About 97,000 people will be admitted to U.S. hospitals today. Behind each admission is a parallel process: determining whether the care delivered can be supported for reimbursement.
Building AI for that task is one of the hardest applied problems I’ve worked on. Success isn’t measured by what a model can generate. It’s measured by financial outcomes: whether a claim is paid, requires rework or is denied.
I’ve spent my career on both sides of that challenge, first as a physician and later in AI research at Google DeepMind and Google Health. One lesson has become increasingly clear: utilization management is not simply a documentation problem. It’s a complex intersection of clinical judgment, payer behavior, operational workflow and financial risk.
The challenge isn’t reading charts
At first glance, utilization management seems well suited for AI. The inputs are largely unstructured clinical records and the goal is to determine whether documentation supports a level-of-care decision.
The reality is much more complex.
Clinical records are created to support patient care, not reimbursement. The evidence needed to support medical necessity often emerges over time as a patient’s condition evolves. A case that appears borderline on admission may be well supported 48 hours later. Understanding that progression requires more than reviewing a single note. It requires following a patient journey across the entire hospitalization.
Then there’s the feedback problem. Utilization management is supposed to be concurrent, with decisions made in real time as care happens. But the feedback on those decisions isn’t: outcomes arrive weeks or months later through approvals, denials, appeals and reversals. Training systems against delayed and incomplete feedback creates challenges that don’t exist in most traditional AI use cases.
Why we built something different
Many AI tools focus on summarization, generation or question-answering. Those capabilities are useful, but they don’t solve the fundamental utilization management challenge.
Our goal was not to build an automation script. It was to create a sophisticated intelligence layer that empowers utilization review teams to identify the cases most likely to require attention before they become denials.
Continuous review across the episode of care is the standard UM teams have always aimed for, but doing it consistently, on every active case, as new documentation lands in real time, is difficult to sustain manually at scale. Phare Utilization Management (Phare UM), R1’s AI-driven utilization management solution, is built to do exactly that: as new information enters the record, the assessment updates. A note added today may provide critical evidence that was missing yesterday.
Just as important, Phare UM evaluates two distinct questions simultaneously. First, does the documentation support the current level of care? Second, how likely is the case to encounter a payer challenge?
Those signals are related but not the same. A clinically appropriate case can still face reimbursement challenges. Surfacing both allows utilization management teams to focus their attention where it can have the greatest impact.
The output is not an automated determination for every case. For situations the system can assess with high confidence, it can move directly to a determination. For everything else, the more ambiguous or high-stakes cases, it is a prioritized worklist that highlights supporting evidence, areas of risk and opportunities for intervention. Clinical judgment remains central to every decision.
Why domain expertise matters
One of the biggest misconceptions in healthcare AI is that access to large language models is enough to solve complex operational problems.
In utilization management, technology is only part of the equation.
Revenue cycle systems need to learn from real payer outcomes, understand how utilization review teams work and reflect the realities of evolving reimbursement policy. That knowledge doesn’t exist in foundation models alone.
At R37, R1’s AI lab, physicians, nurses, revenue cycle operators and AI researchers work together to define what constitutes a defensible case and continuously refine how the models perform in production. Every disagreement, denial, appeal and outcome becomes a learning opportunity.
That combination of technology, operational expertise and outcome data is difficult to replicate from the outside.
Moving upstream
Historically, healthcare organizations have managed reimbursement challenges after the fact, identifying issues only after a denial occurs.
A better approach is to identify denial risk while the patient is still receiving care.
That’s the philosophy behind R1’s revenue operating system, Phare and the broader effort to bring utilization management, authorization and reimbursement workflows into a connected operating model. The objective is simple: address documentation and reimbursement risks earlier, when intervention is still possible.
The opportunity is to scale what utilization review teams already do best. With comprehensive visibility into which cases require attention, they can work more strategically, ensuring the care patients need is also the care they and their hospital get reimbursed for.
Healthcare organizations face growing pressure to improve efficiency while managing increasingly complex reimbursement environments. In utilization management, the most effective AI systems will deliver value that is measured in outcomes: prevented denials, reduced patient burden and clinical expertise focused where it matters most.
That is ultimately the outcome that counts.
Discover how leading health systems are modernizing utilization management.