Sixty Percent Appeal Win Rate Exposes Prior Auth Failures

Insurance companies are testing AI systems to make prior authorization decisions without human review.

This is presented as efficiency, and it might be. But the question nobody is asking in real time is whether the efficiency is real or whether we are watching the same mechanism that failed twenty years ago simply get faster.

In the early 2000s, UnitedHealth and Aetna deployed automated denial systems to process prior authorization requests. These systems operated on black-box logic, applying coverage rules at machine speed and producing systematic rejection of care that doctors deemed necessary.

When patients appealed these denials, they won more than 60 percent of the time—meaning the systems were rejecting treatments that secondary human reviewers agreed should be approved. The denials persisted because the incentive structure rewarded rejection. A claim denied is a claim that costs nothing. State regulators eventually forced insurers to require human review before automated denials took effect. The rule worked because it aligned incentives with outcomes.

Sophistication is orthogonal to whether a system is designed to reject or to approve.

The new AI pilots do not mention what changed in the underlying incentive structure. The companies deploying these systems have not become less profitable when they deny care. The financial pressure to reduce approved claims has not disappeared. The technology is better, faster, more sophisticated—but sophistication is orthogonal to whether a system is designed to reject or to approve. What matters now is whether the new regulations that constrained the old systems remain in place, or whether "AI" and "automation" and "efficiency" have become permission to rebuild the machine that already failed once.