Diego Orofino

The 27% is heavily concentrated in functions sitting on top of already-structured data. Claims edits, scheduling, revenue cycle. The functions still stuck in pilot land are the ones where the underlying source is a PDF, plan documents, PA criteria, formulary tier logic. Structure precedes scale.

27% of healthcare organizations are now deploying AI at scale across multiple functions. And 50%+ are running active AI pilots. A year ago that number was a rounding error. The pilot era is over. #HealthcareAI #HealthIT #DigitalHealth healthcareittoday.com/2026/04/26/bon…

That headcount is the real CMS-0057 ROI argument, not the regulatory deadline. The labor exists because medical-necessity criteria still live in PDFs. Da Vinci CRD/DTR/PAS only collapse the workflow once the upstream plan documents are themselves machine-readable. Otherwise you just move the manual step.

The 20-to-1 ratio in hospital settings versus physician-run practices is a useful gut check, but the number that haunts me from my own reporting is this: a single academic medical center employs 60-80 people solely for prior authorization, at $70-90K fully loaded cost per head. One function. One hospital. Around $6-7 million annually before you touch anything clinical. That ratio gap you're describing is partly a staffing problem and partly a billing complexity problem, and those two causes have very different solutions. Recruiting fixes neither. Where it gets strange is that administrative workers are only 20-25% of hospital FTEs total, so even a complete administrative automation wave leaves the larger labor cost problem completely intact. The 75-80% who perform irreducibly physical work, moving patients, transporting specimens, managing environmental services, don't appear in the efficiency conversation at all, even though structural nursing shortages and $11.6 billion in 2022 travel nurse spend are the numbers actually threatening solvency for nonprofit systems running at 1-3% margins. The chart may be imprecise, but the instinct behind it is tracking something real. I wrote through the full structural picture here: onhealthcare.tech/p/the-labor-pr…

The "touch once" framing has held up for clinical data through TEFCA. The harder lift is coverage data, where SBC and SPD content still gets re-extracted by every broker, TPA, and HR platform every plan year. Once those documents are emitted in machine-readable form aligned with US Core, the same compounding kicks in.

“Touch once, make available to everybody.” That’s the future of health data. Watch the full conversation with @mcuban on interoperability, transparency, and modernizing healthcare 👇
 youtu.be/5-5oejpK8wE
 #HealthData #HealthIT

Vertafore now AI-imports SBCs into BenefitPoint. 45 min saved per plan, they say. The interesting question isn't "can we extract SBCs" anymore. That's commodity. The question is: can you cite the source page and bbox for every extracted value? #BenefitsTech

@_ACHP The path to zero runs through structured benefit data. If member eligibility, formulary tier, step therapy, and medical-necessity criteria are emitted in machine-readable form, routine cases get auto-approved and human review concentrates where it should. The nonprofit model proves the outcome side. The infrastructure side is what makes it scale.

There’s a better model. ACHP’s own @CeciConnolly & Dr. Charles Bloom lay it out in @NEJMcatalyst: prior auth should be moving toward 0, not celebrated as a profit driver. Our nonprofit health plans prove you can deliver better outcomes without making denials a business strategy. achp.pub/49x2b72

United Health Group (UHG) recently reported their Q1 2026 results, exceeding earnings expectations with $111.7 billion in revenue—a 2% year-over-year increase. They attribute their positive performance to “improved medical cost management.” Just so we are all clear - medical cost management is code for prior authorization, downcoding, care denials, etc. They take pride in denying care because it looks good on their earnings report.

@NedOLutz That overturn rate is the signal. Once CMS-0057-F goes live the appeal-to-denial gap becomes machine-comparable across plans, and the PDFs that house the actual criteria stop being a defensible audit trail. The plans that emit structured criteria and dated decision provenance are the ones that will hold up when the discovery questions arrive.

Centene just told the feds: 1 in 8 Medicare Advantage prior auths denied last year. 94%+ of appeals get overturned. Public record now under CMS-0057-F.

@healthapiguy The Lewis trap is the real play here. Same dynamic shows up in coverage litigation. Arbitration and individual claim handling block aggregation, which insulates decision logic from systemic discovery. The defensible operators are the ones whose plan documents, criteria, and decision provenance can hold up when discovery does land. Path two looks right.

Epic filed their Reply in Veeva Systems v. Epic this week, closing the MTD briefing. The Reply is tight and surgical - Epic quotes Veeva's own opposition back at itself three times in the first paragraph and builds the whole case around those concessions. But the bigger story is what happens next. Both realistic outcomes for Veeva require pulling Epic employees more closely into the case, which triggers Epic's arbitration clauses and class-action waivers (see Epic Systems v. Lewis). Veeva's whole briefing strategy has been about avoiding exactly that trap. Full analysis on Daily Ping, including why I still think path two is the most likely outcome despite Epic's push for with-prejudice dismissal.

Medicare's AI prior auth pilot turned 2-week approvals into 4-8 week waits at Washington hospitals. AI on top of unstructured benefit data isn't faster. It's just slower with extra steps. Cite the source or skip the AI. #HealthTech #PriorAuthorization

Indiana banned AI as the sole basis to downcode claims. Utah requires insurers to disclose AI use in authorization to enrollees starting 2027. The 2026 rule for AI in coverage: if it touched the decision, you better be able to show your work. #HealthcareAI

@NickHealthAI Audit rights on paper are easy. Audit rights with structured access to formulary tier logic, prior auth criteria, and rebate guarantee math are where transparency actually shows up. Most plan sponsors discover the gap during litigation, not contracting. Good to see this surfaced earlier in the cycle. Re: x.com/NickHealthAI/s…

rxbasis.com It analyzes PBM contracts for rebate economics, spread pricing risk, specialty control, audit rights, and suggested contract redlines. Thoughtful work in an area that needs much more transparency.

@healthapiguy The interesting boundary is the payer side, not the clinical reasoning. Once a copilot suggests an order, the next question is whether the patient's plan covers it and on what terms. Provenance from clinical recommendation back to the specific plan rule that allowed or denied it is where defensibility breaks today. Re: x.com/healthapiguy/s…

OpenAI launched ChatGPT for Clinicians this week. It's a free AI copilot for verified U.S. physicians, NPs, PAs, and pharmacists: - Clinical search grounded in peer-reviewed literature - Reusable skills for common workflows - CME credits baked in. The interesting thing isn't the product. It's that OpenAI already shipped most of these features in January as ChatGPT for Healthcare. What's actually new here is the go-to-market: free, individual clinicians, no institutional contract required. The rest is the enterprise product with the enterprise layer peeled off. So why the PLG pivot? Institutional sales in healthcare are slow and painful, and since most of the advertised use cases don't require PHI, OpenAI can skip the BAA overhead. The bigger story is competitive: OpenAI is winning consumer but losing enterprise to Anthropic, which now has a 70% win rate on new enterprise deals, just overtook OpenAI's total revenue run rate, and owns 54% of the enterprise coding market via Claude Code. When your enterprise motion isn't clicking, you lean harder on what you're good at. PLG is a consumer-adjacent motion applied to a B2B-flavored vertical. As others have noted, this is also a full frontal assault on OpenEvidence, the Miami-based AI company that's become healthcare AI's breakout story with 40%+ clinician adoption. Functionally, ChatGPT for Clinicians is OpenEvidence with a bigger brand name and deeper pockets. Which raises the real question: can a horizontal platform displace a hyperspecialized default that already has the clinicians' habits? The meta-lesson is about PLG markets themselves. PLG is permissionless by design. You ship it publicly, your competitors kick the tires, and they copy what works. OpenEvidence ran this motion to challenge UpToDate. OpenAI is running it against OpenEvidence. Someone will run it against OpenAI. But the trap cuts both ways. In PLG, features are copyable. Distribution isn't. OpenEvidence's real moat isn't their product, it's the 40%+ of clinicians who already use it daily. Getting them to switch requires more than parity.

@dp_oneill The provenance gap on network changes is the operational tell. Members get an SBC at enrollment, then learn mid-year that a key provider is out. The plan documents and the live directory rarely reconcile. Structured benefit data plus dated network deltas would make these shifts auditable rather than litigable after the fact. Re: x.com/dp_oneill/stat…

Personally, I don't think Medicare Advantage plans should be permitted to shrink provider networks *during a plan year* In other areas, if you sell a financial product to elderly people, and the product doesn't do what the seller explicitly says it will do, we call it fraud, no?

Another HR tech vendor launched a broker partner program this week. Brokers stay the trusted advisor; tech provides the rails. The broker channel is where benefit data infrastructure gets decided in 2026. Build for it, or get priced out of it. #BrokerTech

Aetna: 88% of PA volume on standardized electronic submission. UHC + Cigna: 70%+ by year-end. The standards are winning. Vendors that emit clean, cited, structured plan data downstream get built into the pipes. The rest don't. #HealthTech #PriorAuthorization

Utah, Indiana, Alabama, California, Texas. Five states now require disclosure or provenance for AI in health insurance decisions. If your vendor can't show sources for every extracted data point, that's a compliance gap, not a roadmap item. #HealthTech #AIRegulation

The drug PA proposed rule gets interesting at the intersection with NCPDP and the Da Vinci PAS and CDex IGs. The standards to run most of what CMS-0062-P contemplates already exist. The gap is that formulary, step therapy, and coverage criteria still live in PDFs rather than structured, queryable form, so even a clean API cannot answer what the plan actually requires.

CMS Proposes Major Changes to Drug Prior Authorization Processes: What You Need to Know natlawreview.com/article/cms-pr… #Health #Government #Laws @SheppardMullin

The local-first posture is the right call for HL7 engineer tooling. The piece the industry still underinvests in is provenance on the transformer itself, being able to point back from a FHIR field to the v2 segment and the mapping rule that produced it. That is what audit teams ask for and what most integration stacks cannot answer. Nice to see an MIT-licensed option enter that space.

Shipping integrator today after 24+ years in healthcare software — an open-source AI teammate for HL7/FHIR integration engineers. MIT licensed. Runs fully local by default; no PHI leaves your machine. What it does: - Paste a broken ADT, get a root cause and suggested fix - Translate HL7 v2 ↔ FHIR in either direction - Generate draft transformer code for Mirth, Iguana, Cloverleaf, Rhapsody, Corepoint, or generic JS/Python - Onboard a new trading partner from a batch of samples → inferred spec + draft transformer - Produce a full FHIR migration plan for a legacy v2 interface Eleven skills total. Accessible from a CLI, a stdio MCP server (drops into Claude Code / Claude Desktop / Cursor), a paste-and-go web UI, or a natural-language ask orchestrator. The memory layer is what makes it compound: every vendor quirk, every Z-segment, every trading-partner note you teach it stays local on your machine. The second diagnosis of an Epic ADT quirk is better than the first. github.com/toddegray/inte…

Posted prices only empower shoppers if patients can compare them against their own benefit design. The harder infrastructure problem is aligning the hospital MRFs and the payer TiC files with a member's specific plan, so the number on the screen is the actual out of pocket, not a list price the plan will negotiate away. Patients Deserve Price Tags is the right direction. Machine-readable on both sides is what makes it real.

.@SenBillCassidy highlights how price transparency is essential to empowering patients: “With price transparency, the patient can shop for the most affordable healthcare…” As he also notes, the Patients Deserve Price Tags Act is a bipartisan bill that will deliver true transparency!

Hospital MRF enforcement went live April 1. Median, 10th, 90th percentile allowed amounts. Type 2 NPIs encoded. A named senior official accountable for file accuracy. Price data is getting structured and auditable while benefit plan data is still trapped in PDFs. That gap will not hold much longer. #PriceTransparency

Real time benefit at prescribing is the cleanest example of price transparency actually reaching a clinical decision. The hard part is not the UI. It is that the cost number depends on plan design, formulary tier, accumulators, and PBM logic that are not exposed in a structured form. Getting the benefit layer into FHIR alongside eligibility is the missing primitive.

What if a clinician could see a patient’s prescription costs in real time while prescribing? That kind of transparency helps support better decisions, lower-cost options, and more affordable care. #HealthIT #PatientAccess #Transparency #PriceTransparency @mcuban