HIT Consultant – Read More

Healthcare’s utilization management crisis didn’t start with artificial intelligence — but AI threatens to make it exponentially worse if we don’t change course now.

A recent peer-reviewed study from Stanford University, published in the January 2026 issue of Health Affairs, evaluated 21 predictive and/or generative AI tools currently deployed across the utilization review landscape. The findings should serve as a wake-up call: we’re in the midst of an “AI arms race,” where concerns are that payers deploy tools to accelerate denials, and providers counter with AI-drafted appeals. The entire system risks what the researchers call “supercharged flaws” — particularly during prior authorization — that could entrench rather than resolve longstanding tensions.

But buried within the study’s cautionary findings is a roadmap for a better approach. It starts with rejecting the premise that AI should be a weapon in an adversarial battle.

The Arms Race No One Wins

The current trajectory is unsustainable. If payers blame providers for clinical documentation tools that enable overcoding, and providers blame payers for automated systems that generate unfair denials, what is the end result? More AI tools designed to outmaneuver the other side, with patients and care quality caught in the crossfire.

The Stanford researchers identified critical risks that should concern every payer executive:

• Automation bias: Human reviewers developing excessive trust in AI recommendations, even when clinical judgment suggests otherwise
• Anchoring effects: Case managers who review AI-generated summaries first may be unconsciously influenced toward predetermined conclusions
• Expertise gaps: Users without deep clinical knowledge may fail to catch AI hallucinations or inappropriate recommendations
• Opacity: No comparative studies exist on denial rates or wrongful denials in AI-assisted versus traditional reviews

Perhaps most troubling: the researchers found that even at Stanford Health Care itself, staff using AI tools were often unaware of potential biases or limitations.

What the Evidence Actually Tells Us

Here’s what we know works — and what the Stanford study validates:

Meaningful human oversight isn’t optional. This isn’t about keeping “humans in the loop” as a checkbox exercise. It means structuring workflows so that AI guidance is combined with appropriate clinical oversight, and when needed, physician guidance on level of care decisions. The era of rubber-stamping AI recommendations — from either side — must end.

One-size-fits-all AI fails in healthcare. The study’s concerns about how social determinants are accounted for in predictive models point to a larger truth: generalized medicine is behind us. Precision medicine has demonstrated that outcomes improve when care reflects individual needs. The same principle must guide AI adoption in utilization management — accounting for geographic variation, plan mix, and the complexities of specific populations.

Transparency builds trust. The researchers noted that insurers haven’t shared data on how much time human reviewers actually spend on cases that are ultimately denied. This opacity fuels provider skepticism. When AI tools operate as black boxes — for either payers or providers — they erode the trust necessary for effective care coordination.

The Collaborative Alternative

The Stanford study identified only two “collaborative platforms” among the 21 tools evaluated. This distinction matters more than it might initially appear.

Collaborative AI represents a fundamentally different philosophy: rather than optimizing for one party’s interests, these platforms are designed to align payers and providers around appropriate care decisions based on clinical evidence and historical outcomes.

What does this look like in practice?

Precision-tuned thresholds that reflect the unique characteristics of each provider, plan, and patient population — not generic algorithms applied universally. This addresses the study’s concern about social determinants and disparate impacts.

Decision frameworks that combine AI guidance with structured clinical review protocols. This is the antidote to automation bias: AI informs but doesn’t dictate, and clinical expertise remains central to complex decisions.

Shared analytics that help both parties understand patterns in authorization decisions, appeal outcomes, and areas of misalignment. When providers and payers can see the same data — including where and why decisions are being reversed on appeal — it creates opportunities for process improvement rather than escalating conflict.

Transparent governance around how AI models are trained, validated, and monitored for performance. The Stanford researchers called for stronger governance and monitoring for underperformance. Collaborative platforms make this possible by giving both parties visibility into how decisions are being made.

Rethinking the Entire Continuum

If collaborative AI were to become the industry standard, many of the downstream problems the Stanford study examined — appeals automation, post-payment audits, the proliferation of tools designed to game the system — would be far less necessary.

Consider: What if prior authorization and concurrent review processes were built on shared clinical evidence and transparent decision-making from the start? Then, denial rates would reflect genuine clinical disagreements rather than information asymmetries or process failures. Appeal volumes would decrease. Provider administrative burden would decline. And payers would have greater confidence that approved care is clinically appropriate care.

The study examined AI applications across the full utilization management spectrum: prior authorization, concurrent review, claims adjudication, post-payment audit, eligibility verification, and appeals. Each represents a point where an adversarial approach creates friction, cost, and delays. And, each could be transformed by collaborative intelligence.

The Path Forward

The Stanford researchers concluded with recommendations that should guide every payer’s AI strategy:

• Increased transparency in how AI tools make recommendations
• Meaningful human review, not perfunctory oversight
• Staff training on AI limitations and potential biases
• Monitoring for underperformance and disparate impacts
• Governance structures that ensure responsible use

These aren’t aspirational goals — they’re operational requirements for AI that actually serves patients and the healthcare system.

The study also noted that only three of the 21 tools evaluated offer both predictive and generative AI capabilities. This combination matters: predictive AI can identify cases requiring closer review, while generative AI can improve communication and documentation. But both must be deployed within frameworks that prioritize collaboration (and increased efficiency to reduce unnecessary and shared administrative burn) over competition.

A Choice Point

Healthcare stands at a crossroads. We can continue down the path of escalating AI-driven conflict, where each innovation on one side prompts a countermeasure on the other, and the system becomes more complex, more expensive, and more frustrating for everyone involved.

Or we can choose a different approach — one where AI amplifies human judgment rather than replacing it, where transparency builds trust rather than eroding it, and where technology serves to align rather than divide.

In this sense, the Stanford study provides both a warning and a blueprint. The question is whether we’ll heed it.

For payers, the imperative is clear: vet AI vendors not just on speed or cost savings, but on their commitment to collaboration, transparency, and clinical appropriateness. Demand evidence of outcomes, not just efficiency metrics. And recognize that the cheapest or fastest tool may ultimately be the most expensive if it damages provider relationships and patient trust.

As the researchers noted, we need AI that helps insurers approve requests more efficiently, improves communications with providers and patients, and conserves reviewers’ time for genuinely complex decisions. That’s not the AI arms race — it’s collaborative intelligence.

And it’s the only approach that ensures this technological revolution doesn’t have destructive outcomes.

About Matt Brink

Matt Brink leads strategic health plan partnerships at Xsolis, where he has spent nearly a decade helping build what is now the industry’s largest collaborative network connecting providers and payers during concurrent authorization processes. Matt has been instrumental in architecting and scaling the company’s connected network model — now linking more than 600 hospitals with their payer partners to enable real-time, AI-driven alignment.