The Feed: June 18, 2026
FHIR Just Became a HIPAA Standard. Now It Has to Survive Epic.
CMS proposed FHIR as the first HIPAA standard for prior auth. Whether it changes anything depends on adoption at Epic shops, not on the rule.
Nature Medicine ranked six AIs on medicine. The test was built by the vendor whose model won, the judges were the models being evaluated, and the clinicians scoring it barely agreed with each other.
Baptist Health saved $924K on chart review with Epic’s AI coding assistant. The CDI Specialists Are Still Right but does it matter?
Microsoft is sunsetting Azure API for FHIR in three and a half months. I read it and had no idea if it applied to me. That is the alert.
Off-Label: 32% of public healthtech companies are profitable. 94% of the S&P 500 is. The bolt-on era has a math problem.
The Feed is a quick take on a few of the highest-profile health IT items of the week, with one piece from outside the beat at the close. Free through June, paid after that.
CMS proposed something on April 10 that the trade press has mostly covered as a prior auth rule, and that is the wrong frame. CMS-0062-P would make FHIR the HIPAA Administrative Simplification standard for prior authorization transactions, with an October 2027 compliance deadline, mandatory Inferno conformance testing, and three tiers of payer API oversight. The X12 278 transaction that has carried prior auth for forty years is on a sunset clock. The clearinghouse industry that has made its money translating between X12 and everything else is on notice. Brendan Keeler called it a banger of a proposed rule, and he was not wrong.
Credit where it is due. The standard finally exists. The compliance clock is real. After fifteen years of FHIR being strongly recommended, lightly encouraged, and quietly ignored, CMS picked a date and an enforcement mechanism. That is a foundation move.
Foundations matter but here is the part the trade press is not saying: this rule lives or dies on whether provider orgs actually configure Da Vinci CRD and DTR. Not because CMS says so. Epic shops will have a disproportionate effect on success or failure because Epic is where the prior auth workflow executes for the majority of hospital-based providers in this country. If Epic shops get behind it and light up Da Vinci at scale, payers will have to meet the standard. The volume coming at them will force it. If Epic shops do not, the rule joins the pile. The X12 clearinghouses keep running the practical workflow indefinitely, the way they have through every previous regulatory cycle that tried to dislodge them.
My case for skepticism stems from other initiatives like Appropriate Use Criteria, which was legislatively mandated and repeatedly delayed because implementation was technically unrealistic. It eventually was given the mercy killing it deserved. The pattern is familiar. A CMIO listserv thread from 2021 makes the awareness gap concrete. The opening line: “We didn’t even know Epic could do it.” A CMIO at a major health system, asking whether Epic had ePA functionality. The thread that followed was a collective shrug from CMIOs across the country. CoverMyMeds was clunky. The PBM routing was broken. The pilots were getting abandoned. One CMIO mentioned the HL7 Da Vinci project by name as a promising future thing. Five years later Da Vinci is about to be HIPAA mandatory, and the awareness gap has not closed at anything close to the speed of the regulatory calendar. Epic has the functionality. The UserWeb has the implementation guidance. I went looking for current discussion. Crickets.
This is not a technical problem. It is an activation problem that has been sitting unsolved for at least five years. The tipping point is not in the rule. It is in whether the next Epic fiscal year planning cycle treats Da Vinci as a project or a footnote. If a critical mass of Epic shops activates, FHIR for prior auth becomes the standard de facto and de jure, and the X12 era ends on schedule. If not, October 2027 comes and goes and the rule joins a long list of CMS interoperability mandates that existed mostly in the Federal Register.
Nature Medicine Ranked Six AIs and the Test Was the Problem
Nature Medicine published a paper on June 12 comparing three frontier large language models against two specialized clinical AI tools. The frontier models won across the board. Headlines wrote themselves. Clinical AI is dead, general models took over, procurement implications, vendor implications.
The OpenEvidence response has the right objection underneath it. The benchmarks used in the study include an exam answer key the frontier models may have been trained on. One of the rubrics was built by a vendor whose own model was among the evaluated systems. Some of the scoring used LLM judges, which is to say the models being evaluated were also part of the panel grading the responses. And the clinician-scored portion, the closest thing in the paper to a real-world signal, had inter-rater reliability the authors themselves flagged as low.
This is not OpenEvidence being a sore loser. It is the loser raising a valid methodological question, which is whether the test measured anything other than how well the test was designed for the model that won. The fuller version of this argument deserves its own piece, and I have it in my queue for next week. The short version for this week: every story celebrating which AI won is a ranking exercise on a map that keeps changing.
Baptist Health Saved $924K. The CDI Specialists Are Still Right but does it matter?
Baptist Health deployed Epic’s AI Professional Billing Coding Assistant and reported a 20% reduction in chart review time, with $924K in estimated annual labor savings. Assume the numbers are clean. Even if they have been polished, the direction is real. AI is faster than humans at the narrow workflow of reviewing a chart and producing a coding pass.
A CDI consultant pushed back on LinkedIn last week against ambient AI vendors claiming they had eliminated retrospective queries entirely. The argument was that ambient AI does not know payer-specific clinical validation standards, does not know which RAC DRG targets are active in which market, and does not know what makes a query compliant in this region versus that one. That argument is correct, and it points to a deeper problem the celebratory coverage misses.
Local knowledge in clinical coding is not in any general training set. What gets a query sustained at one payer in one market is not what gets it sustained at a different payer two states over. The CDI specialist who has been doing this for fifteen years has a pattern library that the model does not. You can automate the volume. The judgment that keeps that volume from creating audit liability eighteen months later is somewhere else.
The other problem the celebratory coverage misses cuts the opposite way. Case in point: a BCBS study showing AI coding assistants dramatically increased chargeable codes. None of this is shocking. The metrics have become the incentives and Goodhart’s Law is proving itself again.
The longer-run problem is the one almost nobody is naming. As ambient AI and AI coding assistants generate more of the documentation and coding content in any given health system, the pie will not get bigger. BCBS explicitly stated that they will start just reducing reimbursement. It is revenue cycle grade inflation. On top of that, the training data for the next generation of those tools increasingly becomes content the prior generation of those tools generated. The vendor ends up grading its own work. The benchmark becomes the vendor’s map of the territory, not the territory. This is the same structural problem the Nature Medicine paper has, one floor lower. It will get worse as ambient content scales, and the people who should be sounding the alarm are the CDI specialists whose pattern knowledge is the thing being trained over.
Microsoft Is Sunsetting Azure API for FHIR. I Read It and Did Not Know If It Applied.
Microsoft set a hard deadline of September 30, 2026 for sunsetting Azure API for FHIR, with migration required to Azure Health Data Services. Three and a half months left. The technically correct take is that this is an active infrastructure migration with a hard deadline and the organizations that have not started are behind.
I will be honest. I read this item in my news briefing and I did not immediately know whether it applied to me, to my consulting clients, or to any of the Epic shops I work with. I am a physician Epic consultant. If I do not know off the top of my head whether my organization or my clients are running Azure-hosted FHIR infrastructure, the rank and file analyst three levels below the CIO definitely does not. That gap is the actual story.
After a little digging, here is the version I would give a colleague who asked. If your organization built any FHIR pipeline on Azure in the last several years, there is a non-trivial chance it is running on the product Microsoft is shutting off in September. The replacement product exists. The migration is not trivial. The deadline is not getting moved. The person in your organization who would actually know whether you are affected is probably not the person reading vendor blogs about it. If you are reading this and you are not sure, the next ten minutes of your day is well spent asking your integration team a direct question. The alert function is the value of this item. There is nothing clever about it.
The Foundation Underneath the Week
The FHIR rule is a standard, not a foundation. The Nature Medicine paper ranked tools against a benchmark but my opinion is that its impact in the real world is close to vaporware. The CDI vendors will eventually train their AI on the AI’s own output. Every one of these stories is a layer being celebrated while the floor underneath gets less attention.
The work that would matter, in each case, is the work that gets ignored. Configuring Da Vinci in Epic. Building benchmarks that measure something other than the test designer’s preferences. Preserving the pattern knowledge of the CDI specialists before the model that replaces them is the only thing left to consult. It is boring work, it does not raise at a healthtech valuation, and it is the reason the next consensus will be more honest than the last one.
Off-Label
Off-Label. Not health IT, not my usual lane, but something that made me think. Here’s why it should matter to people who do what we do.
Sergei Polevikov used SpaceX’s IPO as a setup for a brutal piece of math. 32% of public healthtech companies are profitable. 94% of the S&P 500 is. Of 126 healthtech companies he tracked, roughly 100 are AI scribe knockoffs of each other. The market is producing margin-thin bolt-ons to platforms it does not own, and the market calls it innovation.
The point that should land for health IT people is the durability one. A bolt-on to Epic, no matter how clever, inherits the margin profile of being a bolt-on to Epic. Durable margin requires doing something nobody else can do, and the four lanes where that is possible for a third-party vendor in this market are narrow and getting narrower. If you are in healthtech right now and you are competing on speed of execution against Epic on terrain Epic eventually arrives at, you have a window. The window is not a moat. Worth the read.
John Lee is an emergency physician and Epic consultant who helps health systems bridge the gap between Epic’s capabilities and operational reality. He specializes in data architecture, registry optimization, and making Epic’s tools actually deliver results.
If you need help configuring your Epic environment to support these capabilities, connect with him on LinkedIn or via his website.