Are the Blues AI-Ready? Blue Cross vs. the Optum Platform Race

In March 2026, the Blue Cross Blue Shield Association published research blaming hospitals' AI billing tools for $2.3 billion in added healthcare costs. It was a grievance — not a strategy. And it stands in sharp contrast to 1981, when the same Association faced a national-platform problem and built something: BlueCard, the shared claims-routing layer that turned 36 independent regional plans into the reason one in three Americans carry a Blue card today.

This episode asks the contrarian question: in an AI world, is the Blues' patchwork of 36 plans a fatal weakness — or the exact architecture the future rewards?

What we get into:

Why Elevance and Highmark are racing in opposite directions (multi-vendor horizontal vs. Epic single-stack vertical) — and what the other 34 plans aren't doing

The plan-level wins that already shipped (BCBS Minnesota, Illinois, Arkansas) — and why none of them are federated

The federated-learning research that says the Blues' structure is the ideal AI architecture — including a peer-reviewed BCBS Louisiana study where regional models beat national algorithms

The Optum problem: what happens if UnitedHealth builds the AI equivalent of BlueCard before the Blues do

Three concrete signals to watch by the end of 2027

Full companion essay on Substack with sources and the three-signal checklist: https://open.substack.com/pub/danmccoymd/p/blue-cross-built-the-last-healthcare

Watch the video version: https://youtu.be/LCrRFqTTtuo

Connect with me at RocketTools.io for AI Strategy Consulting and podcast or speaking engagements. 

Transcript

SPEAKER_00 0:00

In March of 2026, the Blue Cross Blue Shield Association published a research report blaming hospitals' AI building tools for $2.3 billion in additional healthcare cost. Stat picked it up, healthcare leaders picked it up, Benefits Pro picked it up, hospitals were the villains, AI was the weapon, the blues in the story were the ones holding up a hand and saying, please slow down. Now hold that posture in mind for a second because in 1981, a different version of the Blue Cross Association looked at a different national platform problem, claims rooting across 36 independent regional plans. And this time they didn't write a research report about it. They built Blue Card. They imposed a single national clearing layer on a federation of independent plans. They made it work. And Blue Card is the reason one in three Americans today get their coverage from some flavor of Blue. The conventional story right now is that the Blues are too fragmented for AI. 36 independent plans with their own boards, their own legacy infrastructure, their own state regulators. United Health has Optimum, Aetna has CVS. The Blues have 36 different IT departments running 36 different versions of Teradata. The story says they're going to lose this one. I think the conventional story is half right and half wrong. And the half that's wrong is the more interesting half. So let's actually look at what the blues are doing and what they could be doing if the association rediscovered the appetite that built BlueCard. A quick reset on what BlueCard actually is, because the analogy only works if the history is right. In the late 1970s, you had 36 plus independent Blue Cross plants, each operating in its own geography, each with its own claims processing, each with its own provider contracts. If you were a member of Blue Cross of Wisconsin and you got sick on a business trip in Texas, the paperwork was a nightmare. National employers, the kind of customers the Blues most needed to win against Aetna, couldn't get a single bill or a single member experience. Blue Card fixed all that. The association built a shared electronic clearing layer that let any plan adjudicate claims for any other plan's members at the host plan's contracted rates. By the late 1980s, it was operational. By the 1990s, it was the table stakes infrastructure that let the blues compete for national accounts. It was functionally an early HIPAA standard imposed top-down before HIPAA even existed. The thing to notice is that Blue Card was imposed. The association didn't ask 36 independent boards if they felt like adopting a common claim standard. The association built one and required it. That's the relevant fact for the AI question. Because the natural analogy in 2026 is this: whoever builds the shared inference, the data and safety layer and health insurance is going to control the experience of coverage the same way controlling the claims rail did. The platform is the markup, and nobody is going to ask 36 boards if they want to participate. Let's look at what the largest Blues affiliates are actually building. Two plans are visibly making bets at scale, and the bets are completely opposite. Elevance Health is the largest Blue Cross licensee, formerly Anthem, operating across 14 states and serving members in 23 through their Sydney Health platform. Their AI strategy is horizontal and multi-vendor. They publicly disclose working with OpenAI, Anthropic, Google, and AWS, plus a portfolio of healthcare AI startups, with their AI investments deepening through 2025. Sydney Health runs a virtual assistant that handles provider search, cost transparency, deductible tracking, and Spanish language support across the entire footprint. On the back end, their internal AI tooling now generates postcall wrap-ups across roughly a million customer service interactions per month. And this is the part most observers miss. They publicly committed to implementing the NIST AI risk management framework, which is the federal government's voluntary standard for AI governance. It's a way of telling regulators we're going to govern ourselves. Please don't write the rules. HiMark Health is the second largest, with members in Pennsylvania, Delaware, West Virginia, and New York. Their AI strategy is vertical and single stack. In February 2024, HiMark announced they were hosting Epic's Payer Platform on Google Cloud, integrated claims data with clinical data through Epic's electronic health record system. Their year one projection, closing 2.5 care gaps per member and $2.7 million in annual savings at their owned provider system, Allahaney Health Network. The bet is that the future of payer AI lives inside the Epic ecosystem, where the data already is. These are not two flavors of the same strategy. Elevance is betting that AI vendors stay differentiated and you pick best to breed across a horizontal stack. HiMark is betting that the Epic ecosystem becomes the single source of truth and AI lives inside that walled garden. Whichever one is right, the other one is going to look very wrong in five years. And the more important observation, the other 34 blues plans are mostly quiet, not silent, but you don't see press releases announcing major cloud native infrastructure modernizations or generative AI consumer products at this scale. You see plan level pilots, which brings us to the next thing. Three plan level pilots have actually shipped and they're worth pulling up. Blue Cross Blue Shield of Minnesota rolled out a tool called Blue Care Advisor, an AI-driven member engagement platform that surfaces preventative care opportunities to members who are due for screenings or follow-ups. The reported outcome members using Blue Care Advisors are twice as likely to complete preventative care. That's a real number from a real plan, not a vendor white paper. Blue Cross Blue Shield of Illinois has built predictive readmission models. This is clinical scoring that flags members likely to bounce back to the hospital after discharge, so case managers can intervene before they do. Arkansas Blue Cross Blue Shield built a self-service analytics portal that gives employer groups direct access to their own claims data. I like that, with AI-driven natural language querying. It's a kind of product that in a different industry would be a venture-backed startup. Arkansas built it as a feature. These are all good products. They are not federated. They are not interoperable across plans. A Minnesota member who moves to Texas does not bring their blue care advisor history with them. The model that knows them in Minneapolis does not know them in Dallas, and this is the structural problem. Blue Card solved the claims clearing as problem. It never created a unified member record. The most powerful AI use case in health insurance, longitudinal health modeling that follows a member across plan changes, state moves and care settings, is the use case the blue's federated structure makes hardest. Now here's where it gets interesting because the academic literature has something to say about all this. There is a body of research on something called federated learning that maps almost exactly onto the blues federated organizational structure. And the association does not appear to really have noticed. Federated learning is a machine learning technique that came out of Google's research lab in the late 2010s. The idea, instead of pulling everyone's data into a central server and training a model on it, you distribute the model itself out to each data source, you train locally, and then you aggregate just the model updates back into a shared global model. The data never moves, the intelligence actually does. In healthcare, federated learning is the consensus answer to the HIPAA and data sovereignty problem that has hobbled AI development across the entire industry. A 2020 study in scientific reports, cited over a thousand times, showed that a 10-institution federated medical imaging collaboration produced models that reached 99% of the quality you'd get from a centralized training run while keeping every patient record at the institution that owned it. The literature has converged. Federated learning is the plausible path to multi-institutional medical AI under regulatory constraint. And here's the thing that should make somebody at the Blues Association uncomfortable. There is already published peer-reviewed evidence from a blues plan that's showing that regional AI beats national AI for the use cases that the Blues care the most about. Blue Cross Blue Shield of Louisiana published a study in the Journal of Medical Economics in 2020, showing their internally built risk of hospitalization models, trained on Louisiana specifics claims data, outperformed the national risk scoring algorithms used by other plans. Their models hit an AUC of 0.86. Their conclusion in the paper's own words, member-specific regional data can be used accurately to identify patients with high risk of hospitalization, and they can intervene earlier. That's the federated learning thesis written into a peer-reviewed payer paper. Six years before anyone in the C-suite started talking about generative AI. The blues patchwork structure, which is the conventional story treats as the disadvantage, is maybe the architecture that the academic field has converged on as the correct answer for medical AI under regulatory constraint. The problem is nobody is really building it. Which brings me back to the $2.3 billion upcoding study because once you know what the Blue Cross Association could be doing, the gap with what they are doing actually gets sharper. Here is the full menu of the Blue Cross Association's public AI moves in the last 18 months. Well, as best I can reconstruct it from press releases and industry coverage. One, publish research arguing that hospitals' AI is driving up cost. Two, lobby state legislatures for narrower AI oversight on the grounds that overlapping state-level AI rules would disrupt insurance operations. Three, coordinate analytics through a subsidiary called Blue Health Intelligence to detect AI-driven upcoding across the network. That's the list. There is not, as of this recording, a public Blue Cross Association-led generative AI consumer product. There is not a blues-wide federated learning consortium. There is not a shared safety and evaluation framework that Minnesota and Illinois and Arkansas can plug their plan level pilots into. The association's current position paper on AI uses the phrase encouraging responsible AI use. Notably, soft language for an industry facing an existential platform moment. Compared to what the association did in 1981, they didn't encourage responsible claims processing, they built blue cards. The lobbying play is interesting, though, and worth taking seriously because it is at least a strategy. The association's argument to state legislators is that 50 different state-level AI rulebooks would be operationally impossible for a federated insurer. The implicit pitch is federal preemption, light touch, let the industry self-govern through the NIST framework. That gives plans like elevants, who have already publicly committed to NIST a regulatory tailwind to pilot AI in friendlier states first. It's the federated structure used as a regulatory arbitrage machine. But regulatory arbitrage is a defensive play. It buys time, it does not build a platform. And while the association is buying time, Optum is not. Let me close on the Optum problem because it is the structural risk the Blues entire federation is exposed to. Optum is the technology and services arm of United Health Group. It owns the largest pharmacy benefit manager in the country. And as of late 2023, Optum's own leadership publicly reported the group employed or was affiliated with over 90,000 physicians, more than the entire Mayo Clinic system, making it the largest physician organization in the country. And it operates Optum Insight, which is essentially a consulting and analytics organization the size of a centure. United runs Optimum at roughly half the operating margin of its insurance business, but Optimum is the part of the company that's actually growing. Optimum's AI strategy is what you'd expect from a vertically integrated, top-down single corporate parent operation. Build it once, ship it everywhere, no consensus required. They run an internal claims AI engine that handles millions of claims daily across the United Footprint. They have direct-to-consumer telehealth integrated with their pharmacy, and they have a unified longitudinal data set across pharmacy, claims, clinical encounters, and lab data that no blues plan and no consortium of blues plans currently matches. If Optim builds the AI equivalent of BlueCard before the blues do, the local context advantage that Louisiana and Minnesota and Arkansas plans have built up over decades does not save them. The platform layer eats the local layer. That's what platforms do. It's what BlueCard did to whatever local Wisconsin or Texas claims processing system existed before it did. The Blues had decades to build BlueCard, and they had two decades of relative quiet from non-blues competitors before BlueCard became table stakes. The AI window is shorter. Generative AI consumer products went from research demo to mainstream in roughly 24 months. The infrastructure investment cycle in payer IT is at least five years. The Blues are racing a clock that did not used to exist. By the end of 2027, we will know whether the association still has the appetite that built BlueCard. Three specific signals. First, does the association convert Blue Health Intelligence, currently the upcoding detection analytics arm, into a federated learning consortium? That would be the single most leveraged move available. 20 to 30 blues plans, each training their own member data, each contributing model updates back into a shared inference layer. It's the Blue Card Playbook reapplied to AI. The infrastructure exists, the regulatory groundwork is favorable, it's a press release away. Second, does any plan other than Elevance and Hymart publicly commit to a multi-vendor AI strategy with NIST governance and named partners? Right now we have two visible bets out of 36 plans. If we have five or six within 12 months, that's a federation taking shape. If we still have two by the end of 2027, the answer is structural, the association can't move the network, and the small and mid-plans become acquisition targets. Third, does Optimum announce a payer platform as a service offering aimed at small and mid-sized regional payers? Because if I were Optim, that's exactly what I'd build. Strip out the United Brand layer, license the AI inference engine, the prior auth automation, the claims AI, the consumer chat to any payer who wants it, including the blues. Once a smaller blues plan is running on Optum's AI Rails, the Federation is no longer the Federation. It's a customer base. The Blues have the data, they have the relationships, they have the regulatory cover, by the way. They have a literal published academic paper from one of their own plans, proving the local data advantage is real. The structural disadvantage of the Federation is solvable. Federated learning is the entire field's answer to the problem. What they don't have on the public record, at least, is the appetite that built blue card. In 1981, the association looked at the claims routing problem and decided it was their job to solve it. In 2026, the association looks at the AI platform problem and publishes a research report blaming hospitals. The technology has changed, the business model has changed, the competitors have changed. The question is whether the association has changed. If you like this content, please like and subscribe to my Substack. There's more links in the description.