The Feed: June 25, 2026
AI Coding Tools Were Supposed to Fix Revenue Integrity. Payers Say They're Inflating It.
AI coding tools are capturing diagnoses that are technically correct but clinically insignificant, and payers are now formally blaming them for rising costs
Josh Mandel proved a coding agent can reconstruct Epic’s FHIR output from a single patient’s EHI export, and the data it unlocked goes beyond what Epic’s API exposes
ChatGPT Health is free, handles 230 million queries a week, and Epic’s competing “digital concierge” does not exist yet
Rhode Island’s new AI scribe notification law is one sentence long, and that is exactly the problem
Medicare’s AI prior authorization pilot is producing the errors and delays it was supposed to eliminate, and the contractors want you to know as little as possible about how
Off-Label: A bank’s AI agent approved a $1.4 million credit line at 2:47 AM with no human watching, and nobody knew until the morning standup
The Feed is a quick take on a few of the highest-profile health IT news items of the week, with one piece from outside the beat at the end. The Feed is free through June, then moves to paid.
A Blue Cross Blue Shield audit found something that should worry every health system deploying AI coding tools. At hospitals that adopted AI-assisted coding fastest, the rate of patients coded for acute posthemorrhagic anemia jumped from roughly 4% to over 12% in under three years. Transfusion rates did not move. Chart review at the worst-spike hospital found fewer than one in five of those coded cases actually met the diagnosis.
The AI did not hallucinate. It did not malfunction. It captured a finding that was technically documentable. A patient who bled significantly during delivery, whose hemoglobin dipped, whose lab values briefly crossed a threshold. Ten years ago, the analog brain of the physician on the case would not have coded that as a discrete diagnosis. The level of anemia would not have prompted any acute intervention beyond recommending iron after discharge. It was clinically real and clinically insignificant, and nobody would have made note of it.
The AI made note of it. The physician signed the note. The diagnosis entered the permanent record.
Andrew Lundquist, a physician and technologist, put it plainly: he has done the thing that lets this happen. So have I. Tuesday afternoon, fourth note of six, the system drafts it, it looks right, he signs. All of us do. We do not have the bandwidth to meticulously review every detail. That is the reality, and the AI is operating inside that reality at scale.
The same week, a PwC survey found that nearly 70% of health plans now cite provider AI documentation and coding tools as a top-three driver of rising commercial healthcare costs. The mechanism is the one the BCBS audit just illustrated. AI-assisted coding captures complexity that was always technically present but never documented under the old workflow. Claim values go up. Payers see inflation.
In the last issue of The Feed, Baptist Health reported $924,000 in annual savings from AI-assisted coding, and the observation was that CDI specialists were right about the part that number does not measure. This is the other shoe dropping. The part that number does not measure is now being measured by payers. Both sides are correct. The AI is capturing what was always there, and the billing infrastructure was never designed for technically-correct-but-clinically-insignificant capture at scale. The legacy coding, billing, and level-of-service system was built on a foundation of manual curation and human judgment about what was significant enough to document. The AI does not have that significance filter. Nobody built one, because nobody anticipated needing it.
A Coding Agent Just Reconstructed 90% of Epic’s FHIR Output From One Patient’s Data
Josh Mandel, Chief Architect for Health at Microsoft Research, gave a coding agent a single patient’s EHI export file paired with the same patient’s FHIR output. The agent wrote translators that reconstruct roughly 90% of Epic’s FHIR API output as plain, reviewable code. Then it went further. It built FHIR resources for billing data, financial records, and MyChart messages that Epic’s standard API does not expose at all.
The premise is that traditional interoperability infrastructure is rigid and deterministic. What Mandel built is semantically adaptive. It understands. Or at least it understands better than a sterile ICD-10 code. The agent interprets the source data and generates mapping code that improves with each additional sample. A handful of community-contributed EHI exports could seed a mapping pipeline that dramatically expands data portability without Epic’s participation.
Epic’s FHIR API surface is scoped around operational utility and regulatory minimums. The data flows but there are gaps in meaning. The EHI export, which federal law requires Epic to provide, forces the raw database tables into a computable format. The letter of the information blocking rules allows dumping raw tables. Lack of standardization renders the data functionally useless, at least until now. Mandel just proved that open-source tooling bridges that gap. The bottleneck is no longer Epic’s (or any other vendor’s) willingness to build a semantically rich FHIR endpoint. It is the community’s speed in building parsers for Epic’s publicly documented EHI table structures.
I wrote previously about creating a USCDI data type that could carry markdown, a vector to deliver both volume and quality at scale. Mandel’s project is the demand-side proof that the translation is automatable. The supply-side infrastructure to match it is the piece that still needs building.
ChatGPT Health Is Free. Epic’s Answer Doesn’t Exist Yet.
OpenAI launched ChatGPT Health on June 18 as a free consumer health intelligence product. Over 260 physician evaluators across 60 countries. 230 million health queries per week. A 71% improvement in factuality over the base model. Epic CMO Jackie Gerhart responded publicly with Epic’s own vision: a “digital concierge” tethered to the patient’s actual clinical record. Features on her roadmap slides included dates labeled “November 2025,” “November 2026,” and one that simply read “Future.”
The competitive question is not model quality. It is who owns the patient’s first interaction. OpenAI is building the consumer AI front door that never requires the patient to pick up the phone. Epic is building something that will require entry into their home renovation project with lots of “Future” development promised.
If you are doing an addition on a house, you do not add willy-nilly. You make sure the addition flows and that the underlying plumbing and electrical work properly. That is Epic’s approach. On the other hand, if you are a third party like ChatGPT, you can park an RV on the property easily, but you then have to deal with crossing the yard and living in two places at once. ChatGPT Health is the RV. It is fast, it is free, and it operates outside HIPAA. I have not seen the EULA or privacy terms, but when a consumer health product is free, the business model question answers itself.
Also remember that Epic puts nearly 40% of its revenue into R&D. Because they are privately held, they can make long-term investments without catering to quarterly shareholder expectations. “Thoughtfully developed” likely means slower. Faster with a third party does not mean better. Where do you want to place your bets?
Rhode Island’s AI Scribe Law Is One Sentence. That Is Exactly the Problem.
Rhode Island’s General Assembly passed S2570 on June 11, and Governor McKee signed it June 22. Healthcare IT News called it an “ambient AI scribe opt-out law.” It is not. The full operative clause:
“Any and all healthcare providers and healthcare facilities that employ artificial intelligence (’AI’) to document in-person or telehealth visits shall notify patients of the use of AI for that sole purpose and review the AI-generated documentation for accuracy after the visit.” (https://webserver.rilegislature.gov/BillText26/SenateText26/S2570A.pdf)
That is the entire law. No opt-out mechanism. No specified form of notification. No timestamped attestation requirement. No audit trail. No enforcement mechanism. No penalties.
It is also one sentence of good intentions with the potential to become a regulatory morass. Remember that the Meaningful Use provisions in the ARRA legislation amounted to a few sentences. Those turned into thousands of lines of regulations, years of attestation burden, and a compliance industry that consumed more clinician time than it saved. This bill is the same pattern at the starting line. A vague mandate, good intentions, no operational guidance, and a guarantee that compliance departments will over-engineer the requirements because the statute gives them nothing to calibrate against.
I do not think this spreads to other states, because it is so wildly unreasonable in practice that any state health system association will fight it. But the broader theme holds: legislative virtue signaling that sounds like patient safety and creates operational gridlock is a pattern that predates AI and will outlast it.
Medicare’s AI Prior Auth Pilot Is Producing the Errors It Was Supposed to Eliminate
CMS launched WISeR, the Wasteful and Inappropriate Service Reduction Model, in six states in mid-January. AI-powered prior authorization for 13 medical services in traditional Medicare, where prior auth had historically been rare. The vendors building the system acknowledged an “aggressive rollout from the time of being notified to going live.” Providers called it “horrendous.” A 65-year-old Oklahoma cattle rancher who drives 2.5 hours each way for quarterly spinal epidurals was told he now needs preapproval, adding extra trips and weeks of delay to a procedure he has received for years.
In the last issue covering CMS-0062-P, the point was that the FHIR-based electronic prior authorization standard is the right infrastructure move. WISeR skipped the infrastructure. It went straight to AI-assisted coverage determination without transparency about where the data comes from, what algorithms drive the approvals or denials, or any accountability mechanism when the system gets it wrong. Where there is mystery, there is margin, and these third-party contractors want to incorporate as much mystery as possible. The House Appropriations Committee has already moved to block CMS funding for the program.
A practicing physician in Oklahoma put the stakes plainly: everybody knows that if this pilot shows savings, it will expand to all procedures. What isn’t getting measured is patient harm.
The through-line this week is the same one it always is. AI is arriving faster than the systems it touches can absorb it. Coding tools that capture what was always technically documentable are colliding with billing infrastructure designed for human judgment. A one-sentence law is about to become a compliance regime. A prior auth pilot skipped the plumbing and went straight to the gate. And a coding agent just proved that the interoperability bottleneck was never the technology. In every case, the question is the same: who is doing the unglamorous work to make the foundation match the speed of the layer being built on top of it?
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.
At 2:47 in the morning, an AI agent at a top-10 U.S. bank approved a $1.4 million commercial line of credit with no human in the loop. Nobody knew until the morning standup. The decision was entirely correct. The borrower met every criterion. The agent got rolled back anyway, because the people assigned to supervise it had quietly stopped looking. Russ Pearlman’s commentary on the piece nails it: the override rate started near 8% and drifted below 2% over ten weeks. The institution celebrated it as the agent getting smarter. It was not. The supervisor was catching less. We instrument the agent to the decimal and the supervisor not at all. The agent gets a dashboard. The supervisor gets a chair. This is a cautionary tale for health systems deploying Epic’s Agent Factory or any agentic clinical workflow. They will be running the same experiment soon at scale, and I doubt organizations have budgeted for the governance and oversight.
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 (https://www.linkedin.com/in/johnleecmio/) or via his website (https://www.hitpeakadvisors.com/).