AIは事前承認の問題を解決するのか、それとも悪化させるのか？Will AI fix prior authorization—or make it worse?

Prior authorization, the process by which health insurers require sign-off before they will cover a prescribed treatment, procedure, or medication, has become one of the most contentious friction points in American medicine. Now artificial intelligence is being woven into that process, and the central question is whether the technology will ease a widely criticized bottleneck or amplify its worst failures. The stakes are high because these decisions directly affect whether patients receive timely care and how much administrative burden falls on physicians.

The appeal of automation is straightforward. Prior authorization is slow, paperwork-heavy, and expensive for everyone involved. Physicians and their staff spend hours each week submitting requests, responding to denials, and navigating appeals, while patients can wait days or weeks for approval of care their doctors have already deemed necessary. Insurers argue that machine learning systems can review requests against clinical guidelines far faster than human staff, flag straightforward approvals for near-instant clearance, and reduce the labor costs associated with manual review. In principle, an AI system that rapidly approves routine, well-documented requests could shorten wait times and let clinical reviewers focus on genuinely ambiguous cases.

The concern raised by critics, patient advocates, and some lawmakers is that the same tools can be tuned to deny claims at a scale and speed no human workforce could match. When an algorithm processes thousands of requests, a systematic bias or an overly restrictive rule set does not affect one patient but potentially many. Investigations and lawsuits in recent years have brought this risk into public view. UnitedHealthcare faced litigation over an algorithm known as nH Predict, which plaintiffs alleged was used to cut off rehabilitation and post-acute care for elderly patients, with reporting suggesting a high rate of the tool's denials were overturned on appeal. Cigna drew scrutiny over a system that reportedly allowed medical directors to reject large batches of claims in seconds without individually reviewing patient files. The companies have disputed characterizations of these systems, but the cases crystallized fears about automated denial at volume.

A recurring theme in the criticism is the gap between how these systems are described and how they function in practice. Insurers frequently state that AI is used only to approve or to assist human reviewers, and that a qualified clinician makes any final decision to deny. In reality, the degree of meaningful human oversight appears to vary, and it can be difficult for regulators or patients to verify how much scrutiny a denied claim actually received. Transparency is limited, since the underlying models and the clinical criteria they encode are often treated as proprietary.

Regulators have begun to respond. The Centers for Medicare and Medicaid Services issued rules aimed at speeding up prior authorization decisions and requiring more disclosure, and several states have passed or proposed laws restricting how automated tools can be used to deny medically necessary care. Some measures specifically require that a licensed clinician, rather than software alone, be responsible for adverse decisions. Industry groups, including some insurers themselves, have also announced voluntary commitments to reduce the volume of services subject to prior authorization, a sign that political and public pressure is shaping the landscape independent of the technology itself.

The broader context is that prior authorization automation sits within a wider push to apply AI across healthcare administration, including claims processing, coding, and clinical documentation. Many of these applications are less controversial because they streamline back-office work without directly gating patient care. The difference with prior authorization is that the output is a decision about coverage, where errors or aggressive cost-control tuning translate into denied or delayed treatment.

Whether AI ultimately improves or worsens the system is likely to depend less on the algorithms themselves than on how they are governed. The same predictive capability that could clear routine approvals in seconds could, if optimized primarily for cost reduction, produce denials faster than any appeals process can correct them. Outcomes will hinge on the quality of oversight, the transparency of the criteria, the strength of appeal rights, and whether regulators can audit these systems effectively. For now, the technology appears to be a force multiplier that can push the process in either direction, and the answer to the question in the headline remains unsettled.