If you practice medicine in 2026, the phrase “prior authorization” probably makes you tired before the sentence ends. You have sat on hold with a payer. You have written an appeal letter for a patient who obviously needed the procedure. You have read a denial and wondered whether anyone on the other end actually opened the chart.
Now the Centers for Medicare and Medicaid Services (CMS) is testing something different. A pilot program inside Traditional Medicare hands a large share of those judgment calls to an artificial intelligence model instead of a human reviewer. Physician organizations have called the underlying logic a black box. Two members of Congress have moved to shut the pilot down before it spreads to the rest of the country. The vote ahead is not really about one program. It will set the rule for how Medicare AI prior authorization behaves for the next decade. This piece walks through what the pilot actually does, what the peer-reviewed evidence says, and what a clinician can do in clinic this week.
Understanding the Medicare AI Prior Authorization Pilot
The program in question is the Wasteful and Inappropriate Service Reduction model, known as WISeR. Until WISeR, only Medicare Advantage plans were using automated systems to triage prior authorization requests. Traditional Medicare patients sat largely outside that loop. WISeR changes that. The pilot covers six states and roughly 13 services that CMS has flagged as low-value, including knee arthroscopy for osteoarthritis and certain skin substitute applications. A 2026 review by Mahajan A et al in J Am Acad Dermatol walks through the pilot’s specific implications for dermatologic care and the skin substitute denials clinicians are already starting to see.
The administrative argument is simple. Medicare loses billions each year on services with limited real benefit. An algorithm can flag a request faster than any human reviewer, route the borderline cases up the chain, and let the clean ones through. A recent informatics overview by Keng C et al in J Med Syst makes the case that AI and automation, applied well, can take real load off both clinicians and patients across exactly these workflows. The detail that bothers physicians sits one layer beneath that promise. The contractors who run the AI under WISeR are paid in proportion to the dollars they save Medicare. The contract language for that payment design is averted expenditures. Once you see the term, the rest of the policy story reads differently.
How the Averted Expenditures Incentive Reshapes Denial Behavior
Pay-for-denial has an uncomfortable history in private insurance, and it is the exact pattern WISeR imports into Traditional Medicare. The contractor’s revenue rises when a request is denied or scaled down. When the underlying decision engine is a model rather than a person, that incentive does not just shape what gets approved. It shapes which training data, which appeal threshold, and which automated re-review policy ends up baked into the production pipeline.
In effect, the contract turns the AI from a neutral triage tool into a financially interested actor. The contractor may quietly pass a request through to a physician reviewer only when the model already leans toward denial. The model may tighten its own decision boundary as new claims come in, because tighter boundaries score better against the contract metric. The public has no visibility into how often that retraining happens, what the appeal-overturn rate looks like, or whether the model treats rural and urban claims the same way. Physician groups have spent two years asking CMS for that transparency, with little to show for it.
Why Physician Groups Are Calling WISeR a Black Box
The term “black box” gets thrown around loosely, so it is worth saying what physicians actually mean by it here. They mean three specific gaps. First, no clinician outside the contractor knows the features the model uses to weigh a claim. Second, no clinician outside the contractor knows the threshold at which a request shifts from auto-approve to human review to auto-deny. Third, no clinician outside the contractor can see which version of the model issued a given denial, which means a peer-to-peer call about a denied case is, functionally, a call about a decision no one in the room can fully explain.
A 2026 analysis by Mello MM et al in Health Affairs walks through exactly this terrain across the broader payer landscape. The authors lay out how opacity of algorithmic determinations, weak humans-in-the-loop arrangements, and automation bias combine to make AI-driven utilization review hard to govern responsibly, even before incentive design gets layered on top.
The pilot also does not give beneficiaries a parallel layer of recourse. A Medicare patient denied a procedure by an algorithm can still appeal through the standard Medicare appeals chain, but the appeal does not interrogate the AI itself. It interrogates the paperwork the AI produced. The appeal asks whether the documentation supports the denial, not whether the algorithm should have produced the denial in the first place. That asymmetry sits at the heart of the physician objection.
The Congressional Resolution of Disapproval
In May 2026, Senator Ron Wyden and Representative Suzan DelBene introduced a resolution of disapproval under the Congressional Review Act (CRA) targeting the WISeR pilot. The CRA is a narrow tool. It lets Congress overturn a federal regulation by simple majority within a limited window after the rule takes effect, and it bars the agency from issuing a substantially similar rule without explicit congressional authorization. If the resolution passes both chambers and is signed, WISeR ends, and CMS cannot quietly re-launch it under a different name in the next budget cycle.
The political math is real. A floor vote forces every senator and every house member to take a public position on whether automated denials belong in Traditional Medicare. Even if the resolution fails to pass, the vote itself has consequences. It puts a marker in the record that physicians and their professional societies can point to the next time CMS proposes a similar program. The American Medical Association, the American Hospital Association, and several specialty societies have written letters in support, which is a rare alignment for groups that often disagree on payer policy.
What Peer-Reviewed Evidence Says About AI in Medicare Prior Authorization
The strongest empirical case against expanding Medicare AI prior authorization comes from a 2026 analysis in NPJ Digital Medicine. Raza S et al titled their paper “Medicare advantage becoming a disadvantage with use of artificial intelligence in prior authorization review,” and the framing is not accidental. The authors examined how AI-driven prior authorization in Medicare Advantage interacts with patient case mix and clinical nuance, and they document a pattern they call algorithmic disadvantage.
The mechanism is worth describing in clinical terms. A prior authorization model is trained on past approvals and denials. The training data already encodes the historical biases of payers, including underutilization in populations that have historically had less access to imaging, less access to specialist visits, and less access to documented follow-up. When the model learns from that distribution, it inherits the same biases and applies them prospectively, including to populations the original payer rules were never built for. The model also tends to miss individual clinical nuance, such as multiple failed conservative therapies or atypical presentations that a human reviewer would have caught in two minutes of chart review.
The authors close on a policy point clinicians should read carefully. Without external audit, public reporting of denial and overturn rates, and a meaningful human-in-the-loop requirement, automated prior authorization in federal payer workflows is unlikely to produce the cost savings CMS expects, and is likely to produce measurable harm in the patient populations least equipped to navigate the appeals process.
Clinical Services in the Crosshairs
The 13 services WISeR targets are not random. They are the procedures CMS has flagged in past Office of Inspector General reports and Medicare Payment Advisory Commission analyses as having high volume and questionable marginal benefit in certain patient subgroups. The list includes knee arthroscopy for osteoarthritis, skin substitute applications for chronic wounds, certain spinal injections, certain electrical nerve stimulation devices, and a handful of cardiology and pain management procedures that have been the subject of overuse debates for years.
The clinical problem is that “low-value on average” is not the same as “low-value for this patient.” A knee arthroscopy in an active 62-year-old with mechanical symptoms after a documented meniscal injury is a different procedure from a knee arthroscopy in a sedentary 78-year-old with osteoarthritis and no mechanical findings. Both can be coded the same way. Both can hit the same denial pipeline. Only one of them is the procedure the literature on low-value care was actually written about. WISeR does not currently publish the patient-level features the model uses to distinguish those cases, which means a clinician cannot write a stronger preauthorization packet by addressing the model’s actual concerns.
What a Black-Box Denial Looks Like at the Bedside
Picture a 70-year-old patient in your office on a Tuesday afternoon. Active, golfs twice a week, lives independently, had a meniscal tear three months ago, has been through six weeks of physical therapy with documented failure, now has persistent mechanical locking. You document everything in the visit note. You submit the prior authorization. Forty-eight hours later, an automated denial comes back citing low-value care criteria, with a generic reference to a published guideline. There is no specific reason in the letter, no specific feature of the patient’s case the model weighed, and no specific corrective action you can take.
What do you do next? You can appeal. You can request a peer-to-peer. You can re-submit with additional documentation. Each of those steps costs you and your staff measurable time, and the patient may be in pain through every one of them. The literature on prior authorization burden suggests that the time cost is precisely the lever payers rely on. If appeals were free, the volume of overturned denials would rise. Because appeals are expensive in staff time, many appropriate denials are never appealed, and the model’s overall denial rate looks lower than the medically appropriate rate would actually be.
How Physicians Can Navigate and Appeal Automated Denials
There are practical steps that work. They are not magic. They do not undo the structural problem the WISeR pilot creates. They do meaningfully improve the odds for individual patients while the policy debate plays out.
Start with documentation built for an algorithmic reviewer, not only for a human one. Put the conservative therapies, their durations, and their documented failures into structured fields rather than free-text narrative whenever the EHR allows it. Include the specific guideline language the procedure satisfies, with a citation, in the prior authorization packet itself. The model will not read your prose. It will read your codes and your fields. A 2025 analysis by Lavoie-Gagne O et al in Arthroscopy makes a similar point about standardizing prior authorization workflows and aligning documentation with the way automated reviewers actually parse claims.
The second move is the peer-to-peer. It is your only direct contact with a physician on the contractor side, and it is the step where automated denials are most often overturned. Prepare for it the way you would prepare for a brief consult. Have the chart open. Have the imaging visible. Have one or two sentences in mind about why this specific patient does not fit the pattern the model flagged. Peer-to-peer availability under WISeR varies, so build the request into your workflow the moment a denial lands.
Document your appeals systematically. Many practices now keep a per-payer denial log. When the WISeR pilot publishes its first interim report, those logs will be the evidence base your specialty society uses to push CMS for changes. Your state medical association may already be collecting denial data for the federal record. Reach out and contribute.
There is also a useful conversation to have with the patient before any of this. The Medicare beneficiary may not realize that the denial was issued by an algorithm. They may assume a person reviewed the chart. Telling them honestly that an automated system produced the first decision, and that a clinician on the contractor side will only review it if you request the peer-to-peer, changes the conversation. The patient may choose to be involved in the appeal in a way that strengthens it, such as filing a parallel complaint with their Medicare Beneficiary Ombudsman or contacting their congressional representative. Patient-led complaints carry weight in the CRA debate that physician complaints alone often do not, because the elected representatives in question vote with constituents in mind. The chart note that wins an appeal is the one that describes the patient first and the procedure second.
Conclusion
The vote ahead on the WISeR pilot is the most consequential moment for Medicare AI prior authorization since the original Medicare Advantage prior authorization rules were written. The peer-reviewed evidence already shows that algorithmic disadvantage is real, that it falls hardest on the patients least able to navigate it, and that the pay-for-denial structure baked into the pilot makes the problem worse rather than better. The Congressional resolution of disapproval is not the only path forward, but it is the path that puts every elected representative on the record at a moment when professional societies are aligned and the empirical case is strong. In the meantime, physicians can keep documenting for the model, prepare carefully for peer-to-peer, and contribute to the denial logs that will shape the next round of policy. The clinical community has organized on this kind of issue before and prevailed, and the same patient-centered case can carry this one across the finish line.
References
Raza S et al (2026). Medicare advantage becoming a disadvantage with use of artificial intelligence in prior authorization review. NPJ Digital Medicine. PMID: 41639200. DOI: https://doi.org/10.1038/s41746-026-02387-x. PMC: PMC12979811.
Mahajan A, Whittelsey M, Nambudiri VE, LaChance AH (2026). The Wasteful and Inappropriate Service Reduction model and its potential impacts on dermatologic care. J Am Acad Dermatol. PMID: 41935786. DOI: https://doi.org/10.1016/j.jaad.2026.03.099.
Mello MM, Trotsyuk AA, Mahamadou AJD, Char D (2026). The AI Arms Race In Health Insurance Utilization Review: Promises Of Efficiency And Risks Of Supercharged Flaws. Health Affairs 45(1):6-13. PMID: 41494115. DOI: https://doi.org/10.1377/hlthaff.2025.00897.
Keng C, DiGiorgio A, Ehrenfeld JM, Spear J, Miller BJ (2025). Unburdening Patients and Clinicians Through Automation and Artificial Intelligence: Informatics Strategies for Reducing Administrative Burden. J Med Syst 49(1):128. PMID: 41055647. DOI: https://doi.org/10.1007/s10916-025-02265-1. PMC: PMC12504360.
Lavoie-Gagne O, Woo JJ, Williams RJ 3rd, Nwachukwu BU, Kunze KN, Ramkumar PN (2025). Artificial Intelligence as a Tool to Mitigate Administrative Burden, Optimize Billing, Reduce Insurance- and Credentialing-Related Expenses, and Improve Quality Assurance Within Health Care Systems. Arthroscopy 41(8):3270-3275. PMID: 40120727. DOI: https://doi.org/10.1016/j.arthro.2025.02.038.
Licensed physician and clinical AI specialist. Founder and Editor-in-Chief of ZayedMD, a physician-led medical publication covering clinical AI, neurology, metabolic health, and evidence-based patient guidance.
