THE FEED: June 8, 2026
I’m starting a newsletter. It is a quick take on a few of the highest-profile news items in health IT this week, with one piece from outside the beat at the end. The Feed is free through the month of June and moves to a paid subscription after that.
Epic’s AI agents are live in production, and your ungoverned data will quietly decide whether they work
HTI-5 is a mess, and anyone selling you a clean answer on it has a reason to
The AI scribe is the toe in the door; the discrete data it could capture is the real prize
Epic still wins the market, but the 40% of buyers who vanished are the real story
Fifty billion dollars is about to test whether rural health can be fixed with money
Off-Label: we spent a century learning to trust medical evidence, and now we’re automating its fabrication
Somewhere in your Epic environment, there are probably four different definitions of sepsis. One was built for a quality project a few years back. Another came out of a sepsis bundle initiative the year after. A third was created for a one-time research pull and never touched again. They overlap in places, contradict each other in others, and nobody has reconciled them, because reconciling them was never anyone’s job. These are grouper records, the building blocks Epic uses to define a cohort of patients: everyone with sepsis, everyone who had a given procedure, everyone who belongs in a registry. They are mundane and boring. So are thousands of other data assets. They are also about to decide whether your AI works.
At HIMSS, Epic reported that its autonomous agents are live in production (https://himsstv.brightcovegallery.com/detail/videos/digital-checkup/video/6341328679112/epic-s-evp-of-r-d-reports-on-the-ehr-giant-s-success-with-ai). Penny handles prior authorization, Art generates ambient documentation, Emmie manages scheduling. Agent Factory, Epic’s platform for building your own agents inside the Epic environment, is on the near roadmap. The coverage was breathless. The one sober note, buried at the end, was an analyst warning that deploying AI on a poorly optimized build will replicate go-live failures at AI speed.
That warning is the whole story, and it keeps getting eclipsed by shiny-object panacea talk. Agents do not reason about your patients in the abstract. They work by cohorting, by diagnosis, by procedure, by conceptual grouping, and in Epic that means they lean on your groupers. An agent asked to act on septic patients has to know which patients are septic, which means it has to choose one of your four sepsis definitions. Which one does it use? You may not know. Worse, you may not know that you do not know.
This is why Agent Factory will amplify the organizations that did the boring foundational work and quietly punish the ones that did not. Build on governed, deduplicated, well-reasoned groupers and the factory compounds your advantage. Build on the ad hoc sediment most organizations are actually sitting on, and you will spend months, maybe years, wondering why your agents underperform, never quite connecting it to the unglamorous governance you skipped. The agents are not the foundation. They are the floor you lay on top of it, and a floor is only as level as the ground beneath it.
I can’t decipher HTI-5. Anyone Selling You a Clear Answer Has a Reason To.
ONC published its HTI-5 proposed rule this week, and it does several large things at once. It removes 34 of 60 health IT certification criteria. It extends information blocking to cover automated means, including autonomous AI systems. It tightens the Manner and Infeasibility exceptions that vendors have used to wall off third-party access. AHIMA promptly warned that stripping out the C-CDA criteria, the standard for exchanging clinical documents, could let vendors start charging for document exchange that is free today.
I am not a health information management expert, and I will not pretend the fee question is settled. I think AHIMA’s concern is valid. My reaction to this rule is that I can’t figure out which side I’m on, even as someone living in the weeds. Is streamlining 34 criteria a gift or a hazard? It depends entirely on whether your organization runs on the legacy those criteria were quietly holding together. How much should we let technical debt pull us back toward the past, versus enforce a cleaner standard going forward? As a non-HIM informaticist, I do not have a good answer, and that is the point. Anyone telling you HTI-5 is plainly good or plainly bad has either oversimplified it or has money riding on your believing them.
The Scribe Is the Toe in the Door. The Discrete Data Is the Building.
A multisite study in JAMA Network Open found that ambient AI scribes produced modest but real reductions in documentation burden and after-hours charting, with no major hit to appointment volume. It is the most rigorous evidence yet that scribes do something at scale rather than just in a vendor demo. Read the methods, though, and the conclusions rest on a fair number of assumptions and a lot of normalization smoothing in how the team handled the metadata. “Modest but real” is a meh statement to me.
The burden relief is not the interesting part anyway. In my opinion, note creation is the killer app, the toe in the door, the thing that gets these tools adopted in the first place. The real prize comes later, when ambient vendors figure out that the real value lies in scraping discrete, structured data off the patient-physician conversation itself. The things our coding vocabularies handle badly, social determinants of health, early symptomatic surveillance, the signal that never makes it cleanly into ICD or SNOMED or CPT, are exactly the things a good ambient layer could surface and document discretely. A better-sounding note is nice. Semantically rich, ubiquitous, structured data will change healthcare.
Epic Is Still Winning. The Buyers Who Vanished Are the Real Story.
Epic added a net 77 hospitals last year while Oracle Health lost 56, and KLAS now puts Epic at 43.7% of the acute care market, with Community Connect, its model for extending one organization’s Epic to smaller affiliated hospitals, still the growth engine into sub-200-bed facilities. The dominance is not news. The 40% drop in overall EHR purchase decisions is. That decline was driven by a shift toward faster-ROI AI investments, which is to say organizations are buying AI instead of doing the foundational EHR work, and there is far more of that work left undone than anyone wants to admit. It is the grouper problem again, this time written into the budget.
I have a more hopeful thread. Organizations like OCHIN show another path for the small hospitals that cannot stand up an army of analysts: a federation of members all running Community Connect and governing it together, Epic without the overhead price tag. OCHIN does yeoman’s work, and it is now branching into acute care inpatient offerings. For a large segment of the delivery system, that model may matter more than anything Epic announces from a main stage.
Fifty Billion Dollars Is About to Test Whether Rural Health Can Be Fixed With Money
All 50 states received awards under the $50 billion Rural Health Transformation Program, and a majority of state proposals target data sharing, interoperability, and EHR upgrades, with several building new or modernized health information exchanges. Rural and critical access hospitals are where technology has most plainly left an important part of the delivery system behind, so I want this to work.
I am skeptical that it will, at least without serious bumps. My observation is that money like this gets mired in bureaucracy and middlemen before the foundation ever changes. The organizations with the least capacity to govern a build are the ones about to receive checks to expand one, which is how you recreate the four-definitions-of-sepsis problem at fifty-state scale. The OCHIN federation model, and others like it, is the most plausible way the money reaches the ground without every tiny hospital needing its own analyst team. Epic has at least paid lip service to this with its Garden Plot initiative, though I do not know of an organization that has actually executed on it yet. The legal analysis attached to the program flags HIPAA and data governance as upfront risks, and that is the tell. Governance is being treated as a footnote, when it should be the whole book.
Five stories this week, and they are five versions of of the same theme. The agents, the deregulation, the scribes, the market shift, the fifty billion dollars: every one of them is a shiny layer set on a foundation almost nobody is being paid to build. The industry has gotten very good at funding the floor and very bad at leveling the ground. The organizations that figure out the order of operations, ground first and floor second, are going to look like geniuses in three years. They will just be the ones who did the boring work while everyone else was talking about agents.
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.
Eric Topol sat down with Helen Pearson to discuss her new book Beyond Belief, which traces over eighty years of medical evidence, from eminence and expert hunch to the evidence-based medicine movement that got its name at McMaster in 1991. The conversation runs through Dr. Spock, SIDS, the Cochrane Collaboration, the failures of evidence during COVID, and a sobering note on AI’s capacity to manufacture convincing false evidence at scale. Interestingly, Dr. Topol noted (~20 min mark) how often his patients on statins complain about muscle pain. I thought, “Is captured? Would an ambient solution help with this?” We spent a century painfully learning how to know what is true in medicine and we still struggle. We are now building tools that can fabricate the appearance of truth faster than we can check it. Anyone deploying AI on clinical data should sit with that tension before the next demo, not after.
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.
"That warning is the whole story..."
Could not agree more. If the only thing worse than no data is bad data, then the only thing worse than bad data has got to be AI layered on top of it.