Healthcare AI Weekly Deep Dive

OpenEvidence just confirmed a $250 million raise at a $15 billion valuation. Let that number sit for a second. This is a clinical decision support tool. Not an EHR. Not a device. A software layer that sits between the clinician and the evidence base, and the market just said it's worth fifteen billion dollars. That's not a funding round. That's a category declaration. The implicit bet is that AI-native CDS will eat the legacy clinical decision support market entirely, and that health systems will pay for access to something that gets smarter as it gets used. The Anthropic and Penn Medicine partnership announced alongside this signals something important: OpenEvidence is not trying to be a standalone tool. It's building a clinical intelligence layer that partners with the foundational model providers and the academic medical centers that generate clinical credibility. That's a smart architecture. It also means the distribution strategy is relationships first, direct sales second. Health systems that have tight academic partnerships are already in the conversation. Community systems are not. For consultants, the immediate question is whether a client's current CDS vendor, UpToDate, VisualDx, whatever they're running, can compete with a platform that has $250 million to build clinical reasoning into every workflow touchpoint. The honest answer is probably not in a five-year window. The strategic question is whether to advise clients to pilot OpenEvidence now, negotiate data use terms carefully, and avoid exclusivity commitments until the competitive landscape stabilizes. The second-order effect: clinical AI makes evidence retrieval cheap, which means fake evidence retrieval also gets cheap. A $15 billion valuation creates enormous pressure to show ROI through utilization metrics. Watch for usage dashboards that look impressive but don't trace to diagnostic accuracy or outcome improvement.

Epic has given its customers Art for clinicians, Emmie for patients, and Penny for revenue cycle. And the most telling signal from this week's reporting is not which systems are turning these tools on. It's which systems, like St. Luke's in Boise, are making deliberate decisions about what stays off. That's maturity. The vendor playbook is always the same: ship features fast, let early adopters absorb the validation risk, then use adoption data as proof of market fit. Health systems that take every feature as it ships are doing the vendor's QA work for free. The systems holding off are asking the right questions. When Art suggests a medication, who is responsible if it's wrong? When Emmie handles patient messaging, what happens when the message is medically significant and the AI misclassifies it? When Penny flags a claim for appeal, what's the false positive rate and who eats the write-off when the appeal fails? These are not hypothetical questions. They are the actual governance questions that determine whether an AI tool creates clinical value or creates liability dressed up as productivity. The deeper issue is that Epic has an incentive to show adoption breadth because adoption data is a competitive moat. 'X thousand health systems are using this tool' is a sales argument. It is not a clinical evidence argument. For consultants advising Epic customers, the most valuable thing you can deliver right now is an evaluation framework that separates clinical evidence tools from workflow automation tools from patient engagement tools, and applies different governance thresholds to each. The failure mode to avoid: approving a tool based on vendor-provided metrics, watching utilization climb, and then discovering eighteen months later that the clinicians are clicking through AI suggestions without reading them. That's not adoption. That's alert fatigue with a new interface.

Ortet launched this week with $500 million from Thoreau, a technology holding platform, and a founding team that includes AI leaders from Genentech, Meta, and AWS. The pitch is a full-stack platform for health: compute infrastructure, frontier research, and model development. Kyunghyun Cho, a respected ML researcher from NYU, is involved in the scientific direction. This is not a vendor selling a point solution. This is a bet that healthcare needs its own frontier AI lab the way defense needed DARPA. The architecture is interesting. Most healthcare AI companies either build on top of GPT or Claude, or they fine-tune open models on clinical data. Ortet is betting that healthcare needs purpose-built foundation models trained on proprietary clinical data from the ground up, with its own compute stack underneath. That's a very expensive thesis, and it's also probably correct. The problem with building clinical AI on top of general-purpose foundation models is that the safety properties you need for clinical use cases are not the same as the safety properties optimized for consumer use cases. A model that's great at writing emails and passing bar exams is not necessarily great at knowing when it doesn't know something clinically important. The Ortet thesis is that you need to build the safety properties into the foundation, not bolt them on afterward. Whether they can execute is a separate question. The $500 million commitment from Thoreau is notable because Thoreau is not a traditional healthcare VC. It's a technology holding platform, which suggests a longer capital horizon and less pressure for a quick exit than a typical venture-backed startup. For health systems, Ortet is a 'watch this' for now. In two years, if they have health system partnerships and deployed models, it becomes an 'act now' on whether to participate in their data consortium.

AKASA's move into autonomous inpatient medical coding is the most significant revenue cycle AI development this week, and it deserves more attention than it's getting. The distinction between inpatient and outpatient coding matters enormously. Outpatient coding is complex but relatively structured. Inpatient coding, particularly DRG assignment and clinical documentation integrity review, is where the money is, where the compliance risk lives, and where human judgment has historically been irreplaceable. AKASA is claiming their autonomous platform can handle that. Here's what 'autonomous' means in practice: the AI assigns codes, reviews clinical documentation, flags discrepancies, and produces a final coding output without a human in the loop for every case. That's different from AI-assisted coding where a human reviews and approves. Autonomous means the system is making the call. The implications are significant. On the upside: inpatient coding is expensive, shortage of qualified coders is real and worsening, and the consistency of AI coding could actually reduce audit risk compared to a team of humans making different judgment calls. On the downside: when an autonomous system codes incorrectly and the error pattern is systematic rather than random, you don't discover it until a payer audit or OIG review. And by then you've overcoded or undercoded at scale. The second-order effect is the one to watch: autonomous coding makes coding cheap, which means the bottleneck shifts from coding capacity to documentation quality. If the AI is coding what's in the chart, and what's in the chart is incomplete or inaccurate, the autonomous system will faithfully code the wrong thing. This is the documentation integrity problem that has always existed, now amplified. For consultants, this is a conversation about whether your client's CDI infrastructure can feed an autonomous coding system. If the docs aren't there, the AI can't help.

Cigna's $3 billion productivity initiative is being covered as a technology story. It's actually a margin story, and health systems should read it that way. When a major payer announces AI-driven productivity, the productivity gains accrue to the payer. The costs shift somewhere. They shift to providers through faster and potentially more frequent denials. They shift to patients through tighter prior auth logic. They shift to the healthcare system through reduced manual review that historically caught errors on both sides. The initiative has three stated components: automating and modernizing processes, better managing suppliers and vendors, and making employees more efficient. Translation: fewer humans reviewing claims means fewer humans who might exercise judgment in a provider's favor. AI-driven supplier management means tighter contract enforcement. Employee efficiency means more cases per staff member, which means less time per case. This is not a criticism of Cigna specifically. It's a description of what 'AI productivity' means in a payer context. Every major payer is doing this. Elevance has been doing it for two years. UnitedHealth has Optum. Aetna has its own automation stack. The difference is that Cigna is now doing it at announced scale with a specific dollar target. For health systems, the strategic response is to build prior auth and denial management capabilities that can match the pace of an AI-augmented payer. Manual denial management processes built for a human-paced payer will not work against a system processing claims at AI speed. This is an argument for investing in tools like AKASA on the provider side before the payer side gets so fast that human-speed appeals become economically unviable.

McKinsey dropped a number this week that is going to travel: AI can now perform up to 22% of US outpatient care. This is exactly the kind of headline that generates board questions, vendor urgency, and bad decisions if you don't decompose it immediately. Here's what to ask when a client sends you this article. First, what counts as 'outpatient care' in this analysis? If it includes scheduling, pre-visit documentation, after-visit summaries, and administrative follow-up, that's very different from clinical assessment and treatment decision-making. AI has been handling scheduling for years. That doesn't mean it's delivering care. Second, 'can perform' is not 'is performing.' Technical capability and deployed clinical utility are separated by FDA clearance, liability frameworks, EHR integration, clinician trust, and patient acceptance. McKinsey's analysis of AI capability at the model level does not account for any of those deployment factors. Third, 22% of outpatient care volume is not uniformly distributed. The tasks AI handles well are the high-volume, low-complexity, well-documented tasks. The 78% it can't handle includes the diagnostically ambiguous, the socially complex, and the cases where the patient's words don't match the chart. The VBC implication is actually interesting and underreported. If AI can reliably handle 22% of routine outpatient tasks, that frees clinical capacity for the high-acuity, high-complexity patients who drive disproportionate cost in value-based contracts. The strategic play is not to use AI to reduce visits. It's to use AI to free the clinicians who can actually move the needle on the top 5% of cost drivers in a VBC panel. That's a different framing, and it leads to different technology investments.

The PHTI survey of 321 purchasers, spanning 106 health plans, 115 employers, and 100 health systems, found that a majority are now using performance-based contracting for digital health tools. This is a structural market signal, not an anecdote. The digital health industry spent the last decade selling on promise and potential. The payer and employer communities have now collectively decided that the promise era is over. You get paid when you deliver. For health systems on the purchasing side of this equation, the PHTI data is a template for their own procurement. If employers and payers are requiring outcomes-based contracts from digital health vendors, health systems should be demanding the same. The question is what outcomes. Engagement metrics, app downloads, and session completions are not outcomes. Blood pressure control, A1c reduction, readmission reduction, ED utilization change: these are outcomes. The critical gap in most digital health contracts is attribution. A vendor will claim credit for any improvement that occurs while a patient is using their tool. The contract needs to specify what comparison population or control condition defines success. Without that, you're not measuring outcomes. You're measuring coincidence. For consultants, this survey is a business development tool. Any health system that hasn't updated its digital health procurement framework in the last 18 months is operating with pre-PHTI standards. The value-add conversation is helping clients build a contracting framework that specifies outcome metrics, attribution methodology, minimum performance thresholds, and clawback provisions. That's a defensible, billable project with a clear deliverable.

Capstone Partners' September 2026 Healthcare IT Market Update is the most useful single data point for understanding where the market is going. A 93.8% surge in AI-enabled healthcare IT M&A year-to-date, with 179 announced or completed transactions, tells you the buyers have decided that AI capability is worth paying a premium for. The strategic acquirers are not buying revenue. They're buying training data, clinical validation relationships, workflow integration depth, and AI talent. Those are the moats that matter in healthcare AI, and you can't build them fast enough organically. For health systems, the M&A wave creates a real problem: vendor selection is no longer a point-in-time decision. A company you select today may be acquired by a competitor, a private equity firm, or a large health IT vendor within 18 months. The product roadmap shifts. The support model changes. The pricing model changes. The integration you built gets deprecated. Health systems need to bake acquisition risk into their vendor evaluation process in a way most of them don't today. That means asking vendors directly about acquisition discussions. It means requiring contractual protections that survive acquisition, including price lock, feature continuity commitments, and data portability rights. It means preferring interoperable, standards-based integrations over deep proprietary ones that are harder to migrate. The second-order effect of the M&A surge is category consolidation. When the market settles, there will probably be three to five dominant platforms in each major healthcare AI category: ambient documentation, revenue cycle automation, clinical decision support, care management. The companies that don't get acquired will either be the winners or the ones that couldn't find a buyer. Knowing the difference requires looking at actual deployment scale, not just funding announcements.

Price transparency rules have been in effect for several years now, and the industry consensus is that the machine-readable files payers produce are a mess: incomplete, inconsistently formatted, often months out of date. Payerset is betting that AI can make sense of that mess at a scale and speed that human analysts can't. The platform pulls in negotiated rate data, claims data, policy documents, and market benchmarks and presents them through an AI research assistant that a health system's managed care team can query directly. If it works, this is genuinely significant. The information asymmetry in payer-provider contract negotiations has always favored payers, who have full visibility into what they're paying across their entire network while health systems typically only know their own rates. Price transparency was supposed to level that playing field. It mostly created a compliance exercise. An AI assistant that can actually synthesize that data into actionable benchmarks changes the negotiating dynamic. A hospital CFO who can say 'based on your machine-readable files, you're paying the regional median for this service line at competitor X while offering us 12% below median' is negotiating from a position of information parity for the first time. The risk is the data quality problem I mentioned. The AI is only as good as the underlying price transparency files, and those files have well-documented quality problems. Payerset will need to be transparent about data coverage gaps. The tool that tells you what it doesn't know is more valuable than the tool that confidently answers every question.

Huron's public growth target is the kind of signal that should land on every competitor's radar. Double-digit growth through 2029 anchored on AI and managed services is not a modest ambition. It's a statement that Huron intends to be a primary AI implementation and operations partner for health systems, not just a project-based advisor. The managed services component is the strategically important piece. Project-based consulting is high-margin but episodic. Managed services creates recurring revenue, deeper client dependency, and data access that compounds over time. A firm running a health system's AI operations has access to performance data, clinical workflow data, and outcome data that informs their next engagement before it's even sold. That's a significant competitive advantage, and it's why every major firm is trying to get into managed services. The healthcare AI managed services market is not yet winner-take-all. There's room for multiple players. But the window to establish a position is closing as health systems sign longer-term agreements. For independent consultants and smaller advisory firms, the competitive response is not to try to match Huron's scale. It's to be the advisor who helps clients decide whether they need managed services at all, and if so, how to structure agreements that preserve strategic flexibility and don't create vendor lock-in. That's a different value proposition, and it's a defensible one. Milliman's formation of an AI solutions practice this week is a parallel signal: actuarial and analytics firms are building AI practices that will compete for the analytics and modeling work that has historically gone to the big strategy firms.