Cedars-Sinai puts OE in Epic 🏥, AI scribes fail the equity test 🤐, Healthcare costs spike 7.9% 📈
Cedars-Sinai Puts OpenEvidence Inside Epic
Cedars-Sinai announced a systemwide deployment of OpenEvidence, integrated directly into Epic.
Not a pilot. Not a “check it out if you want” link. An enterprise deployment where the AI reads the active EHR view and contextualizes medical literature to the specific patient in the chart.
Cedars-Sinai plans to layer its own care pathways and protocols on top of OpenEvidence’s medical literature search. Patient data is used only for contextual queries and never stored by OpenEvidence or used for model training. This follows Mount Sinai’s enterprise-wide Epic integration from April.
😤 “This is just a fancy UpToDate replacement.” You should give it a try then and see for yourself.
😤 “Two health systems is a trend piece, not a movement.” Fair. But 650,000 physicians using it without those health systems is the movement. The institutions are catching up to their own clinicians. That’s the story.
⁉️ “Their own practice guidelines get put in there?” I think that is the interesting part too, along with management of staleness of those guidelines.
[BTW: I didn’t realize Daniel Nadler wrote a book of imagined ancient love poems until yesterday … hard to know everything that is going on]
📡 Builder’s Radar
A Family Doctor Pauses the AI Scribe — and Exposes the Equity Gap Nobody Tested
Dr. Gigi Magan published “Voy a Pausar” this week — an essay about deliberately pausing her ambient AI scribe before sensitive conversations with undocumented Spanish-speaking patients.
The patients self-censor when recording is present. Speech recognition has structurally higher error rates on non-native English. Consent forms are in English. Research on ambient scribes in FQHCs is essentially nonexistent.
A Nature Digital Medicine review this month confirmed the gap: limited functionality with non-English speakers is a documented barrier to ambient scribe scaling, and the equity gap between high- and low-resource organizations is widening.
😤 “Speech recognition is getting better — this will fix itself.” It won’t fix itself if patients stop talking. The error isn’t in the transcription. It’s in the disclosure. A patient who withholds their domestic situation or their medication non-adherence because a recording is running — that’s a degraded clinical record, not an improved one.
💡 80/20: If you’re building or deploying ambient scribes, build a pause workflow. Not “turn it off and back on” — a deliberate, visible pause that signals to the patient: this part is just us.
Healthcare Costs Hit $37,824 per Family — 7.9% Spike Is the Highest in a Decade
The 2026 Milliman Medical Index dropped this week. Healthcare costs for a family of four rose 7.9% to $37,824 — the highest annual increase in more than a decade.
Outpatient facility costs rose 7.5%. Pharmacy costs rose 14.8%. Those two categories account for 69% of the increase.
Buried in the report: Milliman analysts are watching AI-enabled billing, which they say could influence hospital cost growth by capturing services that weren’t previously billed.
😤 “AI billing optimization raises costs?” It doesn’t create new services — it captures existing ones that were missed. Whether that’s “raising costs” or “billing accurately” depends on which side of the claim you’re sitting on.
💡 80/20: The 14.8% pharmacy increase is the number to watch. GLP-1s, specialty drugs, and site-of-care shifts are driving it. If you’re building anything adjacent to medication management or PBM workflows, the cost pressure is your tailwind.
Oura’s Healthcare Pivot: 6x Engagement in Medicare Advantage
Oura’s partnership with Essence Healthcare is producing engagement numbers that should make every digital health company uncomfortable. 77% of MA members wearing Oura rings check in five or more days per week. That’s 6x the engagement of prior MA wellness programs.
Cigna is now partnering with Oura on preventive health and chronic disease management. Oura has 40 PhDs and MDs on staff, an FDA-pipeline hypertension study, and a Dexcom CGM integration for metabolic health.
This is a consumer wearable that quietly became a clinical-grade engagement platform — and payers are noticing.
💡 80/20: The 6x engagement number is the wedge. If you’re building anything in the MA engagement space, benchmark against Oura’s 77% five-day-a-week figure. That’s the new floor for what a payer considers “engaged.”
AI Gave the Wrong Diagnosis — and the Model Wasn't the Problem
John Lee this week: two AI diagnostic tools (OpenEvidence and Epic’s Cosmos Diagnosis Advisor) were used on the same ED patient. Both gave plausible differentials. Both were wrong. The right answer came hours later from an ICU physician who pulled a piece of patient history nobody had documented yet — so neither tool could have caught it.
The metaphor: AI is an astrolabe. The instrument works, the math works — but the stars (the medical record) are blurry, drifting, and incomplete.
“The ICU physician who got my patient’s diagnosis did it by asking a question nobody else had asked. No model trained on what is already documented can do that, because the answer relies on data and documentation that doesn’t exist yet. The answer had to be created.”
❓What do product do you think lies adjacent or between the blindspots of OE and Cosmos Diagnosis Advisor? I think there is something, but I can’t put on finger on it yet.
Nvidia Posts $81.6B Revenue, $58.3B Profit — AI Chip Boom Continues
Nvidia’s Q1 FY27 earnings: $81.6B revenue (up 85% YoY), $58.3B profit. Data center revenue hit $75.2B. Next quarter guidance: $91B. The compute substrate that clinical AI runs on is not slowing down.
OpenAI IPO Filing Reportedly Imminent
Multiple outlets report OpenAI could file its IPO prospectus before the week is out. Separately, Anthropic is reportedly hitting $559M operating profit this quarter with compute costs dropping from 71¢/$ to 56¢/$. The foundation model layer is turning profitable — which means the clinical AI tools built on top have a more stable substrate than the “AI companies are burning cash” narrative suggests.
🎙️ From the Pods
Guest Loren Adler from Brookings breaks down the unintended consequence: the No Surprises Act’s arbitration process may be anchoring out-of-network rates higher than pre-Act levels in some markets. The law fixed the patient’s bill but may have inflated the system’s bill.
💡 Builder take: If you’re building price transparency or cost estimation tools, the No Surprises Act arbitration data is now a source — and the gap between “patient-facing price” and “system-level cost” is a product opportunity.
🎙️ Health Tech Nerds Radio — “How Gyde Is Enabling the Shift from MA Broker to Trusted Advisor”
Will Johnson of Gyde on how AI is reshaping the Medicare Advantage broker model — from transactional enrollment to longitudinal advisory. The insight: the broker who retains a relationship with the member post-enrollment captures retention value that the current model throws away.
💡 Builder take: MA enrollment is a $3B+ commission market. The “broker as ongoing advisor” model is an AI-enabled workflow no one has nailed yet. The retention economics make the build math work.
What are you building this week? Reply and tell me — I read every one.
— Kevin