A new pilot programme in the United States is using AI to determine whether certain Medicare-covered treatments should be approved, and early reports suggest the system is already creating delays and unexpected denials of care. The six-year Wasteful and Inappropriate Service Reduction (WISeR) programme operates in Washington, Arizona, New Jersey, Ohio, Oklahoma and Texas. Doctors seeking approval for 15 selected procedures must submit clinical reasoning through an online portal, where AI assesses whether the treatment meets Medicare requirements.
The programme was introduced to reduce wasteful Medicare spending and prevent unnecessary procedures. HHS estimates that billions of dollars of Medicare spending may go toward services with minimal benefit, so the administration argues that AI could make prior authorization faster and more consistent. But doctors in the participating states report the opposite in some cases: procedures that previously took a day to authorize can take weeks, while some patients have faced repeated denials and additional appointments. In Washington, one report cited average waits of 15 to 20 days, with some patients waiting even longer.
The biggest concern is transparency and accountability. Physicians interviewed by Science Focus describe the AI systems as effectively black boxes, with limited visibility into the data and reasoning behind individual decisions. There are also concerns about incentives because technology companies participating in the programme receive compensation partly connected to savings generated by reducing inappropriate care. CMS says the system includes safeguards: AI denials must receive human clinical review, companies are monitored for accuracy and overturned decisions, and poor-performing participants can face corrective action. Some providers also say their systems are designed to use AI to identify requests that can be approved quickly, rather than allowing AI itself to make a final “no” decision.
The larger lesson is that AI can be useful in healthcare without necessarily being appropriate as the final decision-maker. AI can help organize medical records, identify patterns in scans and check whether documentation meets established requirements, but deciding whether an individual patient should receive treatment involves context that may not be captured in a standardized dataset. The WISeR experience therefore illustrates a central challenge for healthcare AI: the question is not simply whether an algorithm is accurate, but where it should be placed in the decision-making chain, how its reasoning can be challenged, and who remains accountable when it gets a patient's care wrong.
