When Deborah Darling’s husband needed treatment for a rare autoimmune disease, their insurer denied the claim. The denial came fast — faster than any human reviewer could have reasonably assessed the case. What Darling didn’t know at the time was that the decision may have been shaped not by a physician or even a claims adjuster, but by an algorithm trained to say no.
She’s far from alone.
Across the American insurance industry, artificial intelligence systems are now embedded at nearly every stage of the claims process — from initial filing to final denial. These tools promise efficiency and consistency. What they’re delivering, according to a growing body of evidence from regulators, journalists, and patient advocates, is something more troubling: mass claim denials executed at a speed and scale that no human workforce could match, often with little meaningful review of individual circumstances.
The trend has accelerated sharply. As Digital Trends reported, insurers are deploying AI not merely as a support tool but as a primary decision-maker, one that can process and reject claims in seconds. The publication detailed how companies including UnitedHealth Group and Cigna have faced lawsuits and regulatory scrutiny over their use of algorithmic systems that critics say are designed to maximize denials rather than accurately assess medical necessity. The numbers are staggering. UnitedHealth’s AI model, known internally as nH Predict, was alleged in a federal lawsuit to have a 90% error rate — meaning that when patients appealed denials, the original decision was overturned nine times out of ten. Yet the system kept running. And the denials kept coming.
This isn’t a story about technology failing. It’s a story about technology working exactly as intended — for the companies deploying it.
The mechanics are straightforward enough. Insurers feed historical claims data into machine learning models. These models identify patterns: which diagnoses tend to correlate with expensive treatments, which patients are statistically less likely to appeal, which providers submit claims that can be flagged for review. The AI then applies these patterns to incoming claims, generating recommendations — or, in many cases, outright decisions — about whether to approve or deny coverage. The speed is extraordinary. Cigna’s system, called PXDX, was reported by ProPublica to have enabled doctors working for the insurer to reject claims at an average rate of 1.2 seconds per case. That’s not a review. That’s a rubber stamp on an algorithm’s output.
The human element hasn’t disappeared entirely, but it has been reduced to a formality in many instances. Physicians technically sign off on denials, but the volume makes genuine case-by-case analysis impossible. When a doctor is processing hundreds of denials per hour, the AI is the decision-maker. The doctor is the signature.
State regulators are starting to push back, though unevenly. Colorado enacted a law in 2021 requiring insurers to test their AI systems for bias and discrimination. California’s Department of Insurance issued a bulletin in 2024 warning companies that using AI to deny claims without individualized review violates existing state law. But enforcement remains patchy. Most states lack the technical expertise to audit these systems, and insurers have been reluctant to disclose the inner workings of their models, citing proprietary trade secrets.
The federal response has been similarly fragmented. The Centers for Medicare and Medicaid Services finalized a rule in early 2024 requiring Medicare Advantage plans to base coverage decisions on individual patient circumstances rather than algorithmic predictions alone. The rule was a direct response to reports — including a damning investigation by STAT News — that Medicare Advantage insurers were using AI to cut off post-acute care for elderly patients, sending them home from nursing facilities before they were medically ready. But Medicare Advantage represents only a slice of the market. Employer-sponsored plans, individual market plans, and Medicaid managed care plans operate under different regulatory frameworks, many of which have no specific provisions addressing AI-driven decision-making.
And the technology is getting more sophisticated. Newer models don’t just flag claims for denial — they predict which patients are likely to fight back. This is where things get particularly dark. If an algorithm can identify that a patient in a certain demographic, geographic area, or income bracket is statistically unlikely to appeal a denial, the insurer faces almost no downside in rejecting the claim. The patient gives up. The insurer saves money. The model learns that this type of denial is “successful” and replicates it.
It’s a feedback loop that rewards aggression.
Consumer advocates have been sounding alarms for years, but the issue gained mainstream traction after the December 2024 killing of UnitedHealthcare CEO Brian Thompson outside a Manhattan hotel. The suspect, Luigi Mangione, reportedly harbored grievances related to the insurance industry’s treatment of patients. While the act was universally condemned, the public reaction was telling — social media filled not with sympathy for the executive but with stories of denied claims, delayed treatments, and bureaucratic cruelty. The incident exposed a depth of public anger toward health insurers that many industry observers found startling.
UnitedHealth Group has denied that its AI systems operate without human oversight. In public statements, the company has emphasized that technology assists but does not replace clinical judgment. Cigna has made similar claims. But the lawsuits tell a different story. In the class-action complaint against UnitedHealth, plaintiffs alleged that the company’s AI model was deployed specifically because it produced denial rates that aligned with corporate financial targets, and that human reviewers who overrode the system faced internal pressure to conform to its recommendations.
The insurance industry’s trade group, AHIP (America’s Health Insurance Plans), has argued that AI helps reduce fraud, speed up legitimate claims, and lower administrative costs — savings that theoretically get passed on to consumers. There’s some truth to that. Fraud detection is one area where pattern-recognition algorithms have genuine utility, identifying billing anomalies and suspicious claim patterns that human reviewers might miss. But the same technology that catches fraud can also be tuned to catch — and reject — expensive but legitimate claims. The difference lies in how the model is trained, what outcomes it’s optimized for, and who’s checking its work.
Right now, almost nobody is checking its work.
The opacity of these systems is itself a regulatory challenge. When a human claims adjuster denies a claim, the reasoning can be examined, questioned, and challenged. When an AI denies a claim, the reasoning may be embedded in a neural network with millions of parameters — a black box that even its developers can’t fully explain. Patients receive denial letters that cite policy language and medical criteria, but the actual decision pathway remains hidden. You can appeal the denial. You can’t appeal the algorithm.
Some insurers have begun experimenting with generative AI as well, using large language models to draft denial letters, respond to patient inquiries, and even handle parts of the appeals process. Digital Trends noted that this expansion of AI into patient-facing communications raises additional concerns about transparency and accountability. When a patient receives a carefully worded denial letter, they may assume a medical professional wrote it after reviewing their case. Increasingly, that assumption is wrong.
The legal system is slowly catching up. Beyond the UnitedHealth and Cigna lawsuits, a wave of litigation is working through state and federal courts challenging the use of AI in coverage decisions. Plaintiffs’ attorneys have begun retaining data scientists and machine learning experts as witnesses, a development that would have been unthinkable five years ago. Courts are grappling with novel questions: Does an AI-generated denial constitute a coverage decision under ERISA? Can an insurer claim trade secret protection for a model that directly affects patient care? If an algorithm is biased against certain demographic groups, does that violate civil rights law?
These cases will take years to resolve. Patients can’t wait that long.
In the meantime, a cottage industry of denial-fighting services has emerged. Companies like Claimable and Waystar offer tools that help patients and providers contest AI-generated denials, sometimes using their own AI to identify weaknesses in the insurer’s reasoning. It’s an arms race, algorithms versus algorithms, with patients’ health caught in the middle. Some hospitals have hired dedicated denial management teams, adding administrative costs that ultimately get passed back to patients and employers through higher prices.
The fundamental tension is this: insurers have a legitimate need to control costs and prevent fraud, but they also have a fiduciary obligation — and in many cases a legal one — to provide the coverage their policyholders have paid for. AI makes it trivially easy to tip that balance toward denial. The technology doesn’t care about the patient on the other end of the claim. It optimizes for whatever objective function it’s been given. If that function prioritizes cost reduction over accurate adjudication, the results are predictable.
And they’re exactly what we’re seeing.
The political dimension is intensifying too. Several members of Congress have introduced legislation that would require insurers to disclose when AI is used in claims decisions, mandate human review of all denials, and establish federal standards for algorithmic accountability in healthcare. Whether any of these bills gain traction in a divided Congress remains uncertain. But the issue has unusual bipartisan appeal — nobody likes getting a claim denied, regardless of party affiliation.
For now, the industry continues to expand its use of AI. Investments in claims automation technology are growing at double-digit rates. Startups promising to help insurers “optimize” their denial processes attract significant venture capital. The incentive structure is clear: every denied claim that isn’t appealed drops straight to the bottom line. AI makes it possible to deny more claims, faster, with less human labor. From a pure business perspective, the math is irresistible.
But math isn’t medicine. And efficiency isn’t justice. The question facing regulators, courts, and ultimately the public is whether the American insurance system will allow algorithms to become the final word on who gets care and who doesn’t. So far, the answer has been a quiet, bureaucratic yes — delivered at 1.2 seconds per case, with no one watching.
That’s starting to change. Whether it changes fast enough is another matter entirely.
WebProNews is an iEntry Publication