Cityblock Health's CEO argues that the dominance of billing AI in U.S. health care is built into fee for service payment, and that only a value based payment model makes patient facing AI the rational investment.
Health systems in the United States are spending roughly 60 cents of every health-care AI dollar on software that bills, codes, and adjusts risk on insurance claims, while the sliver that actually reaches patient care keeps shrinking. The pattern is not a failure of intentions. It is what fee-for-service reimbursement predicts.
That is the case Dr. Toyin Ajayi, co-founder and CEO of Cityblock Health, made on the latest episode of the KFF "Business of Health" podcast, and it is the spine of a new Cityblock report on AI in Medicaid. Cityblock runs a value-based primary-care model serving roughly 100,000 Medicaid and dual-eligible members (people who qualify for both Medicare and Medicaid because of age, disability, or low income) across ten states. The company's argument is that the AI arms race inside American health care is a mirror of the payment system that funds it, and that the only way to change where AI dollars flow is to change how the system gets paid.
The 60-percent figure comes from 2024 reporting by SVB and Fierce Healthcare on venture investment patterns, not from a Cityblock measurement. Ajayi uses it to make a structural point. In a system that pays providers per procedure, per code, and per risk-adjustment category, the AI tools that make those workflows faster and more accurate are the AI tools that earn back their cost. The same logic is what makes ambient documentation, prior-authorization bots, and coding copilots the obvious buys for hospital systems, and what makes "AI that calls a member about a missed asthma refill" a harder business case to underwrite.
Menlo Ventures' 2025 State of AI in Healthcare report tracks the same pattern from the demand side. Healthcare AI spend hit roughly $1.4 billion in 2025, about triple the 2024 figure, and adoption is running 2.2 times the broader economy. By Menlo's tally, 27 percent of health systems had implemented domain-specific AI tools, 18 percent of outpatient groups, and 14 percent of payers, with eight healthcare AI companies reaching unicorn status. The largest named rollouts are back-office. Kaiser Permanente's Abridge ambient documentation now spans 40 hospitals and more than 600 medical offices, billed as the largest generative-AI deployment in U.S. health care. Mayo Clinic has committed more than $1 billion across 200-plus AI projects. Advocate Health is live with 40 AI use cases, including Microsoft Dragon Copilot, with a projected 50-percent reduction in documentation time.
These are not bad projects. The friction is in what they do not touch. Ambient scribes free physicians from keyboards, but they do not change who gets a callback. Risk-adjustment AI extracts more revenue per member, but it does not keep that member out of the emergency room. In a fee-for-service chassis, optimizing the parts of the system that already exist is the rational capital move, and the venture ledger reflects it.
Cityblock is the case for the alternative math. Because the company is paid a fixed monthly amount per member and is on the hook for total cost of care, AI that improves outcomes lowers Cityblock's expenses and raises its margin. The company's report is, in effect, a manifesto for AI built around the patient relationship, anchored on six company-stated principles: equity first, solve pressing problems, trust before data, AI as a relationship engine, behavior change as the goal, and scalable compassion. The principles are company framing, not independent validation, and Cityblock has not published clinical outcome data tied to the AI deployments.
Two operational examples do most of the work. The first is what Cityblock calls "agentic AI" outreach: systems that listen for housing instability, food insecurity, and other social-determinants-of-health signals in member calls and messages, then route a structured summary to a human care coordinator who can act on it. The second is an ambient scribe that natively translates real-time Spanish encounters into English clinical notes, an unlocks-the-front-door feature in a Medicaid population that is often Spanish-primary at home and English-primary in the chart.
The pattern matters most where the structural case is sharpest. Federal eligibility churn, workforce attrition, and shrinking state Medicaid budgets are squeezing the safety net at the same moment AI spend is accelerating across the rest of the system. Ajayi's argument is that AI built for care, not for claims, is one of the few levers that can lower cost in this population by reaching the highest-need members whose care drives most of the spending.
The honest version of the argument is narrower than the slogan would suggest. Fee-for-service has been the engine of U.S. health care for decades, and the AI investment pattern is a faithful shadow of it. The companies that win on ambient documentation, risk adjustment, and prior authorization are solving real problems for the customers who pay them, and they will keep winning as long as those customers get paid the same way. The Cityblock version of AI only becomes the default when more of American health care is paid the way Cityblock is paid. The 60-percent figure will not move on its own.
The watch item through the rest of 2026 is whether Medicaid managed-care contracts, Medicare Advantage Star Ratings, and ACO REACH benchmarks shift the value-based pool fast enough to move venture dollars with it. If they do, the AI line items in next year's Menlo report start to look different. If they do not, the next Cityblock-size deployment will keep waiting for a payment model that makes it obvious.