Prior Auth Does Not Need a Black Box. It Needs an Agent That Shows Its Work.

The Pressure Is Real. The Response Matters.

Prior authorization has always been slow, manual, and fragmented. What has changed is the regulatory environment around it.

CMS-0057-F took effect on January 1, 2026. The rule applies to all prior authorization, regardless of how payers process it. Standard PA decisions must now be returned within seven calendar days and urgent decisions within 72 hours. Every denial must include a specific reason. And beginning March 2026, payers must publicly report PA metrics, including approval rates, denial volumes, and average processing times.

These requirements do not prescribe technology. But meeting them consistently, at scale, with the accuracy and transparency CMS now demands, will be difficult to sustain through manual workflows alone. That reality is driving payers toward AI. The 2024 AMA Prior Authorization Physician Survey found that 61 percent of physicians already believe AI is increasing denial rates. Ninety-three percent report that PA delays access to necessary care.

AI in PA is no longer a question of if. It is a question of what kind, and whether that AI earns trust or erodes it further.

The Black-Box Problem

Much of the AI entering PA today follows a familiar pattern: ingest a request, run it through a model, return a decision. The clinical reasoning behind the output is often opaque. Researchers at Stanford and the Health Affairs journal have documented concerns about limited transparency, algorithmic opacity, and insufficient human oversight in AI-driven insurance decisions. A federal class action lawsuit against one major insurer alleged a 90 percent error rate in its AI-based PA determinations.

That is the black-box trap. It produces faster decisions, but not necessarily defensible ones. In a regulatory environment where every denial must carry a specific, documented reason, and where PA metrics are now publicly visible, speed without explainability is a liability.

What a Glass-Box Alternative Could Look Like

An emerging concept gaining traction in the industry is agentic AI: systems designed not just to generate a decision, but to work through the problem step by step with traceable, auditable logic. In a PA context, this could mean reading clinical documentation against the plan’s medical necessity criteria and benefit rules, adapting when payer policy changes to eliminate denials caused by outdated logic, and orchestrating the full workflow from eligibility verification through decisioning and provider notification.

In this model, complex or low-confidence cases would escalate to human clinical reviewers with the evidence, policy match, and rationale already assembled. Every step would be documented, not as an afterthought, but as a core function of how the system operates.

The building blocks are emerging, and players are moving fast. PA platforms, UM vendors, appeals tools, and member-services agents are accumulating across the enterprise, each with its own model, its own audit format, and its own contract. CMS has launched the WISeR pilot program, testing AI-driven PA decisions for specific Medicare services across six states. What most payers do not yet have is a unified governance story across all of it.

The 2026 Deloitte US Health Care Outlook reinforces the momentum: over 80 percent of health care executives expect agentic AI to deliver significant value, and 70 percent of health plans are prioritizing it for utilization management and prior authorization. The trajectory is clear, even if full maturity is still ahead.

Why Payers Should Be Evaluating This Now

For providers, the PA conversation centers on reducing administrative burden. The AMA reports that physicians and their staff spend an average of 13 hours per week on prior authorization requests. The provider’s goal is straightforward: less paperwork, fewer fax chases, less time on hold.

For payers, the challenge is different in kind. Payers must process PA requests faster, with greater accuracy, at higher volumes, while ensuring every decision is defensible. Denial reasons must be specific. Metrics must be reportable. And as of 2026, that performance is no longer an internal operational matter. It is public.

In that context, explainability is not a feature. It is the entire point. Payers who begin evaluating glass-box, explainable AI architectures now, even before the technology reaches full maturity, will be better positioned to meet regulatory demands, rebuild provider trust, and gain a competitive advantage built on transparency rather than speed alone.

The Path Forward

Prior authorization does not need another black box. It needs systems that can show their work: read clinical evidence, apply policy logic transparently, document every decision, and escalate what they cannot resolve with confidence. But rented explainability is still a governance gap. Seeing the reasoning on a single decision is not the same as owning the model, controlling the policy logic, or auditing the architecture across your AI estate. Payers who begin evaluating what they actually own, not just what they can observe, will be better positioned for what the regulatory environment is now requiring.

The Same Engine, Different Badges

Strip away the branding, and every PA agent on the market is the same thing under the covers: a large language model with instructions, wrapped in a vendor’s interface. What payers accumulate, as they add specialized vendors for PA, UM, appeals, and member services, is fragmentation, five audit formats and five contracts, with no unified view of how AI is making decisions at exactly the moment CMS is requiring those decisions to be public. ThoughtFocus AI Workforce takes a different position: we build agentic AI for regulated workflows under a model where the payer owns the logic that makes the decisions, not just the right to use it. Renting a black box was the old problem. Renting five glass boxes is the new one. The conversation usually starts with a single-workflow proof of concept, owned by the payer at the end.