The Prior Authorization Arms Race

In October 2024, the United States Senate Permanent Subcommittee on Investigations published a report showing that between 2019 and 2022 the three largest Medicare Advantage insurers increased denial rates for post-acute care and relied on algorithms to help manage these decisions. For one particular insurer, the post-acute denial rate jumped from under nine percent to over twenty-two percent. In 2022, the Office of Inspector General (OIG) of the Department of Health and Human Services reported that, on average, thirteen percent of prior authorization requests denied by Medicare Advantage plans would have most likely qualified for Medicare coverage if the determination had been made by a Medicare employee.

Together, the two reports illustrate a similar change in the process. The choice to approve or deny care is shifting from a person analyzing a chart to a system evaluating a request. Providers, after losing the battle a fax at a time, have invested in their own systems to fight back.

The Fight Became Machine Against Machine

For most of the prior authorization process, the burdens were placed on clinicians and their staff. The American Medical Association (AMA) reported in its 2024 survey that more than eighty percent of physicians had patients leave without receiving recommended treatment while the approval was pending. In that survey, practices estimated they spent on average thirteen hours a week per physician, time that was completely devoid of clinical value.

What is different now is how technology has flipped the roles. A payer can use AI to figure out which denials get appealed and which get ignored. On the other side, a provider can use denials management software to draft, document, and submit an appeal in a matter of seconds. That same provider can use automated medical coding to assign billing codes directly from a clinical note. All of this technology can exist without hiring staff, which is the bulk of the appeal and almost all of the risk.

Considering these implementations separately leads to logical conclusions. Combined, though, all of these systems create an almost paradoxical reality. A clinician makes a request, one model denies it, a second model decides the appeal is warranted and files it, and in the interim the patient is left waiting.

Policies have tried to address the denial issue. The 2024 CMS prior authorization rule requires payers for government programs to decide urgent requests within seventy-two hours and standard requests within seven days, and to provide a justification for any denial, with the new time limits taking effect in 2026. This addresses the edge of the issue. It does nothing about the larger question of whether a decision should be left to a model at all, and nothing about the appeal model the provider has implemented to counter it.

What the Health System Still Owns

Here is what remains once the rest of the work gets automated. When a health system files an automated appeal, it is submitting a claim to a payer for a specific patient, and it answers for that claim. Let a model do the coding, and the health system is making a representation to a government program or a commercial payer, and it owns the truth of that representation either way. Enforcement of federal payment integrity in the healthcare billing system has been around for a long time. A model that codes a task more quickly does not change who is answerable for signing off on the results.

This is where governance has to be something operational, or it becomes a dusty document. A denial model the payer will not explain and an appeal model the provider never audits are the same problem. Each generates a confident answer that gets accepted, often because the reviewer is tired at the end of a long day in the clinic. Neither gets caught until a patient is hurt or an auditor shows up.

The Data Underneath Decides Everything

An appeal is only as strong as the record it pulls from. Coding is only as accurate as the document it reads. When the underlying data is inconsistent across systems, incomplete, or out of date, automating on top of it does nothing but worsen the problem and add a gloss.

The Agency for Healthcare Research and Quality (AHRQ) has contended for some time that the clinical record is a safety tool, and a record created for billing efficiency rather than clarity will confuse any model that examines it. A health system that intends to apply AI to the revenue cycle must first answer a less than attractive question. Does our data capture what actually occurred with the patient well enough that a machine could act on it without filling in the gaps.

Most systems are unable to answer that question confidently today. That is the true challenge, and it sits underneath every vendor demonstration.

I learned this in back office billing, long before anyone coined the term AI. There was always a problem with certain codes, and certain insurers always denied the claims. I started to track this in a notebook, and what first appeared to be a problem with numbers turned out to be a problem with definitions. Two departments used the same code to represent completely different concepts. No model would have solved that. A faster system would have submitted the discrepancy at high volume and considered it productive. The staff worked around it by hand for months because no one went looking for the cause, only the symptom.

What Leaders Should Build Before Automating

Stop thinking of revenue cycle AI as disjointed tools and treat it as an ongoing program with owners. Assign responsibility for the integrity of the automated denials and appeals, and for the codes a model assigns, the way a thoughtful operation assigns responsibility for any high-volume process involving money and patients. Sample the model output the way you would any other process you cannot completely observe, watching for a decline in accuracy and for changes in what gets denied and what gets overturned.

As always, the foundation is data. It needs to be of good quality before broad automation occurs, since every model relies on the quality of the data beneath it. This is the bulk of our work with health systems, and it is critical to the long-term success of the system even though it sits as a low priority in the vendor demonstration.

Insurance companies and health systems will keep enhancing their models. The challenge is, and will continue to be, ownership of the outcome, whether it is automated or not.

Context and Sources

This edition relies on the work of the United States Senate Permanent Subcommittee on Investigations regarding Medicare Advantage denials, OIG reports on Medicare Advantage prior authorization denials, the AMA prior authorization surveys, AHRQ studies on the clinical record as a safety mechanism, and federal payment-integrity enforcement. It continues themes from issue 13, "From Prior Auth to Predictive Care," which examined this change from the payer perspective, and the 06-15 edition on what clinicians still retain after an AI has drafted the work.

Christopher Hutchins Founder & CEO, Hutchins Data Strategy Consultants