The AI insurance adjuster deciding whether your claim gets paid may spend less time on your case than it takes to read this sentence. A landmark 2023 investigation by ProPublica and The Capitol Forum found that Cigna medical directors used an algorithm called PXDX to deny more than 300,000 payment requests over two months, spending an average of 1.2 seconds on each one, with one former Cigna doctor summing up the process to reporters as “we literally click and submit.” Three years later, the technology behind that speed has only become more central to how American insurers decide what gets covered.
What an AI Insurance Adjuster Actually Does Before You Ever See a Denial
The mechanism ProPublica documented at Cigna is representative of how algorithmic claims review works across the industry. Software scans an incoming claim, flags a mismatch between the diagnosis code and the tests or procedures Cigna’s own list considers acceptable for that condition, and routes the flagged claim to a medical director for what the company calls review.
According to the former employees ProPublica interviewed, that review typically involved no access to the patient’s actual medical file, no consideration of the specific clinical circumstances, and no individualised medical judgment at all, just an electronic signature applied in batches. Internal Cigna documents estimated only 5 percent of patients would appeal a PXDX-driven denial, a number that made the system’s economics work regardless of how many of those denials were actually correct.
UnitedHealth’s Version Reached Further Into Patients’ Lives
A separate and, according to litigation, more consequential system operates inside UnitedHealth’s Medicare Advantage business. The insurer’s subsidiary naviHealth built an algorithm called nH Predict to estimate how many days of skilled nursing or rehabilitation care a patient would need after a hospital stay, using historical data from similar patients rather than an individualised clinical assessment of the patient in front of a doctor.
A Senate Permanent Subcommittee on Investigations report found that UnitedHealth’s denial rate for post-acute care claims rose from 1.4 percent in 2019 to 12.6 percent in 2022, a ninefold increase that coincided directly with naviHealth’s expanded role in managing those claims.
The plaintiffs suing UnitedHealth allege the algorithm’s predictions were used to cut off coverage at a predetermined point regardless of a patient’s actual recovery, sometimes forcing families to pay tens of thousands of dollars out of pocket to continue care their own doctors still recommended. A federal magistrate judge ordered UnitedHealth in March 2026 to produce internal documents on nH Predict, including records related to its AI review board and the compensation structure for care coordinators, after finding that a 2024 Senate investigation’s denial-rate data could serve as evidence supporting the plaintiffs’ claims.
Optum, the UnitedHealth subsidiary that now runs the tool, disputes the characterisation directly, telling reporters that medical necessity determinations are made by qualified physicians following federal guidance, not AI, and that naviHealth functions as a care-support tool rather than a coverage decision-maker.
The Gap Between What Companies Say and What Gets Reversed on Appeal
Both companies maintain that a licensed physician makes every final coverage decision, and in a narrow technical sense that is accurate: a human signature is required somewhere in the process at both companies. What the litigation and reporting on both systems has surfaced is a gap between that formal requirement and what actually happens when a medical director is expected to clear tens of thousands of cases a month.
Reporting on the nH Predict litigation found that appealed denials tied to the algorithm were overturned roughly 90 percent of the time, a reversal rate that is difficult to reconcile with a system working as intended, since it implies the great majority of contested denials should never have happened in the first place.
That reversal rate matters enormously because so few people ever generate one. Fewer than 0.2 percent of denied claims are formally appealed industry-wide, meaning the 90 percent reversal figure describes only the small, self-selected group of patients with the health literacy, time, and stamina to fight back. The much larger group who simply pay the bill or go without the care never finds out whether their denial would have been overturned too.
California’s Answer: Keep a Physician in the Loop, Legally
California has moved further than most states to address this gap directly. Senate Bill 1120, known as the Physicians Make Decisions Act, took effect January 1, 2025, and requires that any California health plan or disability insurer using AI or algorithmic tools in utilization review base coverage decisions on an individual patient’s actual medical history and clinical circumstances, not solely on a group dataset, and prohibits AI from displacing a licensed physician’s judgment in the process.
The law does not ban AI from the claims process. It requires that AI-assisted determinations be based on individualised clinical criteria rather than population-level statistics, and that the final coverage decision remain attributable to a licensed physician who can be held professionally and legally accountable for it.
State Senator Josh Becker, the bill’s author, has argued that an algorithm cannot grasp a patient’s full medical history the way a treating physician can, and that treating it as though it could risks serious, sometimes life-threatening errors. Whether that legal requirement changes what actually happens inside an insurer’s review software, versus simply changing what the paperwork says happened, is precisely the question the pending Cigna and UnitedHealth litigation is now trying to answer through discovery.
Why This Connects to a Much Broader Pattern in AI Decision-Making
The specific failure mode here, a system that technically satisfies a human-review requirement while functionally removing individualised judgment from the process, recurs across nearly every domain where AI now makes or shapes consequential decisions about people. LiveAIWire’s coverage of the AI doctor dilemma in diagnostic medicine found the same underlying technology producing dramatically different real-world outcomes depending entirely on how much genuine human oversight surrounds it, with some deployments improving diagnostic accuracy substantially and others missing critical conditions at alarming rates.
An AI insurance adjuster sits inside that same spectrum: the technology’s usefulness is not fixed, and whether it functions as a genuine second opinion or as a rubber stamp depends on choices insurers make about time, incentives, and accountability that have nothing to do with the underlying algorithm’s technical sophistication.
A closely related pattern shows up in how AI evaluates financial trustworthiness more broadly. LiveAIWire’s reporting on AI credit scoring’s persistent discrimination gap found that algorithmic lending systems marketed as fairer than human loan officers still produced measurably unequal outcomes for minority borrowers, detectable only through the kind of rigorous, independent research that an individual applicant, or an individual patient contesting a denied claim, has no practical way to conduct on their own behalf.
Claims Denial and Premium Pricing Are the Same Underlying Story, Told Twice
An AI insurance adjuster deciding whether to pay a claim and an AI underwriting model deciding what premium to charge in the first place are, functionally, two ends of the same pipeline. LiveAIWire’s coverage of AI-driven insurance premium pricing found that algorithmic precision at the pricing stage systematically prices the highest-risk customers out of coverage entirely, while the claims-denial systems examined here determine what happens to the customers who make it through underwriting and then actually need the coverage they paid for.
Both stages are optimising the same objective, minimising the insurer’s payout relative to premium collected, and both produce the identical distributional pattern: customers who most need the product are the ones an AI system is most likely to price out or deny. LiveAIWire’s broader reporting on AI’s expanding role across personal finance found this same pattern recurring in credit, budgeting, and now claims: genuine efficiency for the mechanical parts of a financial decision, and a persistent gap in cases that depend on an individual’s specific, messy circumstances rather than a population average.
Humana Shows This Is an Industry Pattern, Not One Company’s Mistake
Cigna and UnitedHealth are not isolated cases. Humana faces a similar lawsuit in Kentucky over its own use of the nH Predict algorithm, with plaintiffs making the same core allegation: that a predictive model built on aggregate historical data was used to cut off post-acute care at a predetermined point regardless of a specific patient’s actual recovery trajectory. The fact that two of the largest Medicare Advantage insurers in the country ended up relying on the same third-party algorithm, built by the same naviHealth subsidiary before UnitedHealth acquired it, suggests the underlying business logic is an industry standard practice rather than a single company’s aggressive outlier decision.
That concentration matters for anyone trying to assess how widespread this problem actually is. A predictive tool built once and licensed or deployed across multiple major insurers means a single flawed assumption in the underlying model, about how quickly patients with a given diagnosis typically recover, for instance, can propagate into denial decisions affecting millions of people across several different companies simultaneously, without any of them needing to have independently developed a bad idea. It also means regulatory scrutiny of one insurer’s practices, such as the discovery UnitedHealth now faces, may end up revealing information relevant to how the same underlying technology functions at its competitors.
What This Means for You
If your claim is denied, the evidence gathered here supports a specific, practical response rather than simply accepting the letter at face value. Ask explicitly whether the denial involved an algorithmic or AI-assisted review, and if you are a California resident, cite SB 1120’s requirement that the decision reflect your individual medical circumstances rather than a population dataset.
Appeal every denial you believe is wrong, since the roughly 90 percent reversal rate documented in the nH Predict litigation suggests the base rate of incorrect algorithmic denials may be far higher than insurers’ own numbers imply, and the overwhelming majority of people who could successfully appeal never do. Request your full claim file and any documentation of how the denial decision was reached, since ProPublica’s reporting found that even asking pointed, specific questions about whether a human genuinely reviewed a case can surface information insurers do not volunteer.
The honest verdict on AI insurance adjusters mirrors nearly every other domain examined in this piece: the technology is not inherently the problem, and used to flag genuinely miscoded claims or surface routine approvals faster, it can reduce friction for patients and providers alike.
The problem is a business model that rewards speed and denial volume over individualised review, deployed by companies that control both the algorithm and the appeals process a patient must use to challenge its output. Nothing about that arrangement requires you to accept a denial as final before you have actually tested it.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.