PITFALLS.md 👀, Oracle scrubs into the OR 🔪, Assort banks $120M 🏦
Oracle Health scrubs into the OR with Theator
Oracle Health is extending ambient AI into the operating room, partnering with Theator to capture surgical video, cross-reference it against the EHR, and auto-generate a structured operative report before the surgeon leaves the room.
Theator’s “Surgery-to-Text” engine has already run on 600,000+ procedures across 150 operation types, deployed at Mayo Clinic and UHealth Miami.
The scribe didn’t stay in the clinic note — it walked into the OR, where the documentation has always been the surgeon’s least favorite part of the case.
😤 “Surgeons already dictate fast. Who needs this?” The value isn’t speed, it’s the structured, billing-optimized, video-grounded record — the op note that actually matches what happened on the table, not what the surgeon remembered an hour later.
💡 80/20: Ambient documentation is migrating procedure by procedure. If you build clinical tools, watch where the next “highest-friction note” lives — endoscopy, cath lab, IR — because that’s where the next Theator-shaped wedge opens.
Mount Sinai makes “AI governance” its own purchase
Mount Sinai is deploying Signal 1’s AI Management Platform to centralize governance, performance monitoring, and ROI tracking across roughly 120 AI tools now in use or under evaluation.
When a health system has 120 models running, “is this one still working?” stops being a question you answer per vendor and becomes a platform you buy.
The observability layer is becoming its own procurement category — the monitor is turning into the product.
😤 “This is just a dashboard with a press release.” A dashboard that tells a CMIO which of 120 models silently drifted last quarter is not nothing. Calibration rots while accuracy holds — somebody has to watch that, and “somebody” is now a line item.
Assort Health banks $120M to run the patient journey
Assort Health raised a $120M Series C at a $1.2B valuation (Menlo Ventures) to scale agentic AI across scheduling, intake, and referrals — landing the same week competitor Prosper AI took $30M from a16z.
Two nine-figure-ish rounds in one week for “AI runs the front office” is a land grab, not a coincidence.
😤 “Another voice-agent raise. We get it.” The tell isn’t the raise, it’s the category consolidating — patient access is becoming a platform fight, and the losers become features inside the winners.
Medicare’s AI is also learning to say no
The other half of the same capability: KFF Health News reports that CMS’s WISeR model — AI-assisted prior auth now live in traditional Medicare across six states — is snarling patients and doctors in errors and delays. Oklahoma’s vendor, Humata Health, says 88% of clinically-supported requests get an immediate yes; clinicians on the ground call the rollout “horrendous.”
The same agentic engine funded to watch a chronic panel for shared savings is, on the other side of the payer wall, being funded to deny the claim.
😤 “88% auto-approved sounds good, actually.” It’s the other 12% — and the speed of the denials — that lands a real patient in a real delay. “Mostly right at scale” is exactly the failure mode every deployed clinical agent shares.
💡 80/20: If you build anything that touches utilization, know which side of the wall your tool arms. The clinician-builder edge is reading both the workflow and the incentive — and being honest about which one your model is actually optimizing.
Quick hits
Utah’s Doctronic standoff turns collaborative — Modern Healthcare reports the fight between Utah’s medical board and Doctronic’s autonomous-AI care pilot (which renews prescriptions under human review) is moving from confrontation to a shared framework — an early read on how states will regulate AI that makes clinical decisions.
OpenEvidence vs. Nature Medicine, round two — the June 12 study showing generalist frontier models beat specialized medical tools has a new development: OpenEvidence reportedly demanded a retraction and apology. The supplementary reviewer comments are worth reading — the paper got better under “Reviewer #2,” which is the quiet case for peer review when a $20 commodity answer keeps beating the specialized RAG.
🛠️ From the Workbench
The productivity pattern: build narrow “loops,” and pair every skill with its scars
Nate Jones argues the invisible work in your day is the wiring between apps — remembering, checking, following up — and the move isn’t one all-knowing assistant, it’s small recurring loops, each with memory, sources, safe actions, and hard boundaries (the household loop that drafts the message but stops before sending).
Pair it with a pattern making the rounds: for every SKILL.md you write for an AI tool, keep an append-only PITFALLS.md — Trigger / Wrong behavior / Correct behavior — so the tool stops repeating the same mistake. The skill file says what to do; the pitfalls file is the institutional memory of what went wrong.
😤 Haters
“This is generic productivity advice with a healthcare coat of paint.” Fair on the loops framing alone. But the PITFALLS.md pattern is exactly the missing artifact in clinical AI deployment — the place where “the model omitted the penicillin allergy that one time” lives so it never happens silently again.
“I’m an employed physician. I can’t run agents on anything real.” Correct — and you don’t need to. Build the loop on your own non-clinical wiring (the manuscript-status checks, the meeting prep) on synthetic or personal data. You’re learning the harness, which is the part that transfers.
🎙️ From the Pods
🎙️ Vital Signs — “Chris Altchek on the Future of Chronic Disease Care”
The Cadence founder’s sharpest line: chronic disease management is automatable because it’s rules-based — clear rules are what let AI actually deliver care instead of just summarizing it. It’s why he started there and not in the messy middle of diagnosis.
💡 Builder take: Pick the rules-based corner of your specialty first. The automatable wedge is wherever the protocol is already written down.
🔇 Speaker Blindspot: Survivorship/selection bias — “chronic disease is rules-based” is true right up until the heart-failure patient with a potassium of 7.2 and a reason the protocol doesn’t cover. The exceptions aren’t noise; they’re where the clinician still lives, and “rules-based” quietly assumes them away.
🎙️ Podnosis — “Badges, Bots and Blind Spots”
Imprivata CMO Sean Kelly on the next access-management frontier: non-human identity. As AI agents start acting inside hospital systems, each one needs its own credentials, its own audit trail, and zero-trust monitoring — an agent logging in “from the hospital and from China at the same time” should get blocked like any impossible traveler.
💡 Builder take: This is the boring prerequisite under Claude Tag and every clinical agent — give your agent its own identity with an audit log, not a borrowed human login. Accountability starts at the credential.
🔇 Speaker Blindspot: False dichotomy + vendor framing — Kelly declares the old “security vs. usability” tradeoff a “broken paradigm,” which is conveniently also his company’s sales pitch. The tension is real; whether one access-management vendor dissolves it is the claim to verify, not assume.
🎙️ Relentless Health Value — “EP517: The Business of Prior Auths”
Stacey Richter walks the 401-level mechanics: a cheaper, better generic can land in prior-auth purgatory with a higher copay — not for clinical reasons, but because the PBM’s aggregate-rebate math favors the “brand darling,” and a small new entrant can’t out-rebate millions of existing scripts. The “rebate cliff” is a contracting artifact wearing a clinical mask.
💡 Builder take: If you build prior-auth tooling, model the rebate incentive, not just the clinical criteria — the “no” you’re automating around is often financial, not medical.
🔇 Speaker Blindspot: Composition fallacy risk — it’s one carefully chosen case study generalized to “a lot of examples,” with no PBM voice in the room to contest it. The mechanism is real and well-argued, but a solo-narrator episode is a hypothesis with great production value, not a controlled finding.
💡 BTW
💡 BTW: Tamir Wolf, the surgeon who founded Theator (today partnering with Oracle Health to put AI in the OR), was a trauma surgeon in the Israeli Navy SEALs — decorated with a Medal of Distinguished Service for treating wounded teammates under fire. He started the company after his wife and his former boss both got appendectomies in New York hospitals the same week, with opposite outcomes; the variability was the whole thesis.
📅 Upcoming: today (Thu Jun 25, 3 PM ET) — Adoption of AI in Clinical Care: Updates from the HHS RFI (ONC/HHS), the policy signals you’ll be designing around.
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin