Healthcare policy architect in the Senate out 🗳️, OwnChart gives patients their own AI 🏥, More Epic gains 📈
Healthcare’s Policy Architect Gets Voted Out
Senator Bill Cassidy lost his Louisiana Republican primary on Saturday — the first sitting Republican senator ousted in a primary in a decade. He finished third, behind Trump-endorsed Rep. Julia Letlow and state treasurer John Fleming.
Cassidy is a gastroenterologist. He’s not a “healthcare senator” the way most people use the phrase. He’s the person who actually understood how Medicare Advantage risk adjustment worked, why the No UPCODE Act mattered, and what happens to 940,000 Louisiana Medicare enrollees when you restructure supplemental benefits.
He was arguably the most policy-literate healthcare voice in the Senate.
The healthcare implications are specific. The No UPCODE Act — Cassidy’s bill targeting Medicare Advantage upcoding — dies with his seat. More than 60 organizations had already mobilized against it, warning it would undermine care coordination and reduce in-home health assessments. Whatever you thought of the bill, it was a technically serious attempt to address a real problem.
The HELP Committee loses a physician member who could engage with clinical AI governance at the level of actual regulatory text, not talking points.
😤 “This is just politics, not health tech.” Every health tech builder’s runway is shaped by Medicare policy, prior auth rules, and FDA guidance. The person who wrote those rules being replaced by someone who hasn’t read them changes your regulatory environment.
😤 “Someone else will pick up the healthcare wonk mantle.” Maybe. But institutional knowledge in the Senate isn’t fungible. The learning curve on Medicare Advantage risk adjustment alone is years, not months.
💡 80/20: If you’re building in Medicare Advantage, home health, or any space that depends on congressional policy literacy, your risk model just changed.
MACPAC Calls for Human Oversight of AI Prior Auth Denials in Medicaid
The Medicaid and CHIP Payment and Access Commission recommended new federal guidance making clear that automated systems cannot independently deny care in Medicaid.
The recommendation requires a qualified human reviewer to approve any denial or reduction in requested services, even when AI tools are used in the review process. MACPAC says states and insurers are already using automation in prior auth, but regulators have limited visibility into how those systems work.
If you’re building prior auth AI for Medicaid plans, the “fully automated denial” product just hit a regulatory wall.
😤 “Human-in-the-loop requirements kill the whole point of automation.” They kill the wrong point. The value of prior auth AI isn’t in removing humans from denial decisions — it’s in surfacing the right information so the human reviewer makes a faster, better decision. Build the decision-support layer, not the auto-deny engine.
💡 80/20: MACPAC recommendations aren’t binding, but they signal where CMS is heading. If your product’s value prop depends on fully automated denials, pivot now. The winning architecture: AI does the work, human signs the decision.
Epic Gains 77 Hospitals While Oracle Health Loses 56
A KLAS report published May 14 shows Epic gained 77 acute care hospitals in 2025 and now holds 43.7% of hospitals and 56.9% of beds. Oracle Health posted its third consecutive year as the largest net loser — 56 hospitals, 14,676 beds shed.
65% of Oracle Health customers report they are either leaving or “considered vulnerable.”
💡 80/20: If you’re building clinical AI integrations, Epic is the platform that matters. Two-thirds of beds, growing.
You Can’t Retrofit KPIs to Incentivize AI Adoption
A sharp take: the actual advantage AI-native companies have is hiring people who are innately curious about the tools and play with them. You can’t measure your way into that culture.
💡 80/20: Stop building the AI adoption dashboard. Start hiring the person who already uses the stuff.
🛠️ From the Workbench
OwnChart: Open-Source Patient-Side EHR-to-LLM Warehouse
Nick Dawson — former hospital president and former president of the Society for Participatory Medicine — launched OwnChart, an open-source platform that lets patients pull their full record from Epic, athenahealth, and Oracle Health, layer in HealthKit and wearable data, and query all of it with an LLM.
His use case: hours after outpatient eye surgery, he fed the intraoperative note into Claude, which surfaced that his surgeon had improvised mid-operation in a way he would have otherwise missed. The next day he complimented the surgeon on the creativity.
Currently in alpha. GitHub repo here. Built on Hugo Campos’ OpenKP project and the Critical AI Health Literacy skill.
⚠️ Verify: Alpha software. “For the curious, the technical, and the brave.” Patient-facing tools touching real health data require careful data handling. This is localhost and self-hosted — no cloud PHI exposure — but review the architecture before pointing it at real records.
😤 “Patients don’t want to query their own records.” Most don’t. The ones who do are the early adopters who drive product feedback. Nick’s framing — “data without the means to understand it is paperwork” — is the patient.dev thesis in one sentence.
😤 “This is just a niche tool for technical patients.” Every patient portal started as a niche tool for technical patients. The question is whether the architecture scales, and FHIR-based record pulling from Epic/athena/Oracle is the right foundation.
What are you building this week? Reply and tell me — I read every one.
— Kevin