Two stories from this month put the state of AI in healthcare in stark relief. On April 6, UnitedHealth Group disclosed it is committing roughly three billion dollars to enterprise AI, with twenty-two thousand engineers building agents to process claims, flag fraud, select billing codes, and engage members through its Avery chatbot. On April 14, the American Hospital Association published a roundup of six health systems using ambient AI scribes to ease documentation, citing a JAMA study that showed scribes saved roughly thirteen minutes of EHR time and sixteen minutes of documentation time per day across five academic medical centers.
Both are real progress. Neither is a strategy.
If you are a C-suite leader at a health system, an ACO, or a provider organization, the question is not whether to spend on AI. Seventy-five percent of U.S. health systems are already using or planning to use an AI platform, and over half of those who measured ROI report at least a two-to-one return. The question is whether the dollars you put into AI are going to outcomes that move clinical performance, financial sustainability, and patient lives, or whether they are going to documentation cleanup and administrative throughput that your competitors will match within twelve months.
That is the actual race. And most of the industry is running in the wrong direction.
Strip away the noise and there are two dominant AI strategies emerging.
The first is administrative. Payers are leading it. UnitedHealth's investment is the most public, but it is not isolated. The company is targeting prior authorization, claims processing, fraud detection, and member-facing chatbots. The goal is to compress cost-to-serve and accelerate decision-making inside the insurance machine. It is a defensible play for a payer. Their margin is in operational efficiency.
The second is clinical. It is quieter, more fragmented, and harder to scale, but it is where outcomes actually move. UC San Diego Health's COMPOSER tool monitors more than one hundred fifty patient variables in real time and detects sepsis four to six hours before clinicians can. The published result was a seventeen percent relative reduction in in-hospital sepsis mortality and a ten percent relative increase in sepsis bundle compliance. That is a deployment that saves lives, not minutes.
For health system and provider organization leaders, the temptation is to follow the payer playbook. Buy the ambient scribe. Deploy the prior auth automation. Pat the budget committee on the back. But your margin is not in admin cost. Your margin is in clinical outcomes, value-based care performance, and the new code stacks CMS finalized for 2026. AI deployed against documentation will not move any of those.
Look at where AI is generating measurable clinical impact and a pattern emerges. The wins are in three places.
Predictive monitoring at the bedside, where AI catches deterioration before it presents clinically. Sepsis is the proof point, but the same architecture works for heart failure decompensation, hypertensive crisis, and diabetic emergencies.
Continuous monitoring outside the four walls, where AI processes vital sign streams, medication adherence data, and patient-reported symptoms between visits. This is the territory CMS has spent the past three fee schedules signaling toward, including the 2026 Medicare Physician Fee Schedule changes that lowered the RPM threshold to as few as two days and ten minutes and raised CCM reimbursement roughly ten percent.
Clinical workflow augmentation for licensed staff, where AI handles routing, prioritization, documentation, and escalation triage so that nurses, physicians, and case managers spend more time on the work only they can do. This is the layer that breaks the workforce constraint that no amount of recruiting will solve.
The HRSA shortage projection for 2026 is a roughly eight percent gap in registered nurses, about two hundred sixty-three thousand RNs short of demand, with forty percent of the workforce planning to leave within five years. You cannot hire your way through that math. You can only redesign the work, and the redesign requires AI as connective tissue, not as a replacement.
There is a simple question that separates AI investments that compound from AI investments that decay. Does the deployment shorten time-to-intervention, or does it only shorten time-to-bill?
Time-to-bill efficiency is real value. It does not produce competitive advantage. Every health system will eventually deploy ambient scribes, automated coding, and prior auth bots. The vendor market will commoditize those tools inside eighteen months. If that is your AI strategy, your finance team will see a one-time productivity bump and then nothing.
Time-to-intervention is where the durable advantage lives. A patient whose blood pressure drift is caught on day three instead of day thirty does not become a hospitalization. A heart failure patient on the right titration of guideline-directed medical therapy does not become a thirty-day readmission. A diabetic patient whose A1C trends are flagged in real time does not become an amputation. Every one of those non-events is a clinical win, a financial win in any value-based contract, and a quality measure that improves your HEDIS performance.
This is the model Welby Health was built on. Licensed RN case managers paired with AI-powered workflows that monitor vitals, drive medication adherence, and surface clinical alerts to the right person at the right time. The outcomes track. Hypertensive patients on cellular-enabled blood pressure cuffs show a twenty percent decrease in blood pressure. Diabetic patients on smart glucose monitors show over twenty percent reduction in blood glucose within four weeks. Heart failure patients are 5.5 times more likely to adhere to life-saving therapies. Eighty-four percent of clients pass HEDIS measures for systolic blood pressure control after ninety days. Eighty-one percent pass A1C control after the same window.
These numbers are not the point. The point is what produced them. AI deployed at the clinical front line, in the hands of licensed nurses, against specific outcome targets that align to value-based contracts and quality measures. That is a different operating model from buying a documentation tool and hoping the productivity flows downstream.
The 2026 reimbursement environment is not subtle. CMS lowered the RPM time threshold, raised CCM reimbursement, finalized the mandatory Ambulatory Specialty Model for heart failure and low back pain, and is launching the ACCESS model on July 5, 2026 with outcome-aligned payments tied to chronic care performance over a ten-year horizon. The payers are diverging. UnitedHealthcare tried to gut RPM coverage for chronic disease earlier this year and walked it back under industry pressure, but their direction of travel is clear.
What this means in practice is that the AI investments that compound will be the ones that produce measurable clinical outcomes inside the new reimbursement structure. Health systems that deploy AI to compress documentation will see one-time gains. Health systems that deploy AI to compress time-to-intervention, drive adherence, and capture the new chronic care code stack will build a structurally better operating margin and a clinically defensible care model.
The payer AI race is real. It is also not your race. Run the right one.
Healthcare's AI investment is accelerating, but the dollars are concentrated in administrative and documentation use cases that will commoditize within eighteen months. The AI deployments producing real clinical and financial returns are the ones that shorten time-to-intervention rather than time-to-bill, including bedside predictive monitoring, continuous remote monitoring, and clinical workflow augmentation for licensed staff. With CMS lowering the RPM threshold, raising CCM reimbursement, and launching outcome-aligned payment models in 2026, health system leaders need to evaluate AI by whether it moves clinical outcomes inside the new reimbursement structure. Welby's model of pairing RN case managers with AI workflows is one example of this approach, with measurable outcomes including a twenty percent decrease in blood pressure and 5.5 times higher adherence to life-saving heart failure therapies. The wrong AI race is the easy one. The right one is harder, and it is where the durable advantage lives.
