The letter arrived on a Tuesday. It was not addressed to a doctor. It was addressed to a patient — a sixty-seven-year-old man named Gerald, a retired machinist from Scranton who had spent the previous nine days in a skilled nursing facility recovering from a hip replacement. The letter informed Gerald that his insurer had reviewed his case and determined that continued coverage of his skilled nursing stay was no longer medically necessary. He had seventy-two hours to appeal or arrange discharge. The letter was dated the previous Thursday. It had taken five days to arrive.
The appeal process, as explained in the letter, required Gerald to submit a written request with supporting clinical documentation within the seventy-two-hour window that had already, by the time the letter reached him, partially elapsed. His discharge planner — a social worker named Theresa who managed forty-two patients across two floors of the facility — helped him begin the process. The appeal was denied. Gerald was discharged four days ahead of the clinical schedule his orthopedic surgeon had recommended. He was readmitted to the emergency department eleven days later with a wound infection.
This is not a story about an exceptional failure. This is a story about a system working precisely as designed.
I. What Prior Authorization Was Built to Do
Prior authorization is, in theory, a clinical quality mechanism. Insurers require advance approval for certain services — specific drugs, procedures, specialist referrals, post-acute care placements — to ensure that the requested care is appropriate, evidence-based, and necessary. The process was designed to prevent waste: to stop patients from receiving ineffective treatments, to ensure expensive interventions were matched to patients who would genuinely benefit.
This is what it was built to do. It is not what it does.
What prior authorization does, in the aggregate, in the American healthcare system in 2026, is function as a systematic mechanism for delay, attrition, and denial. The physician submits the request. The insurer reviews it — or does not review it, as we will see — and issues a determination. The determination may arrive in hours, or in days, or not at all. If the determination is a denial, the physician or patient may appeal. The appeal takes time. The patient, in the meantime, does not receive the care.
"Prior authorization has evolved from a tool to ensure appropriate care into an administrative weapon deployed against both physicians and the patients they are trying to treat."
— American Medical Association, 2024 Prior Authorization Physician Survey
The American Medical Association has surveyed physicians on prior authorization every year since 2018. The findings do not improve. In 2024, 94 percent of physicians reported that prior authorization delays care. Thirty-three percent reported that prior authorization had led to a serious adverse event for a patient in their practice — a hospitalization, a disability, in some cases a death. The average physician's practice now submits forty-five prior authorization requests per week. Physician offices employ, on average, nearly two full-time staff members dedicated exclusively to prior authorization paperwork.
— AMA Prior Authorization Physician Survey, 2024
These are not numbers describing a system that is occasionally malfunctioning. These are numbers describing a system that has been designed to produce a particular outcome — and that is producing it, reliably, at scale.
II. The Machine That Never Read the Chart
In March 2023, ProPublica published an investigation into CIGNA's automated prior authorization review system, known internally as PXDX. The findings were specific and documented. CIGNA's system allowed its physicians to deny claims without opening the patient's file. In a ten-month period, CIGNA physicians denied over 300,000 requests after spending an average of 1.2 seconds reviewing each one.
One point two seconds. A physician cannot read a patient's name in 1.2 seconds. A physician cannot open a chart in 1.2 seconds. What a physician can do in 1.2 seconds is click a button.
The process worked like this: the PXDX system flagged claims that matched certain diagnostic and billing code combinations. It presented those claims in batches to reviewing physicians. The physicians — licensed clinicians, credentialed and salaried — worked through the queue. The algorithm had already made the determination. The physician's function was to provide the signature that made the determination legally defensible.
The chart was not opened. The patient was not seen. The medicine was not considered. The machine was certain, and the physician agreed.
CIGNA disputed the characterization and maintained that its process complied with applicable regulations. The company settled related litigation without admitting wrongdoing. The PXDX system, in some form, continued to operate.
The CIGNA investigation was not an isolated exposure. In November 2023, a class action lawsuit was filed against UnitedHealth Group alleging that the company had used an artificial intelligence model called nH Predict to systematically override physician judgments for Medicare Advantage patients requiring post-acute care. The complaint alleged that the model had a known error rate — that it generated incorrect discharge recommendations in cases where treating physicians disagreed — and that UnitedHealth had deployed it anyway, at scale, because even a flawed algorithm that denies faster than human review cuts costs.
The lawsuit alleged that patients were discharged from skilled nursing facilities before they were clinically ready. That some of them were readmitted. That some of them died. UnitedHealth denied the allegations. The litigation is ongoing. The AI model is still deployed.
in a ten-month internal audit period covering 300,000+ denials
— ProPublica, March 2023
III. Medical Necessity, As Defined by the Algorithm
The legal and regulatory architecture of prior authorization rests on a phrase: medical necessity. Insurers are required, under federal and most state laws, to cover services that are medically necessary. The Affordable Care Act enshrined this principle. Medicare is built around it. Medicaid is built around it. Every commercial insurance product sold in the United States contains, somewhere in its language, a commitment to covering care that is medically necessary.
What the law does not specify, with adequate precision, is who decides.
Each insurer maintains its own definition of medical necessity. Each insurer maintains its own clinical criteria — the specific conditions under which a given service will be approved or denied. These criteria are proprietary. They are not public. They are not consistently aligned with clinical guidelines published by specialty medical societies. They change without notice. A treatment that was approvable under last year's criteria may be deniable under this year's criteria, without any change in the clinical evidence base.
AMA, 2022: A study of prior authorization criteria for chemotherapy found that insurer clinical criteria diverged from NCCN (National Comprehensive Cancer Network) guidelines in 40% of cases reviewed — with insurer criteria being more restrictive in the majority of divergences.
JAMA Internal Medicine, 2021: Analysis of Medicare Advantage prior authorization denials found that 13% of denied services met Medicare coverage rules and should have been approved — representing millions of improperly denied claims annually.
HHS Office of Inspector General, 2022: Found that Medicare Advantage plans denied 13% of prior authorization requests that met Medicare coverage criteria. Of those that were appealed and overturned, 75% were reversed in the patient's favor — meaning the original denial was wrong.
The clinical criteria that drive these determinations are, in many cases, not written by physicians. They are written by actuaries. They are validated by algorithms. They are operationalized by automated review systems that match diagnosis codes and billing codes to decision trees and return approvals or denials at machine speed. The physician who ordered the treatment is not part of the process until the process has already failed — until the denial has already been issued and the clock on the appeal window has already begun to run.
This is the gothic architecture of the prior authorization system: it presents itself as a clinical instrument while functioning as a financial one. The machine does not speak the language of medicine. It speaks the language of codes. And it has been optimized, with considerable sophistication, to speak that language in ways that produce a particular answer.
IV. The Appeal and Its Attrition
The prior authorization system is built around an assumption that it will not often be challenged. This assumption is, empirically, correct.
The appeal process is available. It is also, by design, arduous. It requires documentation that busy clinical practices may not have the capacity to compile within the required timeframe. It requires physicians to spend time arguing, in writing, with a system that denied the claim without reading the chart. It requires patients — many of them elderly, ill, or recently discharged — to navigate a bureaucratic process whose complexity is not accidental.
The attrition is the point. When fewer patients appeal, fewer denials are overturned. When fewer denials are overturned, the insurer pays fewer claims. The appeal process is not a safety valve. It is a filter. It is calibrated to pass only those cases whose advocates have the capacity, the knowledge, and the stamina to press through it — which is to say, not Gerald, who is sixty-seven years old and recovering from surgery, and whose social worker manages forty-two patients, and whose letter arrived five days after it was dated.
"The system works exactly as intended. The question is what it was intended to do."
— Former Medicare Advantage medical director, speaking on condition of anonymity
When the HHS Office of Inspector General examined Medicare Advantage denials in 2022, it found something remarkable: 75 percent of prior authorization denials that were appealed and overturned had been overturned in the patient's favor. Three in four appeals that succeeded were cases where the original denial was simply wrong. Not disputed. Not a close clinical call. Wrong. The algorithm had said no to something Medicare would have covered. The patient had appealed. The appeal had won.
But most patients do not appeal. The OIG found that Medicare Advantage enrollees appealed less than 12 percent of denied claims. Which means that for every three patients who successfully overturned a wrongful denial, twenty-three patients accepted a wrongful denial without challenge. They went without the care. Or they paid for it themselves. Or they deteriorated. Or they returned to the emergency department.
V. Gold Carding and the Limits of Reform
There is a reform movement. It is real and it is insufficient.
Seventeen states have enacted gold carding laws — statutes that exempt physicians with high prior authorization approval rates from the prior authorization requirement for services within their specialty. The logic is sound: if a physician's prior authorization requests are approved ninety percent of the time, requiring that physician to submit prior authorization requests is not a quality assurance mechanism. It is paperwork. It is a toll. Gold carding removes the toll for physicians who have demonstrably earned the exemption.
The CMS finalized its Interoperability and Prior Authorization Rule in January 2024, effective for most provisions in 2026 and 2027. The rule requires Medicare Advantage plans, Medicaid, and CHIP to respond to urgent prior authorization requests within seventy-two hours and standard requests within seven calendar days. It requires plans to give specific reasons for denials. It requires them to report prior authorization data publicly — approval rates, denial rates, appeal rates — in a format that enables comparison.
Effective 2026: Medicare Advantage plans, Medicaid, and CHIP must respond to standard prior authorization requests within 7 calendar days; urgent requests within 72 hours.
Effective 2026: Plans must provide specific reasons for denials in all prior authorization determinations.
Effective 2027: Plans must implement Prior Authorization APIs using HL7 FHIR R4 standards, enabling electronic submission and status tracking.
Ongoing: Plans must publicly report prior authorization metrics — approval rates, denial rates, average response times, and appeal outcomes — enabling transparency and accountability.
These are meaningful requirements. They are also, in the context of what the prior authorization system has become, modest. A seven-day response window does not address a system that denies first and processes appeals selectively. A requirement to state reasons for denials does not address a system whose reasons are generated by algorithms trained on actuarial data rather than clinical evidence. A FHIR API for prior authorization submission makes the process faster. It does not make the decision-making inside the process better.
The machine will still be certain. It will simply be certain faster.
VI. What the Data Reveals
The prior authorization system has generated an enormous body of data. Approval rates. Denial rates. Appeal rates. Overturn rates. Response times by plan, by service category, by geography. This data is, increasingly, public — required by the CMS rule, compiled by state insurance commissioners, analyzed by advocacy organizations and academic researchers.
What the data shows, in aggregate, is a system whose denials are distributed in ways that cannot be explained by clinical variation. Medicare Advantage plans deny prior authorization requests at rates that differ, plan to plan, by factors of four and five for identical service categories. The plans that deny most frequently are not the plans whose members have the worst clinical outcomes. They are, in some analyses, the plans whose members have the best financial outcomes — for the plan.
The data also shows something about the texture of denials that aggregate statistics tend to obscure: they cluster. They cluster around specific diagnoses. Around specific drug classes. Around specific post-acute care categories. Around specific patient populations — older, sicker, costlier patients whose care trajectories are more expensive and less predictable. The algorithm is not random. It has been trained on something, and what it has been trained on has shaped what it denies.
Understanding these patterns — building the synthetic data infrastructure that allows researchers, compliance teams, and health system administrators to model prior authorization behavior without touching protected patient records — is part of what PatientDatasets.com was built to enable. The denial landscape is navigable. But navigation requires a map.
VII. The Horror in the Spreadsheet
The prior authorization system is not cruel in the way that a villain is cruel. It is not animated by malice. It operates according to logic — the logic of the actuarial table, the logic of the quarterly earnings call, the logic of a system that has optimized for a particular set of outcomes and has done so with considerable competence.
This is what makes it genuinely frightening. Horror that requires a monster is comprehensible. You can identify the monster. You can confront it. What Poe understood — what the best gothic literature has always understood — is that the most durable dread is not located in the exceptional. It is located in the institutional. It is the House of Usher, not the haunted wood: a structure so committed to its own logic that it cannot perceive its own decay. A system so thoroughly optimized for its own functioning that the people trapped inside it cannot see what it has become.
The prior authorization system is that structure. It processes tens of millions of requests each year. It employs thousands of people, most of whom are doing their jobs correctly. It operates within legal frameworks. It responds to regulatory pressure. It has appeals processes and compliance departments and medical directors who believe, sincerely, that what they are doing serves some defensible purpose.
And every year, in physicians' offices and hospital discharge lounges and skilled nursing facilities across the country, patients like Gerald receive letters on Tuesdays. The letters are certain. The machine was certain. And by the time the letter arrives, the window is already closing.