AI can end the AI arms race in health care utilization management

Instead of resolving long-standing tensions, the technology risks amplifying them, creating what the authors call "supercharged flaws," especially in prior authorization.

Embedded within the study's warnings is a map to a better approach.

It begins with rejecting the premise that AI should function as a weapon in an adversarial system.

The Arms Race No One Wins

The analysis reinforces several principles about how AI should be deployed in health care.

First, meaningful human oversight is essential.

This goes beyond the checkbox idea of keeping "humans in the loop."

Workflows must combine algorithmic guidance with clinical expertise, including physician input when determining appropriate levels of care. The era of rubber-stamping AI recommendations has to end.

Second, one-size-fits-all AI rarely works in medicine.

The authors highlighted concerns about how predictive models incorporate social determinants of health, but the issue runs deeper.

Precision medicine has shown that outcomes improve when care reflects individual patient needs.

AI in utilization management should follow the same principle, adapting to geographic differences, plan structures, and the unique characteristics of patient populations.

Finally, transparency is critical for trust.

The study notes that insurers generally don't share data on how much time human reviewers spend evaluating cases that end in denial.

When AI systems function as black boxes, skepticism grows and payer-provider coordination suffers.

The Collaborative Alternative

Among the 21 tools analyzed, the researchers identified only two designed as truly collaborative platforms.

That distinction matters more than it may seem.

"Collaborative AI" reflects a fundamentally different philosophy.

Rather than optimizing outcomes for one party, these systems aim to align payers and providers around evidence-based care decisions informed by clinical data and historical outcomes.

In practice, this approach includes precision-tuned thresholds that account for the characteristics of specific providers, health plans, and patient populations rather than applying generic algorithms universally.

It also incorporates decision frameworks that combine AI insights with structured clinical review protocols, ensuring that technology informs decisions but doesn't dictate them.

Shared analytics represent another critical component.

When both payers and providers can see patterns in authorization decisions, appeal outcomes, and areas of disagreement, the conversation shifts from conflict to improvement.

Transparency around how AI models are trained, validated, and monitored further strengthens accountability.

Rethinking the Entire Continuum

Many downstream problems identified in the Stanford study — automated appeals, post-payment audits, and tools designed to counteract earlier decisions — would be less necessary if collaborative AI became standard practice.

Imagine prior authorization and concurrent review built on shared clinical evidence and transparent decision frameworks from the start.

Denials would reflect genuine clinical disagreements, rather than gaps in information or process failures.

Appeal volumes would fall.

Administrative burdens on providers would ease.

Payers could have greater confidence that approved care is clinically appropriate.

Each point in that continuum across the utilization management spectrum (prior authorization, concurrent review, claims adjudication, post-payment audits, eligibility verification, appeals) can become a source of friction in an adversarial system.

Each presents opportunities for collaborative AI.

The Path Forward

The recommended guardrails around AI adoption in health care aren't abstract ideals: greater transparency in how algorithms generate recommendations, meaningful human review, staff training on AI limitations, active monitoring for bias or underperformance, and governance structures that ensure responsible deployment.

These are operational requirements if AI is to serve patients rather than deepen systemic problems.

The analysis notes that only three of the evaluated tools combine predictive and generative AI capabilities.

That pairing can be powerful.

Predictive models identify cases requiring closer scrutiny, while generative tools improve documentation and communication.

Both must operate within collaborative frameworks that prioritize alignment over competition.

The Stanford analysis offers both a warning and a blueprint.

Whether the industry follows it may determine whether AI becomes health care's next administrative arms race — or the beginning of a more collaborative model of care.

Matt Brink leads strategic health plan partnerships at Xsolis, a company that connects providers' and payers' authorization processes.