AI in Healthcare: Prior Authorization - When Both Sides Automate

Contributors: Wren Keber and Lisa Soroka
To learn more about Wren and Lisa, click here.

 

Intro

In Part Two of our series “AI In Healthcare,” we turn to prior authorization (PA), which can arguably be the clearest test of whether AI can deliver on a promise to reduce or lighten administrative burden, or merely enable payers and providers to each perform more of it.

Both sides agree on the facts: prior authorization runs high in volume, is heavily rules-based, can be slow in execution and frustrating for practically everyone, including patients. Medicare Advantage Organizations (MAOs) alone issued nearly 53 million prior authorization determinations in 2024, according to CMS data. The AMA cites that physicians complete as many as 40 requests per week and independent practices can spend between 10 and 13 hours of practice time reviewing and responding to initial and review requests. This is valuable time spent on administrative tasks, not delivering clinical care. Therefore, we believe no other administrative process presents more of an obvious testing ground for effective automation.

Why Both Sides Automate
Payers automate for arguably defensible reasons. Machine screening applies coverage and medical policies consistently across enormous claim volumes, flags missing documentation, identifies outlier utilization, and resolves straightforward cases; so human clinical reviewers can concentrate on genuine exceptions, escalations and/or complicated requests. Managing finite premium dollars is a payer’s primary obligation and must be treated accordingly, which must also be considered in the context of each health plan’s obligations to cover high quality, clinically appropriate care.

Providers automate for equally defensible reasons. Revenue cycle teams now deploy tools that assemble payer-specific packets, predict which documentation a given plan will require (and under what circumstances), track deadlines, draft appeal language, and escalate the cases that need the rendering provider’s voice. The goal is faster approval and payment, and above all timely and consistent patient access. PA-related care delays are becoming more common because, for example, bad debt associated with PA-related denials has prompted providers not to schedule a patient visit until the request is approved by the plan.

Both providers and payers are right to automate. We believe the question is: “What happens when the two systems start answering each other at scale using AI agents or other automation means?” In the end, both providers and payers have a shared goal: reduce administrative overhead in general, including that associated with prior authorizations, not simply manage the cost of it.

The Emerging “AI-Versus-AI” Workflow
Payer automation now reaches well past initial request screening into claims edits, documentation reviews, pre-payment edits, reauthorization during inpatient stays, and post-payment audits for potentially up to seven years after service. Standardized criteria improve consistency on routine cases. AI handles atypical presentations far less gracefully including pediatric and neonatal complexity, and other requests with fast-moving or co-morbidly complex justification to name a few. A Senate subcommittee investigation traced that gap at one large Medicare Advantage insurer, where post-acute denial rates climbed from 8.7% to 22.7% over three years alongside the adoption of predictive tools.

The dynamic no longer belongs to Medicare Advantage organizations (MAOs) and commercial payers alone. On January 1, 2026, CMS launched the WISeR Model, which applies AI and machine learning to prior authorization for selected services in Original Medicare across six states through 2031. Like medical necessity denials in general, the CMS pilot requires every non-affirmation decision to be made by licensed clinicians. They also made an allowance for consistently compliant providers  - exemptions from PA for following pre-determined rules (often called “gold-carding”).

Providers answer with counter-automation, and the sheer volumes explain why. According to data released by the consulting firm Premier, MAOs denied 4.1 million Medicare Advantage PA requests in 2024, or 7.7% of determinations, up from 6.4% a year earlier. Providers and patients appealed only 11.5% of those denials. Plans then overturned more than eight of ten appeals they received. That pairing points to a large population of denials that are not durable when scrutinized. Most mature automation offerings will reduce the time required to generate an "as comprehensive as possible” initial authorization request and will learn what the overturn reasons are so that initial requests are tailored, automatically, to each payer’s policy. 

What to Measure
Executives cannot measure the impact of AI in prior authorization with consolidated denial rate alone. Healthcare Financial Management Association (HFMA)’s Claim Integrity Task Force standardized six denial metrics so organizations can benchmark consistently and detect change early. In addition to the HFMA metrics, we would add three metrics that measure the impact of automation directly: these should all be tracked by payer, service line, and by patient population (plan ID). So, in all, we suggest these metrics may be a reasonable starting point:

Metric

Automation Moves the Metric

Initial (administrative; at submission) denial rate

Flat

Primary (fully-adjudicated) denial rate

Flat

Denial write-offs as a share of net patient service revenue

Down

Days from initial denial to appeal submission

Down

Days from initial denial to claim resolution

Flat

Percentage of initial denials overturned

Up

PA turnaround time by payer and service line (in days)

Down

Provider (difficult to track) and/or revenue cycle staff hours consumed per PA

Down

Time from order to service (or time from PA submission to scheduled appointment)

Down

 

Data Access, Human Judgment, and Guardrails
Payers increasingly request direct electronic medical records (EMR) access, justified by the efficiency argument. Direct access can end repetitive document chasing and give reviewers a fuller clinical picture. The exposure deserves equal weight. Most legacy payer agreements predate automated decisioning entirely, and they rarely specify scope, minimum necessary data, retention, secondary use, retrospective mining, audit logging, or termination. These are all necessary guardrails that must be in place when a provider is considering granting transactional access to their EMR. 

For any automation that a provider may be considering, we recommend that leaders press on four fronts:

  1. Include outcome requirements in vendor contracts. Specify response times, data-use limits, retroactivity rules, audit rights, and advance notice of model or policy changes.
  2. Make human review an imperative. Define which cases may clear automatically and which must require a clinician; then confirm that exceptions reach an empowered human quickly.
  3. Design smart queues. Route a request by the kind of judgment it demands, not by dollar value. Routing the right requests to the appropriate humans for easy context and action, in order to enable human review that’s high-value, not merely “overseeing the agent.”
  4. Name accountable owners on both sides. Responsibility should diffuse easily among the model, vendor, payer policy, and the reviewer.
  5. Periodic Review. Vendors should be expected to conduct human reviews on a reasonable cadence to ensure that any decisions and outputs from a fully agentic model are operating within the expected boundaries. What about someone to actually overseeing the agentic system on a routine basis (not just occasional small-sample audits)?
  6. Appropriate Risk Management and Insurance. Vendors must bear the responsibility to make sure that agents do not “go rogue” outside the agreed-upon business rules. If there are any damages, the provider or payer (the customer) would expect the vendor to hold them harmless and cover liability through insurance or other protections.

We recommend buying or building tools that eliminate avoidable work and expose root causes. Automating appeals after initial denial treats only the symptom. Lastly, it’s important to negotiate vendor terms with a long horizon and a focus that includes cost of tokens, because the capital backing today’s AI eventually needs a return, and price increases (which could be substantial) are inevitable and need to be anticipated.

Conclusion: Redefine Success During Times of Change
In the prior authorization space, it is important to recognize that regulation goal posts keep moving. Since January 1, 2026, CMS will enforce the requirement for MAOs to 1) decide expedited requests within 72 hours and standard requests within seven calendar days, 2) to give a specific reason for every denial, and 3) to publish their prior authorization metrics as a scorecard. Payer application protocol interfaces (APIs) will be enforced in 2027, plus an April 2026 proposed rule would extend electronic prior authorization (ePA) to drugs by October 2027. Four MAOs separately pledged to cut prior authorization requirements, and have reported they have eliminated about 11% so far.

We believe leaders on both sides should reject the arms-race framing outright. Success is not merely a faster denial or appeal. Success means routine approvals are received quickly, there are fewer unnecessary exchanges among billing provider and payer, there is genuine human review of complex justification, faster time to payment occurs, and patients experience less delay. When AI delivers on all of these success factors, it will have begun to live up to its promise relative to administrative burden and associated costs.

 

Contact Wren at: [email protected]
Contact Lisa at: [email protected]