Healthcare AI Weekly Deep Dive

Commure is General Catalyst-backed, has meaningful clinical AI product penetration, and just terminated a program that paid for referrals to new customers. STAT published an investigation. The company terminated the program. That sequence tells you everything. You don't terminate a program because it was working great and nobody noticed. Referral payment programs in healthcare technology are not inherently illegal. But when the product in question is a clinical AI tool touching billing, documentation, or care decisions, you are operating in territory where the Anti-Kickback Statute applies broadly and the OIG is paying attention. The second-order effect here is the one that matters most for consultants. AI makes clinical product adoption cheap to pitch, which means fake adoption also gets cheap. Referral schemes are one version of that: you're not buying clinical value, you're buying network effects through financial incentive. The Potemkin village just got a loading dock. What your clients need to do immediately: one, identify every AI vendor they have contracted with in the last 24 months. Two, ask each vendor directly whether they operate or have operated any referral payment program tied to customer acquisition. Three, get that answer in writing and route it to legal. Four, check contract indemnification language. If a vendor's sales model later triggers an OIG investigation, you want to know whether your client is holding the bag. The Medicare NTAP story from the same STAT issue compounds this. Medicare pays hospitals when they use newly-authorized AI devices under the New Technology Add-on Payment program. Researchers are already flagging overuse risk. When you combine NTAP incentives with referral-driven adoption schemes, you get a structure that could look very bad under retrospective audit. This is not hypothetical. This is exactly the pattern the OIG writes enforcement letters about.

Epic's UGM is where the product roadmap becomes real. This year the preview signals AI agents and Cosmos expansion as the headline themes. That's a big deal. AI agents inside Epic means Epic is moving from ambient documentation and clinical decision support into autonomous workflow execution. That's a different category of product. The Redesign Health study published this week found that 71% of health systems have adopted an Epic-first AI purchasing strategy. That number is doing a lot of work. It means the competitive dynamic has already shifted. Independent AI vendors are not competing against each other anymore. They're competing against Epic's roadmap, which they can't see until UGM. The nursing documentation pilot at Northeast Georgia Health System, now the fifth health system to launch Epic's nurse AI tool, shows how fast Epic can roll when the product is ready. Five health systems in what appears to be weeks or months. That's not a pilot. That's a go-to-market. The Cosmos angle matters for VBC clients specifically. Cosmos is Epic's de-identified patient data network. If Epic is expanding Cosmos capabilities alongside AI agents, the combination becomes a population health and risk stratification engine that sits inside the EHR workflow. That's not a point solution you can beat with a better dashboard. So here's the challenge for your clients before UGM. One, audit every AI point solution contract against what Epic has announced or is about to announce. Two, identify which vendors are in the App Orchard and which are not. Three, pressure test your AI vendor on their Epic integration roadmap, specifically whether they are certified or working toward certification. Four, build a decision framework now for when Epic native capability is good enough versus when a best-of-breed point solution is worth the integration overhead. Because after UGM, that conversation is happening whether you're ready or not.

Decart is an Nvidia-backed AI infrastructure startup. Anthropic buying it for $6 billion is the largest acquisition Anthropic has ever pursued. The stated rationale is cheaper, faster compute as AI labs hunt for infrastructure efficiency. That makes sense on the model training side. On the healthcare deployment side, it's more complicated. Anthropic has been actively pitching healthcare AI to IPO investors. The PYMNTS story this week explicitly frames Anthropic's healthcare positioning as an answer to public skepticism about AI, which is a very specific kind of investor messaging. Ode and PointClickCare announced a partnership with Anthropic this week for AI-enabled workflows in post-acute and long-term care. That's a real deployment signal. But the Decart acquisition suggests Anthropic's attention is on infrastructure consolidation and IPO positioning, not on building clinical-domain-specific products. Here's the thing. When a foundation model company goes public, its incentives shift from research and deployment toward quarterly revenue targets and gross margin. Healthcare is a notoriously slow-revenue vertical with high compliance overhead. Post-IPO Anthropic may prioritize enterprise financial services, legal, and software development over healthcare. Health systems that have embedded Claude into clinical workflows are not just buying a model. They're betting on Anthropic's post-IPO strategic priorities. That's a vendor concentration risk that belongs in your client's AI governance framework. The second-order read: Anthropic buying compute infrastructure makes Claude cheaper to run, which makes Claude-based clinical AI cheaper to deploy. That sounds good. But cheaper to deploy also means more vendors will wrap Claude in a healthcare product and sell it without clinical validation. The slop-wrapped-in-a-demo problem gets worse, not better, as the underlying model cost drops.

AthenaHealth has deployed athenaAmbient natively across its athenaOne network. No extra charge. 170,000-plus physicians and clinicians. The product captures encounters and generates draft notes, suggested diagnoses, orders, and care-gap alerts. The two-workdays-per-month claim is the headline metric. Let's take it seriously for a second and then poke at it. Two workdays per month is roughly 10% of a full-time clinician's time. If that's real and reproducible, it's one of the most significant efficiency gains in clinical workflow in the last decade. That's the upside case. Now the contrarian read. This metric comes from the vendor's own network data. Ambient scribe saves note-writing time. Right? Everybody agrees on that. What it doesn't necessarily do is improve diagnostic quality, reduce missed diagnoses, or improve patient outcomes. It makes the note faster. The note was never really the problem. The problem was physician cognitive load across too many patients with too little time for genuine clinical reasoning. Faster notes mean you can see more patients, which can mean more revenue per clinician, which health systems love. It can also mean each patient gets marginally less physician attention. That's the tradeoff nobody in the vendor deck talks about. The competitive pricing signal here is real regardless of your position on the outcome question. AthenaHealth bundling ambient AI at no extra charge is the same move athena has made before with other features: commoditize the point solution to protect the platform. Every standalone ambient scribe vendor that sells into the athena customer base is now in a pricing war with zero. Nuance, Ambience, Abridge, Suki: you now have to prove you're worth a separate line item against a free native option. That conversation is happening in Q4 budget season right now.

Google's AMIE system just cleared a milestone that nobody else has hit: real-time clinical video consultation in a research study. The system conducts the encounter, listens, reasons, and responds in real time during a video call. This is not a chat interface. This is an attempt to replicate the structure of an actual clinical visit. The Abbott and Google partnership announced this week is the connective tissue. Abbott's continuous glucose monitoring data feeds into Google Health's AI coaching platform. That's longitudinal physiologic data combined with a clinical reasoning system. The data flywheel for AMIE gets better if Google has access to real-world health data at scale. Here's what this means for the next three to five years rather than the next three to five months. Google is building toward a vertically integrated AI health stack: wearable data collection via partnerships like Abbott, population-level health data via Google Cloud healthcare clients, clinical reasoning via AMIE, and consumer-facing AI health coaching via Pixel and Google One. None of those pieces individually threatens Epic. All of them together, connected and trained on longitudinal data, is a different conversation. The relevant client question is not whether to switch from Epic to Google. The relevant question is whether your clients have a Google Cloud relationship and what data sharing terms they agreed to when they signed it. The AI coaching product controversy this week, where users complained about fabricated health states, is a reminder that consumer-facing health AI from big tech carries reputational risk for health systems that partner with it. That's worth noting in any Google Health partnership discussion.

Researchers analyzed Medicare cost reports and found that four mandatory VBC payment programs were associated with more than $3 billion in aggregated administrative costs compared to non-participating hospitals. This is a real number from real cost reports, not a survey estimate. The programs studied include some of the highest-profile mandatory CMS models of the last decade. Here's the thing about this study. It lands in the same week that WISeR, CMS's new technology-assisted prior authorization pilot in six states, survived a Senate vote to terminate it. WISeR is the next generation of mandatory VBC-adjacent programs. It uses AI to automate prior authorization decisions in traditional Medicare. The Senate voted 46 to 50 to keep it running. That's not a ringing endorsement. That's a program that almost died on a party-line vote. The $3 billion administrative cost finding is exactly the argument program opponents will use against WISeR if the pilot data doesn't show clean efficiency gains. Meanwhile, insurer denial rates remain stubbornly high. KFF reported this week that insurers denied at least one in eight standard prior authorization requests across Medicare Advantage, Medicaid, and ACA marketplace plans last year. AI-generated prior auth appeals make the appeal process cheap. Which means insurers will get buried in volume and potentially auto-deny more. The WISeR model is supposed to cut through this by automating the authorization decision on the CMS side. Whether it does or just adds a new layer of contested decisions is the open empirical question your clients cannot answer yet. What they can do: pressure-test their AI VBC vendors on administrative cost impact, not just clinical quality metrics. The pitch is always about quality improvement. The CFO's question is about whether the total administrative cost went down. Get that number in the contract.

Cedar launched Kora as a platform of purpose-built autonomous AI agents for healthcare revenue recovery. The underlying data assets are substantial: 1.5 billion-plus patient interactions, $13.6 billion in processed payments, 63 million-plus patients served. The 20% outbound call pickup rate versus an industry average around 5% is the headline metric. Let's be precise about what that means. Pickup rate measures whether a patient answers the phone. It does not measure whether they pay. A 4x pickup rate improvement is meaningful if it translates proportionally to collection improvement. If it translates to 4x as many patients saying they'll call back later, it's a vanity metric. The agentic AI model here is the more interesting strategic angle. Cedar is not just automating a single task. They're building coordinated agents that handle different stages of the revenue recovery workflow: outbound calls, eligibility confirmation, payment arrangement. That's a fundamentally different architecture than a single-purpose bot. For health systems, the relevant question is EHR integration depth. Cedar says Kora integrates directly into existing EHRs. The specifics of that integration matter enormously. Does it write back to the patient account? Does it trigger workflow actions in Epic or Oracle? Can it adjust the communication strategy based on the patient's clinical context? Those questions determine whether Kora is a point solution or a platform.

Aetna's internal VBC analysis covers its Medicare Advantage book and shows improvement in preventive care metrics alongside cost savings in VBC-contracted markets versus non-VBC markets. The preventive care metrics likely include AWV completion rates, cancer screening rates, and chronic disease management indicators. Those are the same metrics that AI-powered population health tools claim to move. Context matters here. The MA market is under significant pressure. UnitedHealth Group, Humana, and CVS Health have all reported elevated medical costs and plan exits. Aetna publishing a VBC success story in this environment is also a market positioning move. That doesn't make the data wrong. It means you should understand the methodology. The Franciscan Alliance partnership with an Elevance-backed physician group is the more operationally interesting story. Elevance, which has backed physician group infrastructure through its Carelon subsidiary, is building clinical delivery capacity that sits inside VBC contracts. The AI tools that matter for this structure are risk stratification engines, gap closure automation, and AWV scheduling tools. If your client is negotiating a new MA VBC contract in the next 12 months, the technology stack that supports that contract is a consulting opportunity. The question is whether they're building toward a data infrastructure that makes the contract economically viable, or whether they're signing a contract and hoping the operations catch up.

Kyndryl was spun out of IBM in 2021 as a pure-play IT infrastructure services company. The acquisition of Healthcare IT Leaders, a healthcare-focused staffing and advisory firm, signals that Kyndryl is trying to move up the value chain from infrastructure services into advisory. The framing is AI-led modernization, which is the current packaging for what used to be called EHR optimization and digital transformation. The Guidehouse partnership with Vant4ge announced this week for AI-enabled solutions in public safety and human services shows the broader pattern: infrastructure and advisory firms are all trying to bolt AI practices onto their existing delivery models. The difference is that Kyndryl has large enterprise health system relationships from its IBM legacy. Those relationships create access that boutique advisory firms cannot easily replicate. The risk for clients: Kyndryl's healthcare AI advisory capability is newly acquired. Healthcare IT Leaders has consulting DNA, but the integration into Kyndryl's delivery model takes 18 to 24 months to stabilize. Clients who engage Kyndryl for AI modernization in the first year of this acquisition are essentially beta testing the combined entity.

IBM's CEO admitted publicly that the company did not move quickly enough on AI. Weeks later, IBM announces a strategic partnership with OpenAI and plans to build a thousands-strong AI consulting practice around GPT-5.6 integration. This is a reactive move dressed as a strategic one. The partnership itself is real. IBM brings enterprise security, hybrid cloud infrastructure, and large-scale delivery capacity. OpenAI brings model capability and the GPT-5.6 API. The combination could be powerful for large health systems that need enterprise-grade security governance around AI deployment. The Watson Health ghost still matters here. IBM spent years and significant resources building clinical AI products that did not deliver at scale. The credibility damage in healthcare is real and recent. The consultants IBM is now recruiting into this OpenAI practice are not the Watson Health team, but health system CIOs have institutional memory. Expect IBM to lean heavily on the OpenAI brand rather than the IBM AI brand in early client conversations. The second-order effect: IBM's thousands of consultants being trained on OpenAI tools means GPT-5.6 based solutions will proliferate across IBM's health system client base rapidly. That's a distribution advantage. It also means a lot of early implementations that are works in progress, supported by consultants who are learning as they go.

Deloitte surveyed healthcare CFOs and found a decision readiness gap: CFOs have expanding expectations around AI and business performance but lack the support systems to measure whether AI is delivering value. The survey headline is 'Momentum isn't a metric.' That's a good line. It's also the setup for Deloitte's AI value measurement service offering. The finding itself is credible and matches what we hear from health system finance leaders directly. CFOs are getting board pressure to show AI ROI. They signed contracts with ambient scribe vendors, RCM AI vendors, and clinical decision support tools. Now they need to show the return. Most of them can't, because the metrics in the vendor contract are activity metrics, not outcome metrics. Notes generated per hour. Calls answered per day. Alerts fired per week. None of those translate directly to margin impact or patient outcome improvement without an analytical layer that most health systems don't have built. Here's the opportunity this creates for consultants who are not Deloitte. Deloitte's AI value measurement framework will be enterprise-scale, methodology-heavy, and tied to Deloitte platform tools. Health systems that need a faster, more operationally grounded answer need a different approach. The conversation starts with three questions: what did you pay for each AI tool, what metric did the vendor promise to move, and what did that metric actually do in the 90 days after go-live? If the CFO cannot answer question three, that's your engagement. The Deloitte survey is your door opener. Cite it. Then offer the faster path.