Utilization management—the process of determining whether a hospital stay and its treatments will be reimbursed—is a prime candidate for AI, but it is far harder than it looks. Dr. Martin Seneviratne, co-CEO of R37, R1's AI lab, argues in a sponsored Healthcare Dive article that success is measured by financial outcomes: whether a claim is paid, reworked, or denied. The challenge is not simply reading charts.
Seneviratne points out that clinical records are created to support patient care, not reimbursement. The documentation needed to justify a level of care often accumulates over time, and a borderline case on admission may be well supported two days later. Rules also shift across payers, plans, contracts, and markets, and denial patterns add another layer. Meanwhile, feedback on utilization decisions is delayed by weeks or months, making it difficult to train AI systems in real time.
R1's answer, Phare UM, continuously reassesses each case as new notes arrive. It asks two separate questions: whether the documentation supports the current level of care, and how likely the case is to face a payer challenge. For cases the system is confident about, it can issue a determination automatically; for the rest, it produces a prioritized worklist for human reviewers. Seneviratne stresses that clinical judgment remains central, and that foundation models alone are insufficient—systems must learn from real payer outcomes and the work of utilization review teams. The broader goal is to identify denial risk while the patient is still receiving care, rather than after the fact.