-only Breakthrough designations becomes much easier to imagine.

💡 80/20: The old answer to “when does Medicare pay for this?” was “after FDA, eventually, maybe a year, maybe never.” The new answer is “two months from authorization, if you enrolled Medicare beneficiaries in the IDE.” Reframe: if you’re a clinician-builder advising a startup on a Breakthrough device, the question to ask on day one is not “how do we get FDA?” but “does our pivotal study enroll Medicare beneficiaries, and if not, why not?”

The AMA told Congress this week: AI chatbots should be statutorily barred from diagnosing or treating mental health conditions.

In letters to multiple congressional committees Wednesday, the AMA asked for clear statutory boundaries prohibiting AI chatbots from engaging in mental health diagnosis or treatment — no offering anxiety or depression diagnoses, no recommending medications, no presenting as a licensed clinician. Additional asks: mandatory disclosure that users are talking to AI, suicide/self-harm detection and de-escalation language, escalation pathways, post-market monitoring, serious-incident reporting (especially for pediatric use), limits on advertising to minors, and strict data-retention controls. The context is multiple reported cases of young users dying by suicide after confiding in chatbots that appeared to encourage self-harm.

😤 Haters

“This is doctors trying to gatekeep a modality that’s helping people who can’t access care.” It’s both. The gate is real, and the people falling through the current gap are real too. The AMA letter is not asking to ban chatbots — it’s asking to prohibit them from presenting as licensed clinicians and making diagnosis-and-treatment decisions without human oversight. That line is the same one that defines what clinician licensure is for in the first place.

“Regulation will freeze the useful products out of the market.” Regulation written correctly will freeze out the dangerous products while leaving triage, psychoeducation, and referral routing — exactly the places clinician-built tools could ship today — inside the lines. Regulation written poorly will do the opposite. Which version Congress picks is the question.

💡 80/20: If statutory boundaries land, the non-diagnostic-non-treatment layer is where clinician-built mental health tools get to operate — screening workflows, escalation logic, referral routing, after-visit education. Try: if you’re building in this space, write down where in your product the output stops being “information” and starts being “a diagnosis or treatment recommendation.” That line is your regulatory perimeter, and it should exist in the product today, not in a future compliance review.

🧰 Builder’s Tip

Workflow Pattern — Audit your AI’s retrieval before you trust its answer.

Almanac’s whole pitch is that the retrieval layer is the product, not the model. You can run the same audit on any clinical RAG tool you’re evaluating, including free ones, in about an hour.

1. Pick 20 clinical questions you already know the answer to. Not edge cases — common ones you answer on a normal week. Write down the single best citation for each: guideline, Cochrane review, landmark RCT, UpToDate card, whatever is the canonical answer in your specialty.

2. Run the 20 questions through the tool. Save every answer and — critically — every citation the tool retrieves.

3. Score two things separately. Answer correctness (did it get to the right conclusion?) and retrieval fidelity (is the citation the right one, the wrong one, fabricated, or merely adjacent?).

4. Keep the disaggregated numbers. A tool that is 95% correct with 40% retrieval fidelity is a tool that is trained to sound right, not a tool that is grounded. That’s a meaningful distinction, and it’s the one Almanac is monetizing.

5. Re-run the same 20 questions every 3 months. The retrieval index drifts. The model changes. Your audit is the only thing that catches it.

This is the cheapest, most under-used clinical AI evaluation — most teams skip straight to LLM-as-a-judge benchmarks and never look at what got retrieved. Look at the retrievals first. The answer follows the citation, not the other way around.

What are you building this week? Reply and tell me — I read every one.

— Kevin

OpenAI just made ChatGPT free for every verified US clinician — and shipped the first serious benchmark for real clinician chat tasks.

OpenAI announced ChatGPT for Clinicians — free for any verified US physician, NP, PA, or pharmacist. Reusable “skills” for referral letters, prior auth, and patient instructions. Trusted clinical search with cited peer-reviewed sources. Deep research across medical journals. CME credits auto-tracked from clinical research sessions. An optional HIPAA BAA. Conversations not used for training. The AMA’s 2026 survey says 72% of physicians now use AI clinically, up from 48% a year ago.

Alongside it: HealthBench Professional, an open benchmark for real clinician chat tasks across three categories — care consult, writing and documentation, and medical research. 525 rubric-graded tasks. OpenAI reports physicians rated 99.6% of responses safe and accurate across 6,924 pre-release conversations.

The benchmark sets the evaluation bar for the category. The category, as of this week, is no longer “general-purpose LLM with a medical prompt.” It’s clinician-calibrated chat with a measurable floor.

😤 Haters

“This is just another model with a clinical coat of paint.” The model is a GPT-5.4 variant. What’s different isn’t the weights — it’s the distribution, the benchmark, and the clinician-adjacent primitives (skills, cited search, BAA). Ship enough free clinician AI and the old moat — “we’re the one tuned on medicine” — evaporates. Whether this is a good thing depends on where you sit.

“A Nature Medicine RCT literally just showed LLMs make patient self-assessment worse.” Right — Bean et al. in this week’s Doctor Penguin randomized 1,298 UK participants and found LLMs did not improve (and sometimes worsened) symptom triage vs. normal home resources. Two users with near-identical subarachnoid hemorrhage descriptions got opposite advice. But that study is about patients using general LLMs. ChatGPT for Clinicians is built for clinicians, who bring domain priors patients don’t have. The category distinction matters — and HealthBench Professional is OpenAI’s attempt to measure exactly that distinction.

“The BAA is optional.” It is. That is the single most important sentence in the announcement and the one least likely to be highlighted in any compliance conversation. “Optional” means the default tier is not HIPAA-covered. Any clinician pasting a real patient encounter into a default account before provisioning the BAA is out of scope. Treat the BAA as a dependency, not a feature.

[nota bene: onboarding to this is kinda invasive and requires you to upload your driver’s license (at least)]

📡 Builder’s Radar

The Medicare GLP-1 Pilot just died. The “Bridge” got extended through 2027, and the prior-auth attestation is a PCP problem now.

CMS shelved the Part D piece of the BALANCE Model after UnitedHealthcare and Aetna declined to participate — without them, CMMI couldn’t reach the 80% beneficiary coverage threshold. Instead, the Medicare GLP-1 Bridge was extended through December 31, 2027. The Bridge operates outside Part D: federal appropriation, not plan risk, $50/month copay that does not count toward the $2,100 OOP cap. Wegovy and Zepbound only. Prior auth attestation required — and attestation is clinical eligibility documentation by the prescribing clinician.

😤 Haters

“This is a policy story, not a builder story.” The attestation workflow is the builder story. A panel of 2,000 primary-care adults likely contains 600-800 Bridge-eligible beneficiaries. If the attestation template isn’t in the EHR before July, the clinic either doesn’t do the Bridge or the docs are clinically unsupportable. Someone has to build that template.

“Amazon One Medical already undercut this.” Same week Amazon launched GLP-1 through One Medical at $25/month insured or $149 cash (last week’s issue covered the launch). Every minute of PCP scheduling friction on the Bridge is a referral out to Amazon. The retail-vs-PCP question for obesity medicine is now answered by workflow friction, not by clinical quality.

“Predictive decoration” — someone finally named the thing every hospitalist sees on every patient list.

Anvesh Narimiti, a hospitalist and CI fellow, published a LinkedIn carousel this week coining the phrase predictive decoration — any risk score surfaced in the EHR without a paired decision support path. Epic’s 30-day readmission risk column is his canonical example: it pages case management and checks a box, but does not change a single medication, stay length, or follow-up interval for the acute-stay hospitalist.

Jennifer Goldman (CMIO at Memorial Healthcare) steelmanned the counterpoint in the comments: the same score, read by the primary-care team running express-lane follow-ups, does drive behavior. The score isn’t decoration when a downstream operation is designed around it — but the hospitalist seat may not be the seat at which to evaluate it.

😤 Haters

“So the score is useful, just not to the hospitalist? Fine, that’s not new.” The naming is new, and naming shapes what you build. A score is decoration until a specific seat in the workflow acts on it. The builder question — and the implementation question — is: for whom is this score actionable, and what is the action? If the answer is “case management gets paged,” the score isn’t supporting the hospitalist; it’s supporting case management’s operating model. Design around that.

“This is just Graham Walker’s ‘data slop → model slop → publication slop’ at a different layer.” It’s the downstream half of the same story. Walker’s piece was about a single bad Kaggle dataset generating 124 papers and 1,500 citations; Narimiti’s is about those models becoming a column on a patient list that doesn’t change anything. Upstream: the data doesn’t support the model. Downstream: the model doesn’t support the decision. The full pipeline is the failure.

💡 80/20: A score that doesn’t change a decision you’d make differently is decoration, not analytics. Try: for every AI-generated score, column, or alert you are about to build into a workflow, write the single sentence “If I see X, I will do Y instead of Z.” If you can’t write it, build the action first and the score second.

Cloudflare ran 131,246 AI code reviews in one month. The break-glass override rate was 0.6%.

Cloudflare Engineering published the numbers this week: a custom multi-agent code review system on top of OpenCode, seven specialized agents per merge request (security, performance, code quality, etc.), median 3m 39s per review, $1.19 average cost, 85.7% cache hit rate across 120 billion tokens. Most importantly: engineers overrode the AI reviewer on only 0.6% of merge requests.

😤 Haters

“Code review is not clinical decision support.” It’s not. But the 0.6% override rate is the closest public datapoint for “what does mature agentic adoption look like in a regulated production workflow” that I’ve seen. The structure — multiple specialized agents, per-task cost budget, circuit breakers, override logged and measurable — is the structure a serious clinical-AI deployment needs. Healthcare has no equivalent published metric. That’s a gap.

“Measuring override rate is the wrong metric for medicine.” Partially fair — in medicine the consequence of not overriding can be the harm itself, whereas in code review the CI pipeline catches a lot. But not having an override metric at all is worse than having the wrong one. Build the override telemetry before you ship the agent.

💡 80/20: Override rate is the single most useful trust metric for an agent in production, and almost no health-AI tool publishes one. Try: in the first week after any AI tool you build goes live, instrument override tracking — per-user, per-decision, with a free-text reason. You’ll learn more from the first 100 overrides than from any pre-launch benchmark.

🧰 Builder’s Tip

Tool Spotlight — anthropics/healthcare: the Claude Code healthcare marketplace that has been sitting there since January.

[Here today because of today’s anthropic healthcare webinar.] Anthropic published an official healthcare plugin marketplace at github.com/anthropics/healthcare at JPM26 on January 9, 2026. v1.0.0, no commits since. It ships three Agent Skills — fhir-developer@healthcare (FHIR R4, LOINC, SNOMED CT, RxNorm patterns), prior-auth-review@healthcare (NPI / ICD-10 / CMS Coverage / CPT checks + medical-necessity summarization), and clinical-trial-protocol@healthcare (FDA/NIH-compliant protocol scaffolding) — plus four remote MCP servers covering CMS Coverage, the NPI Registry, PubMed, and ICD-10 codes.

Install the marketplace and a skill in two commands inside Claude Code:

/plugin marketplace add anthropics/healthcare
/plugin install fhir-developer@healthcare

You now have a FHIR-aware assistant with live access to CMS Coverage and NPI Registry and no API keys to set up. That is a real starter kit for a weekend prototype — synthetic patient data via Synthea, a FHIR-aware skill, and MCP access to coverage and identifier lookup, all free, all local to your Claude Code session.

The quiet tell: three months of community silence. The first community PR against anthropics/healthcare — a HIPAA audit skill, a medical-coding abstraction skill, a discharge-summary-extraction skill — lands on the canonical surface with almost zero noise. If you’ve been looking for a concrete starting point for agentic clinical tooling, this is it.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Scrub Capital is looking for scouts. If you’re a clinician-builder, this is how you get into VC without leaving the bedside.

Scrub Capital — the physician-led health tech venture fund that closed Fund 1 oversubscribed beyond $10M in Q1 2026 — is starting a scout program ideal for clinician builders. This is not your average, fund: notably Jon Slotkin, one of the three GPs of the fund, is a neurosurgeon and Chief Medical Officer for Strategy and Growth at Geisinger.

The scout opportunity: bring deal flow from networks the fund can’t access on its own. If you’re a clinician who builds things and also notices what other clinicians are building, this is the intersection. You don’t need to leave your clinical job. You don’t need to raise a fund. You need a thesis, a network, and the ability to spot the clinician-founder who’s three months from quitting their job to build full-time.

I often find you might be able to do more good at scale helping others than starting your own thing. Most clinician-builders think the path to VC is “build a company, get funded, exit, then become an investor.” The scout model skips the exit. You invest your attention and network access, not your capital. And the fund gets something no traditional VC scout can offer — clinical pattern recognition on whether a product actually works at 2 AM.

💡 80/20: If you’ve been building in a clinical niche and you keep seeing the same problems that nobody is solving well, write a one-page thesis: “I believe [category] is under-invested because [specific clinical insight that non-clinicians miss].” That’s the pitch for a scout conversation. The bar isn’t deal volume — it’s deal quality, and you bring the quality by being a subject matter expert.

The ambient scribe is dead. What replaced it is much bigger.

Two things happened in the same week that, taken together, end the ambient scribe as a standalone product category.

First: Abridge announced multi-year content partnerships with NEJM and JAMA. Clinicians using Abridge can now query peer-reviewed evidence — grounded in both published research and the patient sitting in front of them — during the clinical encounter. This is not a documentation tool with a search bar bolted on. This is a clinical decision support system that happens to also write the note. Abridge serves 250 health systems and will support more than 100 million patient-clinician conversations this year.

Second: Ambience Healthcare unveiled a multiyear platform roadmap at its inaugural Apex Summit, expanding from ambient documentation into five domains: clinical workflows, revenue integrity, patient engagement, care orchestration, and clinical research. They previewed “Kait,” an AI patient agent for between-visit care management, and announced performance-based coding contracts that tie their compensation directly to revenue outcomes. Ambience claims 80%+ clinician utilization and NPS above 60.

Read those two moves together. Abridge pivoted up the clinical value chain — from documentation to decision support. Ambience pivoted across the operational value chain — from documentation to everything. Neither company is building a scribe anymore. They are building the operating system layer that sits between the clinician and every other system in healthcare.

The ambient scribe was always a wedge. The encounter is the richest structured data source in medicine — every diagnosis considered, every medication discussed, every referral weighed. The companies that captured it first had the cheapest entry point into clinical workflows. Now they’re using that position to expand into the things that actually make money: coding accuracy, CDS, care coordination, and longitudinal patient management.

😤 Haters

“Scribes still save time — a large study of 1,800 clinicians found 16 minutes saved per 8-hour shift.” True. And modest. One extra patient every two weeks. The time savings alone never justified the vendor price, which is why every ambient vendor is now pivoting to revenue capture. The ABA published a policy brief calling it the “coding arms race” — ambient tools that up-code documentation are the real business model, not time savings. If you’re building an ambient tool and your pitch is still “saves clinicians time,” your competitor’s pitch is “pays for itself through coding uplift.” You’re in the wrong fight.

“These are two companies. The market is bigger than Abridge and Ambience.” It is. And the market just watched the two leaders publicly abandon the category name. Nuance/Microsoft is already there with DAX Copilot inside Epic. Google is pushing Med-PaLM into documentation-and-beyond. The standalone scribe companies — the ones that only write notes — now have exactly zero room. The feature is commoditized. The platform race is on.

“CDS is a graveyard. Abridge won’t succeed where UpToDate alerts failed for 20 years.” Maybe. But the difference is surface area. UpToDate fires alerts that clinicians dismiss. Abridge is inside the conversation — the CDS is grounded in what the patient actually said, not a generic rule. Whether that’s enough to overcome alert fatigue is an open question. But the integration of NEJM and JAMA citations into encounter-level context is a genuinely new approach, not a repackaged alert.

💡 80/20: If you’re building anything in the ambient/documentation space, answer this question today: are you a feature or a platform? If your tool only writes notes, you are a feature that Abridge, Ambience, or Nuance will replicate in a quarter. The viable builder positions are: (1) verification — the quality layer that audits what the platform produces, (2) specialty depth — the domain-specific model for derm, psych, or peds that the generalists won’t invest in, or (3) the thing the platform can’t do — patient-facing, out-of-EHR, or cross-system. Pick one.

→ Full write-up

UT Austin is building the first AI-native hospital from scratch. $750M. Opens 2030.
UT Austin announced the UT Dell Medical Center — a greenfield academic medical facility with AI systems embedded throughout its physical infrastructure from day one. $750M from Michael and Susan Dell, plus $100M from the Coxe family. The CIO, Claus Jensen, framed the design question: “What does a hospital look like if you design the entire clinical, operational, and physical system around intelligence from day one?” The answer includes an Intelligence Performance Center as the hospital’s operational brain, Living Digital Twins predicting patient deterioration 24-72 hours ahead, environmental sensors, integrated robotics, and full MD Anderson Cancer Center integration. Target: top-10 national ranking within a decade.

😤 Haters

“Every new hospital says it’s going to be ‘smart’ and then opens with the same Epic install as everyone else.” Fair historical pattern. The difference here is timing and funding. $850M in committed philanthropy, a greenfield site with no legacy infrastructure to accommodate, and a 2030 opening date that lands squarely in the window where clinical AI tools are mature enough to embed structurally rather than bolt on. The question isn’t whether they’ll have AI — it’s whether designing around it from the start produces a measurably different care delivery model. We’ll know in four years.

“This is an academic hospital project — clinician-builders should care about startups, not health-system construction.” Disagree. This is a four-year product lab with $850M in funding, a mandate to build differently, and a hiring pipeline that will need clinical informaticists, AI product managers, and clinician-developers who understand both the workflow and the toolchain. If you want to build clinical AI inside a system that’s designed to use it — not fight it — this is the rare case where the health-system path might be better than the startup path.

💡 80/20: Watch the UT Dell Medical Center hiring pipeline over the next 12 months. Greenfield AI-native builds need clinical informatics, AI product, and developer roles that don’t exist at most health systems. If you’re a clinician-builder considering the health-system track, this is the profile: a system that was designed for your skillset, not one that’s trying to retrofit it.

Atropos Health launched a 33-million-artifact evidence library that outperformed GPT-5.4 and Claude Opus on clinical questions.
Atropos Health launched Alexandria — a precision Real-World Evidence library with 33 million pRWE artifacts at launch, targeting 2 billion by end of 2026. On 5,000+ real clinician questions, it outperformed GPT-5.4, Claude 4.6 Opus, Llama 4, and Gemini 3.1 Pro by 2-3x. The distribution play is the real story: “First Edition” partnerships with Meta, Microsoft, Heidi Health, Vim, Avo, and others give Alexandria workflow reach to roughly one-third of U.S. physicians and half of major health systems.

😤 Haters

“Real-world evidence libraries are not new. UpToDate, DynaMed, and ClinicalKey have existed for decades.” The evidence format is new. These aren’t curated editorial summaries — they’re machine-generated pRWE artifacts from claims and EHR data. The 2-3x performance claim over foundation models on clinical questions is bold and needs independent validation, but if it holds, it means domain-specific evidence retrieval still beats general-purpose LLMs on clinical accuracy. That matters for every builder deciding between “fine-tune a foundation model” and “build a retrieval layer.”

“Meta and Microsoft as distribution partners sounds impressive but means nothing for clinical adoption.” It means the evidence layer is going to show up inside the ambient scribes and clinical copilots that clinicians are already using. That’s not a search bar — that’s embedded CDS. Whether clinicians trust it is a different question, but the distribution problem is solved.

💡 80/20: If you’re building clinical AI and your evidence grounding strategy is “prompt the foundation model and hope it’s right,” Alexandria just raised the bar. The builder question is: do you build your own retrieval layer, license Atropos, or accept the hallucination risk? For most small teams, licensing a purpose-built evidence layer will be cheaper and more defensible than building one.

⚡ From the Wire

“Healthcare implemented ambient AI backwards.” Angel Mena, MD (CMO of Symplr) said it plainly on This Week Health this week: health systems layered AI onto broken processes without fixing them first. The conversation covers ambient documentation, quality metrics, and the credentialing chaos hiding inside every health system. For builders: if you’re automating a workflow, the first question is whether the workflow itself is broken. Ambient AI on a broken credentialing process doesn’t produce better credentialing — it produces faster bad credentialing.

The clinical history IS the prompt. A physician just said it out loud.

Adam Carewe, MD posted an essay yesterday titled “The Vibe Coding Gap Is Not a Coding Problem”. His core move: line up the clinical history (chief complaint, HPI, PMH, social history, ROS) against the production vibe-coding discipline Anthropic’s Eric describes — “spend 15 to 20 minutes collecting context before writing a single prompt. Explore the system. Build the plan. Name the constraints. Then let the model execute.” The two disciplines are the same discipline. The physician has been running it since internship.

His closing line deserves to be pinned above every clinician-builder’s desk: “System B required us to outsource building to engineers because they held the only skill that could produce software. System C makes domain knowledge the scarce resource and code the commodity.”

This is the thesis of this newsletter in someone else’s handwriting. It’s also a clinical informaticist who saw patients this week saying it cleanly — not a founder deck, not a LinkedIn think piece. When a physician builds the argument, the argument is built. The only thing left to do is build the thing.

😤 Haters

“This is just the same ‘doctors should code’ talk we’ve had for ten years — dressed up in vibe-coding lingo.” Partially fair. The rhetoric isn’t new. What’s new is the economics. Ten years ago, a physician who learned React had bought themselves a six-month path to a mediocre prototype. In 2026, that same physician, armed with the clinical history reflex, can be in a working agentic prototype by Sunday night. The new part isn’t the encouragement; it’s that the cost structure finally matches the encouragement.

“Most doctors I know can’t actually model a software system — they can model a patient.” Correct, but you’ve conceded the point. Modeling a patient IS modeling a system — the inputs are messy, unreliable, adversarial. The output is life-changing. The feedback loop is 48 to 72 hours. If you can disposition a complex polypharmacy patient with an occult sepsis workup, you already have the cognitive substrate for “this tool returns 200 but silently drops the field.” The translation is smaller than it looks.

“The physicians ACTUALLY building are a rounding error — stop pretending they’re a wave.” Also fair. Fewer than 1% of practicing US physicians have a repo they’ve committed to in the last year. But a wave doesn’t need 50%. It needs maybe 2%. When 30,000 of the 1.1M clinicians in the US start shipping specialty-specific tools on agentic stacks, the market looks nothing like it did. We are not there yet. We are also not far.

💡 80/20: The skill your residency drilled into you — do not order the CT without the full history — is now directly monetizable in a market that didn’t know it needed it. Try: write your next LLM-backed tool the way you write an H&P. Chief complaint (one sentence: what does this tool do?). HPI (user story, onset, quality, aggravating/relieving factors). PMH (systems it touches — EHR, scheduler, lab). Social (who uses this and on which shift). ROS (ten-system review of what else could break). Then — and only then — open the editor.

📡 Builder’s Radar

Vertical EHRs are about to get un-seated.
Sam Toole’s two-part series argues that specialty EHRs (think ModMed, NextGen, AdvancedMD) sit on a pile of clinical context and charge per seat — a pricing model that dramatically underprices what the software is positioned to deliver. AI-native entrants can ship better software at lower prices and add consumption-based revenue streams the incumbents can’t match. The PE thesis behind Thoma Bravo-NextGen, Warburg-ModMed, and Francisco Partners-AdvancedMD assumes sticky seat-based contracts. That thesis has a new expiration date.

😤 Haters

“Nobody is actually ripping out their ophthalmology or derm EHR this year — switching costs are brutal.” True, and that’s exactly what makes the window long and the outcome slow-motion obvious. The PE lockup periods outlive the buyer’s confidence. The installed base doesn’t need to switch this quarter; it needs to hear about a cheaper, better-at-documentation competitor and stop paying list price at renewal.

“Consumption pricing in clinical software has been tried and doesn’t work — providers hate variable cost.” Fair in 2019. In 2026 providers have already accepted scribe pricing per encounter, coder AI priced per chart, and prior-auth bots priced per submission. The variable-cost muscle is built. The seat model is the anomaly now, not the default.

💡 80/20: If you’re a physician owner-operator in a fragmented specialty, you are the ideal first customer of an AI-native vertical EHR — and probably also the ideal founder of one. Reframe: the PE-owned incumbent isn’t a competitor, it’s your TAM. Start by auditing what yours bills for per seat vs what it actually does per encounter.

Anthem’s attempt to relitigate its IDR losses got thrown out.
On April 9, Magistrate Judge Karen Scott (C.D. Cal.) dismissed Anthem’s federal RICO, ERISA, and fraud claims against HaloMD and co-defendants — the billing company that helps providers navigate the No Surprises Act Independent Dispute Resolution process. Anthem argued HaloMD flooded IDR with ineligible disputes labeled as eligible. Judge Scott ruled NSA only permits federal review under narrow circumstances (corruption, bribery, destroyed evidence) and that the IDR process already lets participants flag ineligible disputes to the arbitrator. Adopting Anthem’s position, she wrote, would subject “nearly every eligibility determination” to federal court review. Anthem will appeal to the Ninth Circuit. Three parallel BCBS suits remain active in Texas, Georgia, Ohio.

😤 Haters

“This is a minor procedural ruling — insurers will just find another angle.” Partially. But the phrase “procedural ruling” is doing a lot of work. The federal claims — the ones with treble damages and nationwide preemption potential — are gone. What remains is state-law claims that can be refiled in CA state court, where insurers have much less leverage. The fee-structure argument (”arbitrators have bad incentives”) also got explicitly rejected on merits, which is not a procedural loss.

“Builders don’t care about an NSA case.” Builders who touch out-of-network billing care a lot. If you’re shipping anything that routes claims, drafts appeals, or helps providers survive the IDR flood, the legal substrate you’re building on just got considerably firmer. The ruling doesn’t create the opportunity — it protects the existing one.

💡 80/20: If your product touches the IDR process — appeal drafting, QPA calculation, dispute filing — read the ruling itself, not the summary. The judge’s framing of what “bad faith” looks like under NSA will shape vendor contracts for the next 24 months. Try: run your own product’s IDR filings against Scott’s corruption/bribery/destroyed-evidence bar. If your software couldn’t survive that bar on its worst day, harden it now.

Behavioral health visits passed primary care in 2024. Nobody is talking about it.
CDC Data Brief 558 on sources of usual care — re-surfaced this weekend in Jared Dashevsky, MD’s “Where Americans Get Care in 2026” — landed a stat worth stopping on: 66.4 million behavioral health visits vs 62.8 million primary care visits in 2024. For the first time. 90% of adults still say they have a usual source of care, but the age split is stark — 12.2% of 18-34 year olds name urgent care or a grocery-store clinic as their usual source; only 4% of 65+ do. 80% of practicing physicians now work for hospital systems or corporate entities (2023 data). Urgent care centers nearly tripled since 2010 while PCP density dropped 22%.

😤 Haters

“Urgent care replacing PCP isn’t news — everyone has been predicting this since 2015.” Predictions aren’t data. The stat that flipped is data. The moment behavioral health visits exceeded primary care visits is the moment “primary care” stopped being the default front door of American medicine. If you’re building on assumptions from 2019, that assumption is now wrong.

“Agentic AI to ‘expand PCP panel capacity’ is the Dashevsky pitch and it’s basically marketing for his favored vendor.” Somewhat. He’s transparent about which product he’s bullish on. But the diagnosis (supply won’t move; AI panel expansion is the near-term lever) is correct independent of whose tool wins. The bottleneck isn’t model quality — it’s continuity, documentation, and the longitudinal relationship, which urgent care structurally cannot provide.

💡 80/20: The real building opportunity isn’t “replace the PCP”; it’s “be the longitudinal layer that makes urgent care visits count.” Every 18-34 year old who walks into a CVS MinuteClinic is producing a data event that currently disappears into a silo. Try: spend a weekend mapping what it would take to stitch one urgent-care visit into a longitudinal record that a PCP could act on later. That’s the build.

OpenAI shipped a frontier reasoning model for biology. You weren’t on the invite list.
OpenAI introduced GPT-Rosalind — its first frontier reasoning model purpose-built for the life sciences. Deployed under trusted-access terms to Moderna, Amgen, the Allen Institute, and Thermo Fisher. This is a peer story to Anthropic’s Project Glasswing / Mythos cyber-research model — both labs are now shipping narrowed-access frontier models to strategic partners in regulated domains, weeks apart.

😤 Haters

“’Trusted access’ is paperwork. Same model, fancier paperwork.” Partially true — the weights are the weights. But the paperwork is the point. Moderna and Amgen are getting a model-use arrangement that small clinical AI startups cannot match, and the regulatory comfort is real: shared liability, audit trails, indemnification. If you’re a small shop trying to sell bio reasoning into enterprise pharma, the bar just moved.

“There’s no way a bench scientist at Moderna is going to trust LLM reasoning on a drug target over a PhD chemist.” Not today. The observable effect over the next year is not replacement; it’s asymmetry — teams with GPT-Rosalind triage faster, read more papers, propose broader mechanism hypotheses. A group that’s 30% faster at literature-to-hypothesis doesn’t “replace” anyone. It just leaves the teams without it behind in 18 months.

💡 80/20: For clinician-builders outside of big pharma, the read is about tiers of access, not the model itself. Frontier labs are now shipping regulated-domain models with restricted rollouts. Reframe: “frontier model access” is becoming a commercial asset, not a commodity, in biotech/health. Plan for a world where your local-first MedGemma pipeline is the open substrate and the frontier-model calls are a per-query premium paid only when warranted.

🛠️ From the Workbench

isitagentready.com — see if your site speaks agent.
Cloudflare launched a public scoring tool that audits your website for agent compatibility: robots.txt posture, sitemap, MCP discoverability, Agent Skills, OAuth, access control. Drop in a URL, get a score and an actionable fix list. Paired with Radar data showing current web-wide adoption is low. Useful as a fast audit when you’re about to stand up something agent-facing.

⚠️ Verify: This is an uptime/config audit, not a security audit. A high Agent Readiness Score says nothing about whether your app correctly handles deep-merge updates, hallucinated success responses, or prompt-injection through fetched content. Run it for the hygiene layer — do not treat the score as an endorsement.

😤 Haters

“Why do I care about my clinic site being ‘agent ready’? Patients aren’t using agents yet.” Some aren’t. Many are — just not the way you think. Copilot, Perplexity, ChatGPT search, and agentic browsers already fetch doctor pages when patients ask “what should I expect at my derm appointment.” If those fetches return garbage, your office gets skipped. The cost of fixing it (robots.txt, a sitemap, an MCP endpoint for hours/location) is hours, not weeks.

“Cloudflare is grading its own homework — of course every site needs their products.” Partly true. But the open-ended checks (sitemap, robots, a clean API endpoint, OAuth done right) don’t require Cloudflare. They just require you to not be hostile to crawlers that don’t have a 2015 UX.

💡 80/20: Treat agent-readiness the way you treated HIPAA awareness in 2015 — not glamorous, but the building block everything else sits on. Try: run isitagentready.com on your own clinic or product site today. Fix the top 2 items before your next deploy.

🧰 Builder’s Tip

Weekend Project — Build an LLM tool-call safety wrapper (3-4 hours).

This weekend’s build is inspired by Orqisai’s “three hardest problems” post — the three production failure modes that bit every builder who shipped an agent against a real API: (1) LLMs silently PUT partial payloads and wipe your fields; (2) LLMs report “success” on 404s; (3) query params arrive as strings instead of dicts and the request just quietly returns nothing.

The weekend project: write a minimal safe_call(tool_fn, args) wrapper in Python (or TypeScript) that does three things — fetch current state and deep-merge before any write, prefix every non-2xx response with the literal string Error:, and coerce string-shaped query-param inputs into dicts with a logged warning. Point it at a FHIR sandbox (Medplum, Inferno, or the HAPI public server) and run 10 test cases where your agent “updates a Patient resource” while only mentioning the one field it wants to change. Confirm the other fields survive. Starter test: ask the agent to update Patient.gender and verify Patient.birthDate is still there afterward.

You end Sunday with a wrapper you can drop into every subsequent agent project and not re-learn these three lessons the expensive way.

What are you building this week? Reply and tell me — I read every one.

— Kevin

124 Clinical Prediction Papers Were Built on Kaggle Datasets That Nobody Actually Checked

Graham Walker, MD posted a thread on Saturday that deserves to rearrange how you look at clinical ML literature. Australian researchers investigated two Kaggle datasets widely used in “medical AI” work: a stroke prediction dataset and a diabetes dataset that together have more than 400,000 downloads. Neither has verifiable provenance. The stroke dataset’s uploader explicitly notes it should not be used for research. The diabetes dataset has 7 percent duplicate rows and only 18 distinct HbA1c values across a file claiming 100,000 patients — a distribution that is clinically impossible unless the “data” was generated, not collected.

Those two files have seeded at least 124 published clinical prediction papers, roughly 1,500 citations, and one medical device patent co-held by USC and Caltech. Walker frames it as literature laundering: once a sketchy source is cited inside a review article, it stops being sketchy. It becomes “the literature.” Every gatekeeping layer — the repository, the modeler, the peer reviewer, the review-article author — assumed the previous layer had done the due diligence. None of them had.

Walker and Joseph Habboushe, MD MBA — the MDCalc co-founders — are keynoting the Coalition for Health AI summit on this. They are the right messengers because they already do this job for a living: they are clinicians who read the files.

😤 Haters

“This is a peer review problem. It is not an AI problem.” The reviewer failure is real, but peer review is a six-week human process. Model training is a six-hour script. When a sketchy dataset gets cited five times in good journals, it becomes part of the training corpus for the next generation of clinical LLMs — not because the model chose to trust it, but because the citation graph did. This is how bad provenance becomes bad weights.

“Every field has junk datasets. The signal comes through.” In most fields, junk data produces models that fail in obvious ways. In medicine, a model trained on 100,000 synthetic-looking patients can still return clinically plausible predictions — because the distribution was designed to look like clinical data. The failure mode is invisible until someone deploys it. That is the one field where “it looks right” is not enough.

“OK, but this is someone else’s mess. What am I supposed to do with it?” If you are a clinician-builder, this is the job description. The single most common failure mode in clinical ML is not the model architecture. It is the stuff that happened before the first line of Python got written. The people best-positioned to audit that are the ones who already read labs, charts, and imaging reports skeptically for a living. If the CHAI keynote surfaces one takeaway, let it be this: the profession that catches 18 unique HbA1c values across 100,000 rows in 30 seconds is the one sitting at the point of care.

💡 80/20: Before you train, fine-tune, or even cite a clinical dataset, open the CSV. Histogram the key continuous variables. Check cardinality. If a dataset that claims to represent 100,000 patients has 18 distinct values for HbA1c or 12 distinct ages, it did not come from patients. Try: pick one “medical” Kaggle dataset you’ve seen cited in a paper this year, run df.nunique() and df.describe(), and decide in five minutes whether the distribution is clinically possible. That is the audit that should have existed upstream.

📡 Builder’s Radar

Ambience Drew Its Real Map — and Ambient Scribe Is Just One of Five Squares

Ambience Healthcare used its Apex Summit this week to publish a five-domain roadmap: Clinical Workflows, Revenue Cycle/Integrity, Patient Experience, Care Orchestration, and Clinical Research. Point-of-care coding moves from retrospective to real-time, now offered on performance-based contracts where Ambience shares financial risk on coding accuracy. A new patient agent called Kait runs between visits. “Reasoning traces” from clinical encounters get pitched as structured data for computational phenotyping. The company also published enterprise metrics most AI vendors won’t: more than 80 percent clinician utilization, NPS above 60, and a claimed 3:1 operating-margin ROI.

😤 Haters

“Every scribe vendor is announcing a platform. This is a pivot deck.” It is a platform announcement from a company that still sells a scribe — so yes, there is a narrative shift. But performance-based coding contracts are a real commercial commitment. A vendor willing to put its fee at risk on coding accuracy is telling you something different than a vendor who ships a slide deck. The contract model is the part worth watching.

“Reasoning traces as a research substrate sounds like a privacy bomb.” It can be. De-identification at the encounter-narrative level is famously hard — the text itself is the identifier. If Ambience is positioning these traces as research infrastructure, the BAA and data-use terms matter more than the feature announcement. Ask how reasoning traces flow back to the research pipeline before any of this reaches a real chart.

💡 80/20: The scribe is not the product anymore. It is the wedge that gets the contract. Reframe: when evaluating any ambient vendor this year, look past the transcript and ask what the company has committed to ship in the other four domains — and what it will put at risk financially on each.

Hippocratic Put a Voice Agent in the Hallway and Another One on the Phone

Hippocratic AI launched two voice products this week. AI Front Door is a cross-channel patient agent — phone, text, app — that holds longitudinal patient memory across scheduling, billing, and care coordination. Nurse Co-Pilot sits on the inpatient floor, handling admit and discharge education, medication teach-back, and caregiver engagement. Both products were co-developed with Cleveland Clinic, OhioHealth, and Cincinnati Children’s. The pitch: 1 to 4 hours returned per nursing shift.

😤 Haters

“Voice AI in a hospital is a liability magnet. One wrong med-adherence answer and it’s national news.” Not unreasonable. But Hippocratic’s framing — explicit human-clinician checkpoints and EHR documentation of the voice interaction — is the right shape for inpatient deployment. The question is not whether the voice agent is safe in a vacuum; it is whether the checkpoint is enforced in practice when the nurse is busy.

“’1 to 4 hours per shift’ is a vendor claim, not a study.” Correct. The number is the ceiling, not the median. The useful framing is: does the voice agent reduce the specific burdensome tasks your nurses already flag, or does it move time around without reducing it? A 30-minute time-motion study on two shifts will tell you more than the press release.

💡 80/20: The patient call center and the inpatient education binder are both about to become agent-addressable surfaces. Try: before buying into any voice product, sit at the nursing station for an hour and log which tasks the existing staff actively want handed off. The answer is usually three specific ones, not “everything.”

Keebler Health Raised $16M to Build Risk Adjustment That Was Born After Transformers

Keebler Health closed a $16M Series A led by Flare Capital with Sands Capital participating, total raised $23M since 2023. The pitch: the risk-adjustment category was built on legacy NLP that retrofitted itself onto LLMs. Keebler was built LLM-native from day one. The claimed opening: roughly 80 percent of the clinical data that matters for HCC coding is unstructured — prior notes, scanned reports, specialist letters — and only 59.4 percent of chronic conditions get consistently captured across EHR sources. Point-of-care insights sit inside the existing workflow rather than as a retrospective chart pull. RADV audit readiness is the near-term roadmap as CMS sharpens scrutiny.

😤 Haters

“’LLM-native’ is a marketing word. The wrappers all look the same from outside.” True if you are skimming the homepage. Less true when you read the architecture. The legacy-NLP vendors have model chains, ontology lookups, and rule engines glued together over a decade. An LLM-native stack collapses most of that into retrieval plus a single reasoning pass — which is cheaper, faster to iterate, and weirdly easier to audit. The moat argument cuts both ways, though: if it’s easy for Keebler to build, it’s easy for the incumbents to rebuild.

“Risk adjustment is under CMS audit pressure for over-coding, and the whole pitch is ‘find more HCCs.’” This is the sharpest objection. The thing that makes this category interesting also makes it dangerous. If the RADV audit readiness story is a real product — not just a bolt-on feature — Keebler has to show that the HCCs the model surfaces are defensible in a chart-level audit. Otherwise it is just a faster way to book revenue a payer will claw back.

💡 80/20: The ground-up-LLM positioning is going to eat a lot of legacy NLP in clinical revenue cycle this year. Reframe: when your health system’s RCM vendor pitches its “new AI features,” ask whether the underlying data pipeline was rebuilt or whether a transformer was bolted onto the old one. The answer predicts the product’s ceiling.

🧰 Builder’s Tip

Mindset: The Audit Is the Product

Every clinician-builder I know spends the first month of any project wishing they could skip the boring part. Read the CSV. Histogram the continuous variables. Pull 20 charts and compare to what patients actually report. Spend a Saturday reading your vendor’s BAA.

Skip it, and you get the 124-papers problem — clean-looking artifacts built on a foundation nobody checked.

Don’t skip it. The audit is not the overhead. The audit is the defensible part. A dashboard, a model, a clinical workflow — any of these can be cloned in a weekend by someone with a better LLM and a faster GPU. What cannot be cloned is the clinician who read the files, asked the right question, and wrote down why the numbers were wrong.

That is why domain expertise is the scarce input. Not because the coding is hard. Because the skepticism is.

If you want one habit to take into the week: open the data before you open the notebook. Distributions first. Models second. If the histogram looks impossible, the model will be impossible too — it just won’t tell you.

What are you building this week? Reply and tell me — I read every one.

— Kevin

The decades disappear like sinking ships.

You swear you were just in your twenties. You swear the pandemic was yesterday. Then you look up, and a decade is gone.

Every clinician I know has a prototype in their head.

The workflow shortcut you thought of as an intern. The alert you knew was broken as a resident. The discharge script you rewrote in your head on every shift. The handoff template you sketched on a napkin.

They are still in your head.

You said you’d build it after residency.
Residency ended.

You said you’d build it after fellowship.
Fellowship ended.

You said you’d build it when the kids were older.
They’re older.

You said you’d build it when you learned to code.
Someone else shipped it.

You said you’d build it when the EHR opened an API.
The EHR opened an API.

You said you’d build it when AI was good enough.
AI is good enough.

A decade is gone.

Not because you were lazy. Because the friction between your clinical instinct and shipped software was so high that starting felt pointless. Six months to learn React. Another year to learn FHIR. Six more to wrap your head around HIPAA. By the time you could build the thing, your shift had changed, the workflow had moved on, and the frustration had faded.

That friction is gone now.

What took six months takes a weekend.

What took a year takes an afternoon.

What took a decade is sitting in a chat window, waiting for you to describe what you want.

You know the workflow better than anyone on the product team.

You know the edge case because it bit you last Tuesday.

You know the failure mode because you ignored the alert at 2 AM.

That knowledge has always been the hard part. It was never the code.

The decade is going to pass either way.

April becomes April again. And again. You have the same frustrated thought during the same 2 AM handoff with the same broken workflow.

The only question is whether the prototype in your head is also in a repo by then.

What you said you’d build is sitting in your head right now.

It has been for ten years.

Open the laptop.

What are you building this week? Reply and tell me — I read every one.

— Kevin

CMS Just Shipped an App Store for Medicare — and 700 Companies Already Signed On

CMS unveiled the first production rollout of the Medicare App Library this week — a centralized, vetted directory of digital health apps that patients can connect directly to their Medicare data. More than 700 companies have signed onto the voluntary Health Tech Ecosystem initiative, and 50+ apps are either already listed or in the vetting pipeline. Priorities announced: conversational AI, chronic disease management, and eliminating paperwork. Welldoc, Noom, January AI, Flexpa, ZocDoc, b.well, Xealth, Polygon Health, and HealthEx are in the first wave.

The thing to notice: for ten years, FHIR has been plumbing. You fought with it to get a lab value out of Epic. Now CMS is standing up the retail surface on top of that plumbing. Distribution is the thing nobody built for clinician-made apps, and the distribution path just got cut.

😤 Haters

“700 companies signed a pledge. How many apps actually work when grandma tries to connect them?” Probably most will fail the real-world test. The “last mile” of consumer adoption is the thing nobody has cracked, and the vet-and-list approach doesn’t solve it. If the library becomes a directory of dead integrations it’s worse than nothing. But 50 already in the pipeline is not a pledge — it’s shipped code.

“This is just CMS taking credit for work Epic and Apple already did.” Partially true. The networks underneath — TEFCA, MyHealtheData, the Apple Health integrations — were built by other people. What CMS is adding is a discovery surface patients can actually find. That’s the missing piece. You can have the cleanest API in the world and if a 72-year-old can’t find your app, the API is a museum exhibit.

“Voluntary means nothing until there’s a payment code attached.” Right — and there isn’t one yet. But the CPT code for AI cardiac calcium detection landed last week. The direction is clear. Voluntary now, reimbursed soon, and the companies in the first wave get to shape what “shaped by” means.

💡 80/20: The distribution bottleneck for clinician-built tools is starting to lift. If you have been sitting on a tool that could serve a Medicare beneficiary and you thought you’d never find them, re-read that sentence. Try: open the participating-company list, find the nearest non-competitor, and ask how they got vetted. The list of people who know is still small.

→ Full write-up

Joyful Health Raises $22M to Fight Denials — and the RCM AI Wave Keeps Landing

Joyful Health closed a $17M Series A led by CRV (total funding to $22M) to build what it calls “denial intelligence” — claim-level investigation, automated appeal workflows, and recovery across fragmented RCM systems. The company claims to have processed $1.4B in transactions with a 95% recovery rate. Vituity’s investment arm and Commure founder Diede van Lamoen’s Go Global are on the cap table.

😤 Haters

“RCM denial automation is a crowded graveyard. Cedar, Candid, Waystar, Alpha Health, now Joyful — what’s new?” The actual differentiator claimed here is investigation at the claim level, not just pattern-based auto-appeal. Most denial tools play the volume game. If Joyful is genuinely surfacing the payer-specific reason behind a denial rather than regex-matching it, that’s a real wedge. The 95% recovery rate either is or isn’t true — ask for a cohort definition before you believe the number.

“Insurers will just evolve their denial patterns faster than the AI can learn.” Probably, eventually. But the payers’ denial policies move on a quarterly cycle. The model retrains overnight.

💡 80/20: Denial intelligence is one of the cleanest AI-to-dollar loops in all of healthcare — the feedback is fast, the label is obvious (paid/not-paid), and the money is real.

→ Full write-up

DOJ Moves to Bankrupt Zealthy — and the Kyle Robertson Arc Keeps Writing Itself

The Justice Department filed for an asset freeze and receivership against telehealth startup Zealthy and its CEO Kyle Robertson. Allegations include using physicians’ names and DEA numbers to fill thousands of prescriptions those physicians never saw, blocking cancellations, and routing payments through shell companies after losing LegitScript. Robertson was fired from Cerebral in 2022 after a similar Adderall distribution scandal. The DOJ says the remedies it’s seeking may bankrupt the company, and preservation of assets is now the live question.

😤 Haters

“This isn’t about tech — it’s one bad actor.” It’s one bad actor, twice. That’s not noise, that’s a pattern — and the pattern is that the controlled-substance + DTC telehealth stack has structural failure modes that good intentions don’t fix. Every clinician-builder working in that space needs to know what due diligence on a medical director looks like, because “we hired a CMO” is not due diligence.

“This kills telehealth prescribing.” It doesn’t kill it — it kills the part of it that was always going to die. Legitimate telehealth for controlled substances with proper clinician oversight is still a real thing. What’s dying is the growth-hacked, SEO-fronted, physician-name-as-rubber-stamp version. Good.

→ Full write-up

🛠️ From the Workbench

Worki — $2.75M to Map Hospital Workforce at the Task Level

Worki raised $2.75M pre-seed led by Redesign Health and Healthliant Ventures (Tanner Health’s venture arm). The pitch: instead of deploying AI agents against vague job titles, they map the task-level workflows across Workday, UKG, Oracle HR, and ServiceNow to find the places where agents can safely slot in. Deployed at Tanner Health and BJC Healthcare. CTO is ex-Uber/Airbnb AI lead.

😤 Haters

“Another ‘AI for operations’ startup with a vague pitch.” The pitch is actually less vague than most — they are explicitly mapping the before-automation workflow first, then deciding what the agents do. That’s the right order of operations. The wrong order is “here is our agent, now find a workflow for it,” which is 80% of this category.

“Redesign Health companies are studio-built and churn through CEOs.” Also often true. Redesign’s track record is mixed. But Tanner Health co-investing through its venture arm and also running it in production is the kind of signal a pure studio bet usually lacks.

💡 80/20: Before building any agent for a clinical or administrative workflow, map the task-level decomposition first. Who does what, in what system, for how long, producing what artifact. Most agent projects fail here. Try: pick one workflow you want to automate. Write the 15-step decomposition. If you can’t, the agent can’t either.

→ Full write-up

🧰 Builder’s Tip

Tool Spotlight: Synthea — The Synthetic Patient Generator You Should Have Already Installed

If you want to prototype anything on FHIR data without signing a BAA, Synthea is the answer. MITRE-maintained, open source, statistically realistic across 90+ disease modules, outputs FHIR R4 Bundles, C-CDA, or CSV.

Three commands:

git clone https://github.com/synthetichealth/synthea.git
cd synthea
./run_synthea -p 100

You now have 100 synthetic patients with full longitudinal records in output/fhir/. Pipe them into Medplum, your own FHIR server, or a flat DuckDB table. Build the rough prototype on synthetic data. Negotiate the BAA after you have something to show. Nobody builds v0 on real PHI.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Innovaccer Commits $250M to Agentic AI — and Bets the Pricing Model on Actually Delivering

Innovaccer announced a $250 million, three-year investment to expand its agentic AI platform across five healthcare workflow categories: patient access, value-based care, revenue cycle, risk and quality assessment, and utilization management. The platform runs on Innovaccer’s “Gravity” data layer — a unified integration of EHRs, claims, CRM, and finance systems — which enables end-to-end agent workflows that can touch multiple data sources in a single task. Major customers include Kaiser Permanente, Ascension, Trinity Health, and Banner Health.

The money is interesting. The pricing model is the story. Innovaccer is adopting outcome-based pricing: the system charges per successful task completion, not per software license or per seat. CEO Abhinav Shashank put a number on it: “If something is costing you $100 to do from a manual process perspective, we will price it at $20.”

This lands the same week PHTI published a report warning that AI-driven automation is creating “bot wars” — provider AI upcoding documentation while payer AI downcodes and denies, with no net cost reduction for the system. And a Qventus survey of 60+ health system executives found that 70% want a single comprehensive AI partner managing multiple use cases, but only 11% currently have one. Only 4% have achieved scaled AI with measurable outcomes.

😤 Haters

“250M over three years is a marketing number. It’s the R&D budget they were going to spend anyway, repackaged as a commitment.” Probably. Most “X investment in AI” announcements are relabeled operating budgets. But the outcome-based pricing is not a relabeled budget — it’s a structural change to the revenue model. If Innovaccer’s agents don’t complete the prior auth, you don’t pay. That’s a bet, not a press release. Whether the agents are good enough to make that bet profitable is the thing to watch.

“Outcome-based pricing sounds great until you realize ‘successful task completion’ is defined by the vendor, not the customer.” This is the exact right skepticism. If “successful” means “the agent ran to completion” rather than “the prior auth was approved” or “the patient got the referral,” the pricing model is cosmetic. The question for any health system evaluating this: what’s the definition of success, and who adjudicates disputes?

“One platform to rule them all is how Epic became Epic. Innovaccer is trying to build the next lock-in.” Fair. And the Qventus data suggests health systems will walk into that lock-in willingly — 70% actively want a single AI partner. The lesson from the EHR era is that integration convenience creates switching costs. The question is whether this generation of CIOs remembers that lesson or learns it again.

💡 80/20: The PHTI bot-wars critique is the backdrop. Innovaccer’s outcome-based pricing is the response. For clinician-builders, the signal is that “per-seat” AI pricing is going to come under pressure from vendors willing to bet on outcomes. If you’re building an AI tool for a health system, think about what outcome you’d be willing to guarantee — and what that guarantee would cost you if the agent fails. Try: pick one workflow your tool automates. Define “successful completion” from the customer’s perspective, not yours. If you can’t price against that definition, you don’t understand the workflow well enough yet.

→ Full write-up

📡 Builder’s Radar

Abridge Partners with NEJM and JAMA — The Scribe Company Is Now a Clinical Intelligence Platform

Abridge announced multi-year partnerships with NEJM Group and JAMA Network to embed peer-reviewed clinical evidence directly into EHR workflows. The integration lets clinicians ask clinical questions and receive answers grounded in both the medical literature and the patient’s own chart context — without leaving the EHR. This builds on Abridge’s existing clinical decision support partnership with Wolters Kluwer’s UpToDate, which is already live. Abridge now projects supporting over 100 million patient-clinician conversations this year across 250 of the largest US health systems.

😤 Haters

“This is a content licensing deal dressed up as a product launch. UpToDate has been doing CDS for decades.” The product isn’t the content — it’s the context. UpToDate answers generic clinical questions. Abridge answers clinical questions in the context of the specific patient conversation that just happened. That’s the gap between a search engine and a clinical copilot. Whether the context-awareness actually improves clinical decisions is an empirical question nobody has published on yet — but the architecture is different from what existed before.

“Ambient scribing is commoditizing. Abridge is pivoting because the core product is getting squeezed.” Partially true, and that’s exactly the right read. The scribe is becoming table stakes. The valuable layer is what sits on top of the transcript — coding, CDS, quality measures, referral suggestions. Abridge is building that layer. Whether they win it depends on execution, not on the NEJM logo.

💡 80/20: The scribe wars are over. The CDS-on-top-of-the-transcript war is starting. If you’re building anything that consumes clinical conversation data — coding audits, quality reporting, care gap detection — the integration surface is about to get much richer. Try: look at one clinical question you answered today by opening UpToDate in a separate tab. Ask: what if the answer had shown up inside the note, grounded in this patient’s context? That’s what Abridge is building toward.

→ Full write-up

CMS Creates a National Billing Code for AI Cardiac Calcium Detection — The Reimbursement Proof Point Is Here

Bunkerhill Health secured both FDA clearance and a new CMS national billing code for AI algorithms that detect coronary artery and aortic valve calcium on routine contrast-enhanced chest CT scans. Effective April 1, 2026, the billing code creates a reimbursement pathway under the Hospital Outpatient Prospective Payment System (OPPS). The algorithms — developed through a research consortium including UCSF, Emory, and MedStar Health — are the first cleared to analyze calcium on contrast-enhanced CTs, which are already being performed for other reasons. A paired STAT piece frames the scale: 19 million chest CTs are performed annually in the US, and 20-40% of incidental calcium findings currently go unreported.

😤 Haters

“One billing code for one company doesn’t change the FDA-to-reimbursement pipeline for everyone else.” True as a general statement. But the existence proof matters. Every cardiac AI startup pitching investors has been asked, ‘But will CMS pay for it?’ Now there’s a specific answer: this company, this algorithm, this code. The pathway exists. That changes the risk calculus for every subsequent entrant.

“Opportunistic screening on existing CTs sounds efficient until the follow-up costs hit — who pays for the cardiology referral the AI triggered?” The payment question is real and unanswered at scale. The algorithm detects calcium. The health system then has to decide what to do with that information — and the downstream cardiology workup isn’t free. But 20-40% of incidental findings going unreported means patients are walking around with undetected cardiovascular risk right now. The cost question is legitimate. The clinical question already has an answer.

💡 80/20: This is the clearest example yet of an AI algorithm getting both regulatory approval AND a payment mechanism. If you’re building clinical AI, study the Bunkerhill pathway: FDA clearance + CMS billing code + opportunistic screening on existing imaging. The formula is “find the scan that’s already happening, add an analysis layer that catches what humans miss, and get paid for the analysis.” Try: think about what data your patients are already generating — lab panels, imaging, vitals — where a second-pass AI analysis could catch something that’s currently being missed. That’s the Bunkerhill pattern.

→ Full write-up

What are you building this week? Reply and tell me — I read every one.

— Kevin

Glenn Steele ran Geisinger for 14 years — 27,500 employees, 1,800 physicians, one of the most respected integrated health systems in the country. Yesterday he wrote that health systems need to replace “huge numbers” of people with AI within five years. Not a think-tank whitepaper. Not a vendor pitch deck. A former CEO, naming the timeline, in STAT. When the person who managed the workforce says the workforce is the problem, clinician-builders need to decide: are you building the tools that make the transition humane, or waiting to see what happens to your department?

🔬 The Big Thing

Former Geisinger CEO: Health Systems Must Replace “Huge Numbers” of People with AI

Glenn Steele Jr., who led Geisinger Health for 14 years, published an op-ed in STAT arguing that U.S. health systems must deploy autonomous AI to replace large portions of their administrative workforce — and they have about five years to do it. His core observation is structural: at Geisinger, the revenue cycle department that processed bills and reconciled data employed more people than the physician workforce. He says that ratio exists at every health system in America and has gotten dramatically worse over the past two decades.

Steele frames this as survival, not optimization. Hospital margins are thin. Administrative bloat is the structural inefficiency that AI can actually address. He’s not talking about replacing clinicians — he’s talking about the 25,000 non-physician jobs at a system like Geisinger and asking what that org chart looks like in 2031.

😤 Haters

“A retired CEO calling for job cuts from the safety of an op-ed is easy.” It is. Steele isn’t the one who has to manage the layoffs, retrain the workforce, or deal with the union negotiations. But his credibility comes from having run the system. He knows what the P&L looks like. He knows which departments are biggest. The messenger might be comfortable, but the math is real.

“AI can’t replace revenue cycle — it’s too messy, too many edge cases.” Revenue cycle is exactly the kind of work where AI excels: high-volume, rules-based, error-tolerant enough to iterate. Claims processing, coding, denial management — these are pattern-matching problems with structured inputs. The edge cases are real, but the 80% of routine work is automatable now.

“This just means worse patient experience — fewer humans to call, more chatbots.” Maybe. But the current experience isn’t great either. If you’ve ever spent 45 minutes on hold with a billing department, you know the human-staffed version has its own failure modes. The question is whether AI-driven billing is worse than understaffed billing.

💡 80/20: If you’re a clinician-builder, the administrative layer is about to get rebuilt. The tools that bridge the gap — appeal generators, documentation assistants tuned to specific payer requirements, coding audit agents — are the highest-value builds right now. Reframe: the back office isn’t disappearing. It’s becoming software. And software needs builders who understand the clinical context.

📡 Builder’s Radar

Pharma Is Buying the AI Search Bar Your Colleagues Use

Out-of-Pocket’s second installment on pharma advertising through clinical AI tools surfaced a damning historical parallel: Practice Fusion pushed 230 million pharma-sponsored clinical alerts to physicians and paid a $145 million DOJ settlement for it. Now OpenEvidence — used by 757,000+ clinicians — generates revenue through pharma ads served while AI answers load. Gallup data shows a 14-point decline in physician trust since 2021, correlating with increased algorithm adoption in clinical workflows.

😤 Haters

“OpenEvidence is free for doctors — someone has to pay for it.” True. And Google Search is free too. The question isn’t whether ad-supported models exist. It’s whether physicians know the difference between evidence-based results and pharma-influenced results when they’re making prescribing decisions at 2 AM.

“Practice Fusion was different — those were alerts, not search results.” The mechanism differs, the incentive structure is identical: pharma pays to influence clinical decision-making at the point of care. The DOJ didn’t care about the UX pattern. They cared about the influence.

💡 80/20: If you build clinical decision support tools, your business model is a clinical decision. Ad-supported CDS introduces structural conflicts whether you intend it or not. Try: before integrating any “free” clinical AI tool into your workflow, find the revenue model. If it’s pharma-funded, that’s a feature you should disclose to your users.

SimpliFed Raises $10.8M to Build Virtual Maternal Health Platform

SimpliFed raised a $10.8M oversubscribed Series A to expand from virtual lactation support into a full virtual OB provider for low-risk patients. Led by Morningside and Hesperia Capital, with the AHA Social Impact Fund participating. The company is on track to serve 5% of all U.S. births in 2026 and is in-network with most major commercial plans and Medicaid health plans across several states. Clinical outcomes data: 96.6% of patients breastfed at one week versus a national average of 85.7%.

😤 Haters

“$10M is small for a health tech company trying to serve 5% of U.S. births.” It is. But maternal health is capital-efficient when you’re virtual-first and aligned with payer incentives. The real validation isn’t the round size — it’s that commercial and Medicaid plans are already paying.

“Lactation to full OB is a massive scope expansion.” It is a leap. But the patient relationship already exists, the payer contracts are in place, and the clinical team understands the population. The risk is execution, not market fit.

💡 80/20: Maternal health remains one of the most underbuilt categories in health tech despite clear clinical need and payer willingness to pay. If you’re a clinician-builder looking for a vertical, maternal and postpartum care has fewer incumbents and more measurable outcomes than most. Try: look at your own health system’s maternal readmission rates and ask what a virtual touchpoint at day 3 postpartum would cost to build.

NPR: AI Is Arriving in Mental Health — and Therapists Are Striking Over It

NPR reported that Kaiser Permanente is evaluating Limbic, a U.K.-based AI company that builds chatbots trained on cognitive behavioral therapy. This comes after 2,400 Kaiser mental health providers in Northern California went on a 24-hour strike, with AI-driven workflow changes as a key issue. Clinicians report that triage screening — previously a 10-15 minute encounter with a licensed clinician — is now conducted by unlicensed operators following scripts.

😤 Haters

“Kaiser therapists are striking about staffing, not AI specifically.” The strike is about both. The AI evaluation is happening in the context of chronic understaffing. When you don’t have enough therapists and you introduce an AI tool that can do intake, the workforce reads that as replacement, not augmentation. Context matters.

“Limbic is a U.K. company — this isn’t a U.S. story.” Kaiser evaluating it makes it a U.S. story. The largest integrated health system in the country is assessing whether AI chatbots can handle mental health intake. What Kaiser does, others follow.

💡 80/20: The mental health workforce shortage is real — and AI tools that extend clinician capacity (not replace clinician judgment) are the builds that will survive the political and regulatory scrutiny. Reframe: if you’re building mental health AI, design it so the therapist is visibly in the loop. The tool that makes the clinician faster wins. The tool that makes the clinician invisible loses.

🧰 Builder’s Tip

Mindset / Strategy: Build the Transition, Not the Replacement

Glenn Steele says health systems need to replace huge numbers of administrative jobs with AI. UnitedHealth is spending $3 billion to do it. But here’s what neither of them is building: the transition layer. The tools that help a revenue cycle specialist become an AI operations manager. The dashboards that let a billing team audit what the algorithm decided. The training interfaces that bridge “I used to do this manually” to “I now supervise the agent that does it.”

The highest-value builds in the next two years won’t replace humans — they’ll translate between humans and the AI that just took their workflow. If you’re a clinician-builder, that’s your edge. You understand the workflow the AI is automating. Build the tool that makes the handoff safe.

What are you building this week? Reply and tell me — I read every one.

— Kevin

UnitedHealth Group Is Spending $3 Billion on AI — and Avery Is Already Live

UnitedHealth Group is deploying $3 billion into AI across its operations, making it one of the largest AI investments in healthcare to date. The company employs 22,000 software engineers globally, with over 80% now using AI to write code or build agents. The first major product: Avery, a generative AI companion for members that’s agentic, HIPAA-compliant, and already live for 6.5 million employer-sponsored plan members and 160,000 Medicare Advantage members — scaling to 20.5 million by year-end.

Avery isn’t a chatbot that answers FAQs. It learns from member interactions and integrates into insurance workflows: claims status, benefits navigation, prior authorization support. Optum Insight CEO Sandeep Dadlani framed the goal as replacing human-driven processes in claims processing, billing code selection, and fraud detection with AI-driven algorithms.

😤 Haters

“UnitedHealth automating claims decisions with AI is terrifying — they already deny too many claims.” This is the right concern. UHG has faced lawsuits over algorithmic denials before. An AI that processes claims faster isn’t inherently better for patients — it’s faster. The question is whether Avery-style tools increase transparency or just increase throughput. Clinician-builders should be watching this closely because the counter-tool — the thing that audits and challenges AI-generated denials on the provider side — doesn’t exist yet at scale.

“$3B is just a number. Every big company says they’re investing in AI.” Fair, but UHG isn’t announcing a lab or a partnership. They’re reporting that 80% of their engineers are already building with AI agents. This is operational, not aspirational. The spending is already happening.

“This is payer infrastructure — it doesn’t affect how I practice.” It does. When the largest payer automates prior auth and billing code selection, the rules your practice operates under change. The appeal process changes. The documentation requirements change. If you build tools that interact with payer systems, the API on the other side just got a lot more algorithmic.

💡 80/20: UHG is building the AI that decides what gets paid and what doesn’t. Clinician-builders have a window to build the tools that sit on the provider side — audit agents that flag questionable denials, documentation assistants tuned to payer-specific requirements, appeal generators that understand the algorithm’s logic. Try: pick one payer denial pattern you see repeatedly and sketch what an automated appeal would look like.

→ Full write-up

📡 Builder’s Radar

Yuzu Health Raises $35M to Rebuild Insurance Plumbing from Scratch

Yuzu Health raised a $35M Series A led by General Catalyst and Chemistry, with Anthropic’s Anthology Fund participating. The company is a next-generation third-party administrator (TPA) — the behind-the-scenes engine that powers claims processing, payments, and member administration for health plans. Founded in 2022, Yuzu pivoted from building a health plan to building the infrastructure that health plans run on. They’ve processed over $1B in claims across all 50 states, supporting thousands of employers.

😤 Haters

“TPAs are boring back-office plumbing — why should clinician-builders care?” Because the plumbing determines what’s possible. Every health plan innovation — new benefit designs, AI-driven care navigation, real-time eligibility checks — runs through the TPA layer. If the TPA is built on 1990s batch-processing infrastructure, nothing moves fast. Yuzu rebuilding this with modern APIs and unified data is the kind of invisible infrastructure that unlocks what clinician-builders can ship on top.

“Anthropic investing in a TPA? That’s a stretch.” It’s a signal. Anthropic’s Anthology Fund specifically backs companies where AI infrastructure creates leverage. A TPA with unified data and modern architecture is exactly the kind of system where AI agents can automate claims adjudication, stop-loss submissions, and reconciliation — all workflows that currently require humans reading faxes.

💡 80/20: If you’re building anything that touches payer workflows — eligibility checks, claims, benefits verification — watch the TPA layer. Companies like Yuzu are creating the API-first infrastructure that will make real-time payer integrations possible. Reframe: the boring plumbing is what makes the exciting tools work.

🛠️ From the Workbench

Gemma 4 Runs Locally on Your MacBook via LM Studio

Google’s Gemma 4 is a 26B-parameter open model (Apache 2.0 license) that activates only 4B parameters at inference time, making it runnable on consumer hardware. LM Studio 0.4.0 supports it natively — reports of 51 words/sec on a MacBook. It supports tool use, vision input, and a 256K context window. Not medical-specific, but the Apache 2.0 license and local execution make it interesting for clinician-builders who need to experiment with patient-adjacent workflows without sending data to cloud APIs.

⚠️ Verify: Gemma 4 is a general-purpose model, not fine-tuned for clinical use. Do not use for clinical decision-making without thorough evaluation against your specific use case. Local execution reduces data exposure but does not eliminate risk — validate outputs against clinical ground truth before any patient-facing deployment.

😤 Haters

“A general model isn’t useful for clinical work.” Not directly. But a 26B model running locally with tool use and 256K context is a prototyping sandbox. You can test clinical NLP workflows, experiment with structured data extraction from notes, and iterate on prompts — all without an API bill or a data processing agreement.

“51 words/sec is too slow for production.” It’s not for production. It’s for Saturday morning prototyping. The value is zero-cost local iteration before you commit to a cloud API for the real thing.

💡 80/20: If you run Ollama or LM Studio, pull Gemma 4 this week. Use it as a local testing sandbox for clinical NLP experiments — medication extraction, note summarization, structured data parsing. The Apache 2.0 license means you can fine-tune it later if something clicks. If you’re not running local models, you are missing a huge privacy focused aspect of AI building.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Meta Paused Its AI Data Vendor After a Supply Chain Breach. Health AI Should Pay Attention.

Meta suspended all work with Mercor, a $10B AI data startup, after a supply chain attack compromised 4 terabytes of data — including source code, user databases, and SSNs of 40,000+ contractors. The attack vector: a compromised open-source library called LiteLLM with 97 million monthly downloads. Attackers poisoned two versions on PyPI for just 40 minutes. That was enough.

😤 Haters

“This is a crypto/AI startup problem, not a healthcare problem.” LiteLLM is used by anyone routing LLM calls — including health tech companies building clinical AI. If your stack includes any open-source LLM proxy, you share this attack surface. The question isn’t whether you use Mercor. It’s whether your dependency chain includes a package with this kind of single-point-of-failure risk.

“We pin our dependencies and review updates.” Good. But the malicious versions were live for 40 minutes. Automated dependency updates, CI/CD pipelines that pull latest, pre-commit hooks that auto-upgrade — any of these could have caught the poisoned version in that window. The supply chain isn’t just your code. It’s your build process.

→ Full write-up

A VC Just Published the Healthcare AI Survival Checklist

Uma Veerappan of Flare Capital Partners published a framework identifying four factors that separate healthcare AI winners from losers: seamless workflow integration, action-oriented solutions (not dashboards), proprietary data moats, and strong go-to-market distribution. The key insight for builders: ambient AI scribes succeeded specifically because they embedded into the workflow rather than sitting alongside it. The tool that requires a new tab loses to the one that’s already there.

😤 Haters

“Proprietary data moats feel like a recipe for more data silos.” Fair concern. But the argument isn’t about hoarding data — it’s about longitudinal patient context that gets better over time. The distinction matters: a walled garden vs. a learning system with the patient’s full history.

💡 80/20: If you’re building a clinical AI tool, ask the Flare Capital test: does your tool close the loop and take action, or does it surface information and hope someone acts? Reframe: the most dangerous word in health tech is “dashboard.” Build the thing that does the thing, not the thing that shows you the thing.

→ Full write-up

🛠️ From the Workbench

MimiLabs: AI Co-Research Analyst for PubMed and Health Policy

Yubin Park, PhD shared MimiLabs, an AI-powered research tool designed for PubMed literature review, health policy analysis, and Medicare/Medicaid analytics. Think of it as a domain-specific research copilot — rather than generic search, it’s built to understand clinical evidence hierarchies and policy structures.

⚠️ Verify: Evaluate data handling practices before using with any sensitive research data. The tool interfaces with PubMed’s public API, but confirm how queries and results are stored or processed before integrating into institutional workflows.

😤 Haters

“Another AI research tool — how is this different from Elicit or Consensus?” Domain focus. MimiLabs is built specifically for health policy and Medicare/Medicaid analytics, not general academic research. Whether that specialization actually improves results over general tools requires testing with your specific research questions.

“I can just use Claude or ChatGPT with PubMed.” You can. But the value proposition of domain-specific tools is structured output — not just finding papers, but extracting the specific data points (NNT, confidence intervals, policy docket numbers) that health services researchers actually need. Having the data in one Databricks instance is where MimiLabs shines.

💡 80/20: If you do any health policy or clinical evidence research, test MimiLabs against your current workflow on one real question. Try: take the last PubMed search you did manually and run it through MimiLabs. Time both. The delta tells you whether specialization matters for your use case.

🧰 Builder’s Tip

Tool Spotlight: Medplum — Open-Source FHIR Server You Can Run Locally in 5 Minutes

If you’re building anything that touches patient data and you don’t have a FHIR sandbox yet, Medplum is the fastest way to get one. It’s open-source, FHIR R4 compliant, and comes with synthetic patient data out of the box.

npx create-medplum-app@latest my-health-app
cd my-health-app
npm run dev

Three commands and you have a local FHIR server with a React frontend, synthetic patients, and full CRUD on every FHIR resource type. Build your medication reconciliation tool, your patient summary viewer, your CDS hook prototype — all against realistic data without touching anything real. The Medplum team also maintains a Storybook with pre-built React components for common clinical UI patterns (patient timelines, medication lists, diagnostic reports). You’re not starting from scratch — you’re assembling.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Anthropic dropped $400M on nine people and a thesis. Not a product — a team of computational drug discovery researchers from Genentech who launched a startup eight months ago. The Claude-maker is betting that the most valuable thing in healthcare AI isn’t the model. It’s the people who know where the biology breaks. If that sounds familiar, it should — it’s the same bet clinicians.build is making, just at a different scale.

Anthropic Acquires Coefficient Bio for $400M — Its First Biotech Bet

Anthropic bought Coefficient Bio in an all-stock deal reported at $400M. The startup was eight months old with nine employees — both founders came from Genentech’s Prescient Design group, where they worked on computational drug discovery. The team joins Anthropic’s healthcare life sciences group.

The math on this is striking. $400M for nine people is roughly $44M per head. That’s not a product acquisition. That’s Anthropic saying: the people who understand where biology meets computation are worth more than almost any product we could build ourselves. Wow.

😤 Haters

“$44M per head for a startup with no product is insane.” It is. But Anthropic isn’t buying revenue — it’s buying the ability to build domain-specific Claude tools for pharma and biotech. Google did the same thing with DeepMind’s AlphaFold team. The question isn’t whether the price is rational today. It’s whether it looks cheap in three years.

“This is drug discovery, not clinical tools — it doesn’t affect clinician-builders.” Not directly, not yet. But Anthropic is building out a healthcare vertical. Every investment they make in life sciences infrastructure makes Claude better at understanding clinical data, drug interactions, and biological mechanisms. The model you’re building on just got more context about your domain.

“Anthropic is just chasing the next revenue vertical.” Maybe. But against their $380B valuation, this is a $400M bet — roughly 0.1% dilution. That’s a signal, not a pivot.

💡 80/20: The company behind Claude is building a healthcare-specific bench. Every clinician-builder working with Claude benefits from Anthropic understanding your domain better. Try: if you’re evaluating which LLM to build clinical tools on, weight “does the provider invest in healthcare domain expertise?” alongside benchmarks.

→ Full write-up

📡 Builder’s Radar

AI Could Turn Every PCP Into a “Generalist-Specialist”

Bob Kocher, Robert Wachter, and Siobhan Nolan-Mangini published a Health Affairs Scholar paper arguing that AI can collapse the boundary between generalist and specialist. Their thesis: AI-augmented primary care physicians could manage full constellations of chronic conditions across disease-based domains — cardiometabolic, infectious, inflammatory — rather than organ-specific specialties. The catch: it requires reforming medical education, malpractice standards, and credentialing frameworks.

😤 Haters

“We’ve heard ‘AI will replace specialists’ before.” This paper doesn’t say replace. It says augment and redistribute. The claim is narrower and more practical: an AI-equipped PCP handling stable heart failure follow-up frees the cardiologist for the case that actually needs one.

“Credentialing reform and malpractice modernization? Good luck.” Fair. These are decade-long fights. But the paper names the specific barriers, which is more useful than hand-waving about “AI transformation.”

💡 80/20: If you’re building clinical decision support, this paper reframes the user: not “helping a PCP do PCP things” but “giving a PCP specialist-level context for conditions they already manage.” Reframe: the most valuable CDS tool isn’t the one that helps a specialist go faster — it’s the one that helps a generalist go deeper.

→ Full write-up

NYC Hospital CEO: “We Could Replace a Great Deal of Radiologists with AI”

Mitchell Katz, CEO of NYC Health + Hospitals — the nation’s largest public hospital system — told a forum that the network is ready to let AI handle first reads on imaging, with radiologists checking abnormals. He framed it as a cost and access play, particularly for breast cancer screening. The statement was from a March 25 forum, but the coverage wave and community reaction hit this week.

😤 Haters

“Another administrator who doesn’t understand radiology making sweeping claims.” The pushback from radiologists was immediate and pointed. One called it ‘undeniable proof that confidently uninformed hospital administrators are a danger to patients.’ The clinical reality of first-read accuracy, liability, and edge cases is more complex than the CEO framing suggests.

“AI first-reads would increase access to screening.” This part has merit. NYC H+H serves underserved populations where radiologist access is genuinely constrained. The question is whether the access gain outweighs the risk of missed findings — and who’s liable when the AI misses a cancer.

💡 80/20: If you’re building imaging AI, this is a signal that health system leadership is ahead of regulatory frameworks. Try: build the audit layer — the tool that measures AI first-read accuracy against radiologist reads in your specific patient population — before building the AI reader itself.

→ Full write-up

OpenEvidence Lands First Enterprise Deal at Mount Sinai — Embedded in Epic

OpenEvidence completed its pivot from PLG to enterprise, landing a B2B deal with Mount Sinai that embeds its clinical evidence tool directly in Epic’s EHR. The company had built a following with individual clinicians; now it’s selling to health systems. Abridge partnerships are also in play, and clinical trial enrollment is the next adjacency.

😤 Haters

“Another AI tool embedded in Epic — how is this different?” OpenEvidence competes with UpToDate, not ambient scribes. It surfaces evidence at the point of care, which means it’s fighting for the same real estate as the reference tools clinicians already use. The question is whether AI-surfaced evidence displaces the muscle memory of “I’ll just look it up on UpToDate.”

“PLG to enterprise is where startups go to die.” Sometimes. But for clinical tools, enterprise is where you get into the EHR — and the EHR is where clinicians actually live. Individual adoption without system integration creates shadow IT problems.

💡 80/20: The PLG-to-enterprise path is the playbook for clinician-builders who want health system adoption. Try: build for individual clinician delight first, measure usage, then bring those numbers to your CMIO. Mount Sinai didn’t buy OpenEvidence cold — they bought it because clinicians were already using it. Makes me think of Butterfly ultrasound approach.

→ Full write-up

Best AI Agent Scored 1 out of 4 on Self-Assessment

Nate tested four prominent “outcome agents” — Anthropic’s Cowork, Lindy, Sauna, and Google’s Opal — against a framework built on one question: can the agent assess the quality of its own output? The best scored 1/4. The core insight: code has test suites, knowledge work doesn’t. Agents that can’t self-evaluate can’t improve.

😤 Haters

“This is a sample size of four tools with a subjective rubric.” True — it’s one person’s evaluation. But the framework itself is the value: does the agent know when it’s wrong? That’s the question every clinician-builder should ask about their own tools.

“Agents are improving fast — this is just a snapshot.” Agreed. But the structural problem — knowledge work lacks the equivalent of unit tests — won’t be solved by better models. It’ll be solved by builders who define what “correct” looks like for their specific clinical workflow.

💡 80/20: Before deploying any AI agent in a clinical workflow, define your test suite. Not “does it work?” but “what does wrong look like, and will I catch it?” Try: write 5 failure cases for your AI tool before you write a single success metric.

→ Full write-up

What are you building this week? Reply and tell me — I read every one.

— Kevin

🔬 The Big Thing

Your Agent Is 80% Plumbing — The Claude Code Leak Reveals What Production Clinical AI Actually Requires

A detailed analysis of the accidentally published Claude Code source (1,902 files, 512K+ lines) mapped 12 infrastructure primitives that make up the 80% of the system that isn’t the LLM call. Session persistence. Permission pipelines. Context budget management. An 18-module security stack for a single shell command. Error recovery. Developers have already ported the harness to Python and Rust — the patterns are structural, not Anthropic-specific.

😤 Haters

“This is just good engineering 101 — anyone shipping production software knows you need session management and error handling.” Fair. But the gap between “knows you need it” and “has actually built it for an agent system” is where every clinical AI demo dies. Most tutorials stop at the prompt. These patterns are documented now. Use them.

“Clinician-builders don’t need to worry about 18-module security stacks — that’s for enterprise teams.” If your agent touches a medication list, you need permission scoping. If it runs across a multi-patient session, you need context budgets. The scale is different. The primitives aren’t optional.

“The leak was irresponsible and Anthropic should be embarrassed.” Two leaks in one week from a company shipping at Anthropic’s velocity. Development speed outrunning operational discipline is a pattern every fast-moving builder should learn from, not just observe.

💡 80/20: The LLM is the easy part. Your next build session: audit your agent for crash recovery, permission scoping, and context overflow handling. If any of those are missing, that’s why it works in the demo but not in the clinic. Try: run the architecture audit prompt against your own agent stack.

→ Full write-up

📡 Builder’s Radar

Anthropic Launches “Claude Code for Healthcare” Webinar — First Major AI Lab Targeting Physician-Developers

Anthropic announced a webinar for April 23 focused on physicians building with Claude Code — live demos, safety verification, compliance traceability, and Q&A with the Claude Code team. Graham Walker (MDCalc founder) amplified it. This is the first time a major AI lab has created a developer event specifically for clinicians who build. Not executives learning about AI. Physicians writing code with agents.

😤 Haters

“It’s a marketing webinar, not a product launch.” Probably. But the framing matters — Anthropic is acknowledging physician-developers as a category worth targeting. That’s a market signal even if the content is surface-level.

“One webinar doesn’t mean they’ll build healthcare-specific features.” True. Watch for follow-through: FHIR-aware tools, HIPAA-scoped permissions, clinical audit logging. The webinar is the signal. The product roadmap is what matters.

💡 80/20: If you’re building with Claude Code for clinical use cases, register and come with specific technical questions about auditability and compliance traceability. The Q&A is where the real value lives. Try: prepare one question about how Claude Code handles PHI in agent workflows.

→ Full write-up

Gemma 4 Ships Under Apache 2.0 — Four Sizes, Edge to 31B

Google released Gemma 4 in four sizes: an edge model for Raspberry Pi, up to a 31B dense model ranked #3 among open models. First Gemma under Apache 2.0. Optimized for reasoning and agent workflows. NVIDIA announced RTX acceleration for local deployment.

😤 Haters

“Another open model release — the benchmarks all look the same.” The license matters more than the benchmarks here. Apache 2.0 means health systems can fine-tune and deploy without legal review of restrictive license terms. That’s a real barrier removed for clinical use.

“MedGemma still isn’t clinical grade — Google said so themselves.” Right. But Gemma 4’s improved base capabilities should translate to better clinical fine-tunes. And the edge model running on a Raspberry Pi opens up point-of-care possibilities that need zero cloud connectivity and zero data exfiltration.

💡 80/20: If you’re running Ollama or LM Studio, watch for GGUF quantizations of the 31B model. The edge model (E2B) is worth testing for simple clinical NLP tasks — triage classification, symptom extraction — running entirely on-device. Try: pull the 31B when available and benchmark it against your current local model on a clinical task you care about.

→ Full write-up

FDA Raises the Bar: AI “Breakthroughs” Must Solve Problems Physicians Can’t

A STAT analysis shows the FDA is shifting what qualifies as a breakthrough AI device. Early designations went to tools improving physician performance. Now the agency favors algorithms that solve problems physicians cannot address independently — detecting multiple cancers from single images, predicting mortality from signals no human could integrate.

😤 Haters

“This just makes the breakthrough pathway harder for startups.” It clarifies the pathway. Your clinical decision support tool probably doesn’t belong in breakthrough designation anyway — it belongs in the 510(k) lane or outside device regulation entirely.

“The FDA is moving goalposts.” The FDA is doing what regulators should do — raising the bar as the technology matures. “Better than a doctor at this one thing” was appropriate when AI radiology was novel. “Does something no doctor can do” is the right bar now.

💡 80/20: If you’re building a clinician-facing tool, this clarifies your regulatory strategy. Most clinician-built tools (documentation, decision support, workflow automation) don’t need breakthrough designation. The FDA is telling you where the high bar is so you can plan accordingly. Reframe: regulatory clarity is a gift, not a barrier.

→ Full write-up

Happy spring.

— Kevin

🔬 The Big Thing

Epic’s Predictive AI Models Underperform in Real-World Clinical Settings

A meta-analysis by Northwell Health researchers, published in SpringerNature’s Journal of General Internal Medicine, reviewed five of Epic’s out-of-the-box AI tools — the Deterioration Index, Sepsis Model, Unplanned Readmission Model, End of Life Care Index, and Risk of No Show Model. None exceeded an AUROC of 0.79. Confidence intervals for three models fell below Epic’s own published performance stats. The clinical consequence: false positives driving unnecessary diagnostics, antimicrobial overuse, and alert fatigue across the 42% of U.S. acute care systems running Epic.

Epic responded that the study examined first-generation models and pointed to second-generation versions with local fitting capabilities. That’s a fair clarification — but it sidesteps the real problem. Most health systems deployed first-gen models because Epic shipped them as ready-to-use. Local validation requires clinical informatics teams, data infrastructure, and protected time that most hospitals don’t have. The gap between “available” and “validated” is where patients sit.

😤 Haters

“This is just academic researchers dunking on industry. Every new model underperforms at first.” The researchers weren’t testing bleeding-edge tools — they reviewed models that have been deployed in live clinical workflows for years. Underperformance at launch is expected. Underperformance after widespread deployment is a patient safety issue.

“Epic already fixed this with second-gen models.” Maybe. But there’s no independent validation of the second-gen models yet, and no consensus on how health systems should even run those validations. The ONC transparency rules that would have required this kind of disclosure may be sunsetted under the current administration. So the fix is: trust Epic.

“AUROC of 0.79 isn’t that bad.” In a controlled setting, no. In a sepsis alert that fires on a patient who doesn’t have sepsis, waking up a resident, triggering an antibiotic cascade — the clinical cost of false positives compounds fast.

💡 80/20: If you’re building on top of EHR-native AI predictions, don’t inherit their confidence scores uncritically. Build your own validation layer — even a simple confusion matrix on your patient population tells you more than the vendor spec sheet. Try: pull 100 recent alerts from your institution’s Deterioration Index and chart how many led to a meaningful clinical intervention.

→ Full write-up

EFF Sues CMS for Transparency on Medicare’s AI Prior Authorization Pilot

The Electronic Frontier Foundation filed a FOIA lawsuit against CMS on March 25 — and it’s only now getting coverage outside the legal press. The suit demands records about WISeR — an AI system evaluating prior authorization requests across six states, affecting roughly 6.4 million Medicare beneficiaries. Nobody outside CMS knows which vendors built it, which models it uses, or what safeguards exist. The most troubling detail: vendors are compensated partly through denial rates, with up to 20% of savings from rejected authorizations flowing back to them. That’s a financial incentive baked directly into the algorithm’s decision function.

😤 Haters

“Prior authorization has always been opaque. This isn’t new.” It’s not new that prior auth is a black box. It is new that the black box is now an AI making automated decisions at scale, with no disclosed audit mechanism and a financial incentive to deny. The EFF’s involvement signals this is crossing from health IT complaint to civil liberties issue.

“CMS has the right to pilot new programs without full disclosure.” They do. But FOIA exists precisely for programs that affect millions of people and lack transparency. If the model is sound, transparency should be easy.

💡 80/20: If you’re building anything in the prior authorization space, the WISeR pilot is your regulatory weather vane. Whatever transparency and audit requirements emerge from this lawsuit will likely shape the standard for all AI-driven auth tools. Reframe: build your audit trail now, not after regulators require it.

→ Full write-up

Qualified Health Raises $125M Series B — Anthropic Invested

Qualified Health announced a $125M Series B last week, led by NEA, with participation from Menlo Ventures’ Anthology Fund (created with Anthropic), Transformation Capital, and others. The public benefit corporation builds an enterprise AI platform for health systems — workflow automation, agent development, clinical safeguards, and governance infrastructure. Customers include Emory, Jefferson Health, and the entire UT System (8 institutions). UTMB reported $15M+ in measurable impact within six months. The platform now supports 500,000+ users across systems representing roughly 7% of U.S. hospital revenue.

😤 Haters

“Another AI platform raise. What makes this different?” Anthropic’s direct investment through Menlo’s Anthology Fund. That’s not just a check — it’s a signal that the model provider sees governance and clinical safeguards as a necessary layer, not a nice-to-have. Most AI startups sell the model. Qualified is selling the operational wrapper.

“$125M for governance software? Health systems won’t pay for guardrails.” UTMB generated $15M in run-rate impact in six months. The governance isn’t the product — it’s what makes the product deployable in a regulated environment.

💡 80/20: The “AI platform for health systems” category is consolidating fast. If you’re a clinician-builder shipping tools into health systems, study Qualified’s approach to governance and auditability — that’s the bar your tool will be measured against. Try: map your tool’s audit trail today and identify the gaps.

→ Full write-up

Healthcare’s Data Quality Crisis — “We Don’t Know What We Don’t Record”

Emergency physician and Epic consultant John Lee published the sharpest essay I’ve read on healthcare’s data quality problem. A stroke patient arrives with an empty medication list — no data from other systems. Clinicians guess whether to give thrombolytics. Medication lists track what was ordered, not what patients actually take. Safety reporting takes 10+ minutes per incident, so nobody files reports. The only data healthcare collects with genuine fidelity is billing and coding. Lee’s line that sticks: “An algorithmically confident recommendation generated from a dirty medication list is still a recommendation you should not trust.”

😤 Haters

“This is a known problem. Everyone in health IT knows the data is messy.” Knowing it and quantifying the clinical risk are different things. The stroke scenario isn’t hypothetical — it’s what happens every night in every ED. The question isn’t whether you know the data is bad. It’s whether you’re building systems that assume it’s good.

“AI can clean the data. That’s the whole point.” Lee actually agrees — he points to Epic’s Agent Factory as a potential tool for automated medication reconciliation. The catch is that the AI cleaning the data needs to be validated against the same data it’s trying to fix. It’s a bootstrapping problem.

💡 80/20: Before you build any clinical AI tool, spend one shift manually auditing the data it will consume. Pull 20 medication lists and compare them to what patients actually report taking. The gap will reshape your architecture. Reframe: your AI’s ceiling is your data’s floor.

→ Full write-up

🎯 Clinician-Builder Tip of the Day

When you’re prototyping a clinical tool, don’t start with the model. Start with the data audit. Pick the single most important data element your tool depends on — a medication list, a problem list, a lab trend — and manually review 20 patient records for accuracy. Time yourself. If it takes you more than 30 seconds per record to spot a discrepancy, your users will never catch the AI’s mistakes either. That 30-minute exercise will save you weeks of building on a foundation that doesn’t hold.

What are you building this week? Reply and tell me — I read every one.

— Kevin

The Chatbot Is the Bottleneck, Not the Model

Ethan Mollick’s latest analysis cites research showing financial professionals using GPT-4o saw real productivity gains — but the chatbot interface itself created cognitive overload that partially cancelled them out. “Giant walls of text, offers to pursue new topics, and sprawling discussions” hurt less experienced workers most — the exact people who’d benefit most from AI. Mollick argues the next AI leap isn’t bigger models but better delivery: specialized interfaces, familiar platforms, and context-specific tools. Claude Dispatch — send a task from your phone, get results later — is his example of “post-chatbot AI.”

😤 Haters

“This is obvious. Everyone knows chatbots aren’t the final form.” Knowing it and quantifying it are different. The research shows interface overhead measurably reduces AI productivity gains. If you’re building a clinical AI tool and your delivery mechanism is a chat window, you’re leaving value on the table — and the people who need it most (residents, new nurses, rural clinicians with less support) are the ones losing out.

“Purpose-built interfaces are just more expensive to build.” They are. But the vibe coding era changes the cost equation. A clinician who understands the workflow can now build a focused interface for their specific use case faster than an enterprise vendor can ship a generic one. That’s the entire thesis of clinicians.build.

💡 80/20: If you’ve built or are building a clinical AI tool delivered through a chat interface, the Mollick research suggests you’re underselling your own tool. Reframe: what would a single-purpose interface look like for your most common use case? One input, one output, no menu.

AI Products Need Failure Mode Report Cards

Automate Clinic published a concept called AI Failure Mode Literacy — the ability to understand when and how AI tools succeed or fail. Right now, developing this intuition requires what they call an “insane amount” of hands-on use. Their proposal: every AI product should ship with a consistently updated report card documenting its failure modes. Not just accuracy benchmarks, but contextual failure patterns.

😤 Haters

“Vendors will never voluntarily publish their failure modes.” Probably not — at least not the incumbents. But imagine a clinical AI marketplace where the tools that publish failure data earn trust faster than the ones that don’t. Transparency becomes a competitive advantage. The first vendor to ship real failure documentation in a clinical context will differentiate on something no one else is offering.

“This is just model cards rebranded.” Model cards describe training data and performance metrics. Failure mode reports would describe contextual behavior — when does this tool break in practice, under what clinical conditions, with what patient populations? That’s operationally useful in a way model cards aren’t.

💡 80/20: Next time you evaluate a clinical AI tool, ask the vendor: “What does this tool get wrong, and how often?” If they can’t answer, that’s your answer. Try: keep a running log of failure modes for every AI tool you use clinically — even just a shared doc. You’re building the report card that doesn’t exist yet.

→ Full write-up

🛠️ From the Workbench

LangCare — Open Source MCP Server for FHIR + AI Agents

LangCare (formerly AgentCare) is an open-source FHIR MCP server written in Go that connects AI agents — Claude, ChatGPT, Gemini — to Epic, Cerner, and any FHIR R4 EMR. Ships with 40+ clinical agentic skills (medication management, lab interpretation, clinical decision support) and supports 150+ FHIR R4 resources. MIT licensed. Install via npm: npm install -g @langcare/langcare-mcp-fhir. Shared in the HTN community by creator Hari Kolasani — already past 1,600 installs in under three months.

⚠️ Verify: “HIPAA-compliant” is a vendor claim. Before routing any real patient data through this, confirm BAA availability, review the PHI scrubbing implementation, verify TLS configuration, and understand the zero-persistent-storage architecture yourself. Open source means you can audit it — which is an advantage, but also a responsibility.

😤 Haters

“An npm-installed MCP server handling PHI is a security nightmare waiting to happen.” It’s a reasonable concern. But the architecture is a stateless proxy — it translates MCP requests to FHIR calls and passes back structured responses. No persistent storage means no data at rest to breach. The real risk surface is in transit: OAuth2 token handling, TLS implementation, and whether the PHI scrubbing actually catches everything. Being open source means you can verify all of this, which is more than you can say for most commercial FHIR middleware.

“1,600 installs doesn’t mean production-ready.” Correct. But 1,600 installs in under three months for a healthcare-specific MCP server means there’s real demand for this connector layer. If you’re experimenting with AI agents in a sandbox environment with synthetic FHIR data, this is worth testing. Production with real PHI is a different conversation.

💡 80/20: If you’ve wanted to connect Claude or another LLM to a FHIR sandbox but didn’t want to build the plumbing, this is your shortcut. Try: set up LangCare against a HAPI FHIR test server with synthetic data this weekend. Zero patient risk, full agent capability.

→ Full write-up

🎯 Clinician-Builder Tip of the Day

Before you build the feature, build the test. Not a unit test — a clinical scenario test. Write down five specific patient encounters where your tool should help, and five where it should stay out of the way. Run those scenarios against your prototype before you write another line of code. The five “stay out of the way” cases will teach you more about your tool’s actual value than the five where it shines. A tool that knows when to be quiet is more trustworthy than one that always has an answer.

What are you building this week? Reply and tell me — I read every one.

— Kevin

🔬 The Big Thing

Oracle Fires 30,000 People to Fund AI — and May Sell Cerner to Cover the Tab

Oracle began laying off up to 30,000 employees on March 31, notifying workers via a 6 AM email from “Oracle Leadership.” No advance warning. No manager conversation. The cuts span divisions globally, with India reportedly losing 12,000 positions alone. TD Cowen estimates the workforce reduction will free up $8-10 billion in cash flow — money Oracle needs because US banks have roughly doubled interest rate premiums for data center financing since September, and Oracle’s $156 billion AI infrastructure buildout needs capital from somewhere.

The health tech angle is the one nobody’s talking about yet: Oracle is weighing a sale of Cerner, the healthcare EHR platform it acquired for $28.3 billion in 2022. Cerner has already been gutted by repeated post-acquisition layoffs, and Oracle’s cloud infrastructure revenue grew 66% year-over-year while Cerner sputtered. The math is simple: Oracle needs cash for AI, Cerner isn’t generating it, and the acquisition has been described by multiple analysts as a bet that hasn’t paid off.

For clinician-builders, this is a platform risk lesson playing out in real time. Hundreds of health systems run on Oracle Health. If Cerner changes hands again, every integration, every API dependency, every workflow automation built on that platform enters a period of uncertainty. The VA’s troubled Cerner modernization is already a cautionary tale. An ownership change adds another layer of instability to an already shaky foundation.

😤 Haters

“Oracle clarified the 30K number is just analyst speculation — the actual cuts might be smaller.” Oracle issued a vague clarification but hasn’t denied layoffs are happening. Employees across the US, India, and Europe received termination emails on the same morning. Whether it’s 20,000 or 30,000, the scale is unprecedented for Oracle, and the Cerner sale discussion is coming from analysts who cover the company’s debt structure, not from rumor mills.

“Cerner getting sold could actually be good — a focused health IT buyer might invest more than Oracle did.” That’s possible. But the transition period is the problem. Acquisitions of this scale take 12-18 months to close, during which product roadmaps freeze, engineering talent leaves, and clients can’t get straight answers about their contracts. If you’re mid-build on an Oracle Health integration, you’re now building on sand for at least a year.

“This is just corporate restructuring — it doesn’t affect the actual EHR product.” It already has. Oracle has repeatedly cut Cerner headcount since 2022. The VA EHR modernization has been plagued with issues. When you remove that many engineers from a health IT platform, the product degrades whether the org chart says so or not.

💡 80/20: Platform dependency is a clinical risk, not just a technical one. If your tools run on Oracle Health, start mapping which integrations are FHIR-standard (portable) versus proprietary (locked in). Try: audit your API dependencies this week — anything that only works on Oracle Health is now a liability, not a feature.

→ Full write-up

📡 Builder’s Radar

North Korea Just Backdoored the Most Popular JavaScript Library — and Vibe Coders Should Pay Attention

On March 31, a North Korea-linked threat actor (UNC1069) compromised the axios npm package — the most popular JavaScript HTTP library, with roughly 100 million weekly downloads and presence in an estimated 80% of cloud environments. The attacker hijacked a maintainer account, injected a malicious dependency that deployed a cross-platform remote access trojan (WAVESHAPER.V2), and had the backdoor live for about three hours before removal. Wiz has already detected the malicious versions in roughly 3% of scanned environments.

😤 Haters

“Three hours isn’t long enough to matter — most people wouldn’t have pulled the update.” With 100 million weekly downloads, even a three-hour window means thousands of installations. CI/CD pipelines that auto-install latest versions are particularly exposed. And the malicious package was designed to persist — it installs a backdoor, not a one-time exploit.

“This is a general dev problem, not a health tech problem.” It is specifically a health tech problem. If you’re a clinician vibe-coding a patient-facing tool with npm packages — and many of you are — your dependency tree is your attack surface. As one clinician-builder put it: “This is the biggest issue with using CLI or consumer vibe code models without proper harnesses.”

💡 80/20: Your vibe-coded clinical tool inherits every vulnerability in its dependency tree. Try: run npm audit on every project you have in production today, and pin your package versions instead of using ranges. If you don’t know what that means, that’s the point — npm audit is a one-liner that tells you if you’re exposed.

→ Full write-up

Avo Raises $10M to Put an AI Copilot Inside Your EHR

Avo closed a $10 million Series A led by Noro-Moseley Partners, with participation from AlleyCorp. The company started during COVID as a no-code tool helping hospitals operationalize clinical protocols — and has evolved into an LLM-powered AI platform that sits inside Epic, athenahealth, and MEDITECH. Its copilots (Chart Assist, Ask Avo) synthesize patient data, draft documentation, and pull guidelines into the physician’s workflow. Mass General Brigham is among its users. The round also brings a strategic partnership with EBSCO DynaMed to integrate evidence-based clinical decision support directly into the overlay.

😤 Haters

“Another AI documentation tool — the market is saturated.” Avo isn’t just documentation. The DynaMed integration is the interesting part: evidence-based guidelines pulled directly into the EHR context, not as a separate app you have to alt-tab to. That’s a different value proposition than another scribe.

“$10M is small — can they compete with Abridge and Nuance?” Different play. Abridge and Nuance are ambient listening. Avo is an interactive copilot inside the EHR itself. The question isn’t who wins ambient — it’s whether the copilot layer becomes a separate product category. Mass General Brigham thinks so.

💡 80/20: Avo’s trajectory — COVID-era no-code tool → LLM-powered AI platform — is the clinician-builder growth pattern. Try: if you built a clinical protocol tool during the pandemic, ask yourself what it becomes with an LLM backbone. The answer might be your Series A.

→ Full write-up

Headway Acquires Tezi’s AI Team to Reduce Mental Health Admin Friction

Headway, the largest mental health provider network in the US (70,000+ providers, all 50 states), acqui-hired the team behind Tezi — an AI company that built systems combining human judgment with AI agents for complex workflows. Tezi cofounder Raghavendra Prabhu (ex-Google, Microsoft, Twitter, Pinterest) joins as VP of Engineering. CEO Andrew Adams: “AI can help improve the infrastructure around care so clinicians can spend more time with patients and less time navigating complexity.”

😤 Haters

“Acqui-hires rarely produce real product changes — they’re talent grabs.” Sometimes. But Headway has a specific problem to solve: matching 70,000 providers with patients across insurance networks, credentialing, and scheduling. That’s exactly the kind of complex workflow Tezi built AI for. The talent has a target.

“After Jimini and Doctronic, do we need another mental health AI story?” This is the infrastructure play, not the clinical play. Headway isn’t building a therapy chatbot — they’re reducing the operational friction that makes it hard for therapists to take insurance. Different layer, same ecosystem.

💡 80/20: The mental health AI landscape is splitting into clinical tools (Jimini, Doctronic) and infrastructure tools (Headway). Reframe: if you’re building in behavioral health, decide which layer you’re on — the patient-facing conversation or the operational plumbing underneath it. Both need builders. The plumbing might be the bigger opportunity.

→ Full write-up

ONC Is Back — ASTP Name Reverts, and the Office Loses Its AI and Cyber Roles

The Federal Register confirmed what the health IT community expected: ASTP is reverting to ONC. But it’s not just a name change. The Chief Technology Officer, Chief Data Officer, and Chief AI Officer roles — plus some cybersecurity functions — are moving out of ONC and back under HHS’s Chief Information Officer. ONC is being narrowed to focus on two things: getting patients their health data and reducing friction in health record sharing. The community discussion has been substantive, with former ONC officials and standards experts weighing in.

😤 Haters

“It’s a name change. Nobody cares.” The name is cosmetic. The scope change isn’t. Stripping the AI, data, and cyber roles out of ONC means the office that certifies health IT no longer oversees the AI and cybersecurity standards that health IT increasingly depends on. That’s a governance gap worth watching.

“Narrowing ONC’s focus is actually good — they should focus on interoperability.” Fair. ONC trying to be the AI, cyber, data, AND interoperability office was arguably too much scope. But the question is whether the CIO’s office will prioritize health-specific AI guidance the way ONC would have. Internal IT priorities and external health tech standards are different conversations.

💡 80/20: ONC narrowing to interoperability + patient access might mean less regulatory friction for health IT builders in the near term — especially with HTI-5 proposing to cut 50% of certification criteria. Try: if you’ve been waiting for certification clarity before building, the window of regulatory flexibility is opening. Ship now.

→ Full write-up

🎯 Clinician-Builder Tip of the Day

Lock your dependencies. If you’re using npm, pip, or any package manager for a clinical tool — even a prototype — pin your package versions to exact numbers instead of ranges. The axios attack hit because automated installs pulled the latest malicious version. In your package.json, change "axios": "^1.14.0" to "axios": "1.14.0" (no caret). In Python, use pip freeze > requirements.txt and commit that file. It takes five minutes and it means a compromised upstream package doesn’t silently infect your project overnight. Your patients don’t know what a dependency tree is. That’s why you need to.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Claude Code can now run scheduled tasks on Anthropic’s infrastructure overnight

Anthropic shipped scheduled tasks for Claude Code on the web — agents that run on Anthropic-managed infrastructure even when your device is off. Example use cases from the docs: reviewing open pull requests every morning, analyzing CI failures overnight, syncing documentation after PRs merge, running dependency audits weekly. Available to all Claude Code on the web users as of March 30.

😤 Haters

“Running clinical data through Anthropic’s cloud infrastructure for scheduled tasks is a HIPAA problem, not a feature.” This is the right concern. ⚠️ Do not point real patient data at scheduled tasks until you’ve verified BAA availability for Claude Code on the web — this isn’t documented clearly yet. The appropriate use cases right now are for your tools and code, not patient records: overnight documentation audits of your codebase, weekly dependency vulnerability scans, automated issue triage summaries.

“Claude Code is a developer tool. I’m a clinician who builds things, not a developer.” If you use Claude Code to build your clinical tools, you can now automate the maintenance work — the part that eats into your protected build time. That’s the unlock.

💡 80/20: The clinical workflow use cases will come. Today’s immediate value is for the solo clinician-builder who does their own maintenance: nightly alerts when a dependency breaks, weekly summaries of open issues in your project. Try: set up a weekly task that reviews your project’s README and flags anything that’s no longer accurate — clinical tools drift fast and documentation drifts faster.

→ Full write-up

AI agents don’t fail the way you think — and clinical workflows can’t afford the difference

Nate published a March 30 analysis of how AI Skills — the structured, reusable capability specs that agents invoke — behave differently when no human is watching. His finding: Skills built to work when a human oversees the output fail at dramatically higher rates when agents invoke them autonomously in loops. “Fail 10% of the time when you’re watching. Fail 100% of the time when you’re not.” The mechanism: skills written for human-supervised use cases don’t specify their failure modes, because a human can recognize and redirect a bad output. An agent cannot.

😤 Haters

“This is about Microsoft Office Skills — productivity software. Not the same as clinical AI agents.” The mechanism is identical. A skill spec that doesn’t define what to return on a missing input, a malformed result, or an edge case outside its training distribution fails the same way whether it’s analyzing a quarterly report or checking medication dose thresholds.

“10% failure rate is acceptable for most workflows.” Not in clinical workflows. A 10% failure rate on a medication interaction check is one missed interaction per 10 queries. At 30 medication reviews per shift, that’s 3 silent errors per physician per day. The tolerance for silent failure in clinical AI is categorically different from productivity tools.

💡 80/20: Every clinical AI skill you build should be specified as if no human will ever see its output. That means explicit failure returns — not just “return the result” but “return {status: 'error', reason: 'missing_dosing_context'} when the input doesn’t match the expected structure.” Try: write the failure cases before you write the success case. If you can’t enumerate what your skill does when things go wrong, it isn’t ready to run without oversight.

→ Full write-up

🛠️ From the Workbench

AutoBe — an AI agent that actually writes working backends

AutoBe is an open-source agent that takes a natural language conversation and generates a complete backend — data types, API endpoints, function stubs. The interesting part is the harness: it uses type schemas to constrain what the model can output, then runs a compiler to verify the result, then feeds structured error messages back to the model with a feedback loop. The result, per their March 30 writeup, is boosting function calling success rates from 6.75% to 99.8% for the tested models. The mechanism — constrain → compile → isolate → classify → feed back — is a general pattern for making AI code generation actually reliable.

⚠️ Verify: “AI-generated backend” does not mean “audited backend.” For any clinical tool that touches PHI, AI-generated code is the starting scaffold, not the finished product. Before AutoBe-generated code handles real patient data, it needs a security review, input validation for clinical edge cases, and explicit error handling that goes beyond what the generator will produce by default.

😤 Haters

“AI-generated backends for clinical tools are a compliance nightmare.” The generated code is — unchecked. The harness approach improves code quality significantly compared to direct model output, but it doesn’t produce SOC2-ready or HIPAA-audited code. Use it to generate the skeleton and data models fast, then audit before connecting to anything real.

“A 6.75% → 99.8% improvement on a shopping mall benchmark doesn’t translate to clinical API reliability.” True. The test task was specific. But the harness principle — constrain outputs with type schemas, verify with a compiler, feed structured error messages back — is transferable to any code generation task including clinical data models.

💡 80/20: AutoBe compresses the time from “I know what my clinical tool needs to do” to “I have a working backend scaffold to audit and extend.” That’s the legitimate value. Use it as an accelerator for the parts where your engineering knowledge is the bottleneck, not as a substitute for the clinical judgment in the data model design. Try: feed AutoBe a natural language description of your clinical workflow and see what data types it proposes. The gaps in its model will tell you exactly where your domain expertise is irreplaceable.

🎯 Clinician-Builder Tip of the Day

Before your next build session, write one paragraph describing the specific clinical moment your tool is for. Not the feature list — the moment. “It’s 2 AM, a nurse has flagged a potassium of 6.1, the covering resident has three other things going on, and the EHR doesn’t surface the patient’s last three potassiums or their renal function trend.” That paragraph is your system prompt, your design spec, and your evaluation criteria. Everything your AI assistant builds will be better grounded for it. The clinicians who build the best tools aren’t the ones who know the most about software — they’re the ones who can describe the problem with that kind of specificity.

What are you building this week? Reply and tell me — I read every one.

— Kevin

Prior Auth Goes Public: CMS Forces Payers to Show Their Cards Tomorrow

Starting March 31, every Medicare Advantage organization, Medicaid managed care plan, CHIP managed care entity, and qualified health plan on the federal exchanges must publicly report their prior authorization metrics on their websites. This is the CMS-0057-F Interoperability and Prior Authorization final rule hitting its first real compliance deadline. The data includes: percent of prior auth requests approved, denied, and approved after appeal, plus average time between submission and decision. MA organizations report at the contract level. State Medicaid programs report at the state level. Plans report at the plan level. The reporting template is standardized.

😤 Haters

“This is just a checkbox exercise — payers will bury the data in a PDF on page 47 of their website and no one will find it.” Probably true for the first cycle. But CMS standardized the reporting template, and the data is machine-readable enough that someone will scrape it within a week. The first aggregator to build a prior auth payer comparison dashboard wins a lot of attention from practice managers and referral coordinators.

“The data won’t change payer behavior — they already know their denial rates.” Knowing your denial rate and having it compared publicly to every other payer in your market are different experiences. This is the same dynamic that hospital price transparency created — slow to start, but the comparison tools eventually forced real conversations. And unlike hospital pricing data, prior auth metrics are simpler to compare: approve/deny/appeal is a cleaner signal than a chargemaster spreadsheet.

“Practices don’t have time to analyze payer-level prior auth data.” They don’t need to. The opportunity is for builders: a tool that pulls this data, cross-references it with a practice’s payer mix, and surfaces which payers are the most friction-heavy for specific service lines. That’s actionable on day one.

💡 80/20: This is a builder opportunity hiding in a compliance deadline. The practices that need this data most — small groups, independent specialists, FQHCs — are the least likely to go hunting for it on payer websites. Try: build a simple scraper that aggregates the reports as they go live this week, normalize the data, and publish a comparison. First mover advantage is real here.

→ Full write-up

📡 Builder’s Radar

Deepfake X-Rays Are Good Enough to Fool Radiologists — and AI

A study published in Radiology from Mount Sinai found that ChatGPT-generated X-ray images fooled radiologists 25% of the time even after they were warned synthetic images were present. Seventeen radiologists across 12 centers in 6 countries evaluated 264 images. Individual accuracy ranged from 58% to 92%. The multimodal LLMs tested (GPT-4o, GPT-5, Gemini 2.5 Pro, Llama 4 Maverick) did no better — 57% to 85% accuracy. Most concerning: when radiologists weren’t told synthetic images were in the mix, only 41% noticed anything was off. Experience didn’t help — years of practice had no correlation with detection ability.

😤 Haters

“This is a lab exercise — no one is actually injecting fake X-rays into PACS systems.” Not yet. But the researchers specifically flagged cybersecurity risk: if an attacker gains network access to a hospital, synthetic images injected into the imaging pipeline would be functionally undetectable. The litigation fraud angle is also real — fabricated fractures indistinguishable from authentic ones.

“This is a radiology problem, not a builder problem.” If you’re building anything that processes medical images — AI triage, clinical decision support, quality review — you now need to think about image provenance. Watermarking, chain-of-custody metadata, and authentication layers for imaging data just became relevant to your architecture.

💡 80/20: Image authentication is the next infrastructure layer. If you’re building on medical imaging data, start thinking about provenance now — not as a feature, but as a foundation. Reframe: every imaging AI pipeline needs a “was this image real?” check before it becomes a “was this diagnosis real?” question.

→ Full write-up

Advocate Health Ships Epic Agent Factory Prototypes for Pharmacy and Infusion

Advocate Health’s SVP/chief digital and AI officer Andy Crowder described how Advocate is using Epic’s Agent Factory to build four prototypes targeting pharmacy complex order verification and infusion charting prep. The prototypes came out of a three-day Epic Immersion sprint at the Pearl in Charlotte. Plan is production by July. This is the first concrete implementation story since Epic previewed Agent Factory at HIMSS26 in early March — moving from demo to deployment in under a month.

😤 Haters

“Four prototypes in three days sounds like a hackathon, not a production pipeline.” Fair. But the targets they picked — pharmacy verification and infusion prep — are high-volume, rules-heavy workflows where the error modes are well-understood. These aren’t open-ended AI experiments. They’re automation of specific, repeatable clinical sequences where the human stays in the loop.

“This is just Epic customers building inside Epic’s walled garden.” Yes. And that’s the point. The majority of US hospital workflows run through Epic. An agent framework native to the EHR, with access to the data model and order catalog, is a different beast than bolting an external agent onto FHIR APIs. The walled garden is where the patients are.

💡 80/20: Watch the pharmacy and infusion use cases closely — they’re the canary for agentic workflows in clinical settings. If Advocate ships these in July, the playbook for “3-day sprint → 4-month production path” becomes repeatable. Try: identify one rules-heavy workflow in your system that could be the next sprint candidate.

What are you building this week? Reply and tell me — I read every one.

— Kevin