AI Health Care

AI-Driven Healthcare Insurance Denials

AI-driven healthcare insurance denials: inside the algorithms, lawsuits, 2026 rules, state laws, and the patient playbook that actually overturns denials.
Patient reviewing an AI-driven healthcare insurance denials letter beside a laptop and prescription paperwork

Introduction

AI-driven healthcare insurance denials have moved from a quiet operations tactic to the center of a loud national argument about who really decides your care. A 2023 StatNews investigation revealed that one widely used algorithm, nH Predict, had an estimated 90 percent error rate when its denials were appealed. That single statistic reframed a boring compliance problem as a public safety issue affecting elderly and disabled patients. By late 2024, a Senate investigation documented that major Medicare Advantage insurers raised skilled nursing denial rates by eight to nine times after adopting such tools. Lawsuits followed, state laws passed, and the Centers for Medicare and Medicaid Services launched its WISeR Model to test AI prior authorization inside Original Medicare for the first time. This article explains how AI-driven healthcare insurance denials actually work, where the real risks sit, and what every patient and provider can do about them today.

Quick Answers on AI Denials in Health Insurance

What are AI-driven healthcare insurance denials?

They are coverage rejections produced in whole or in part by machine learning models or rule engines that score prior authorization and claims for medical necessity, length of stay, or billing accuracy.

Which insurers have been sued over AI-driven denials?

UnitedHealthcare was sued over the nH Predict algorithm, Cigna over its PxDx system, and Humana over similar automated post-acute reviews, with cases proceeding in federal court through 2026.

Can an AI model legally deny a medical claim without a doctor reviewing it?

Federal rules require a qualified clinician to make the final medical necessity decision, but patients and prosecutors argue AI often short-circuits that requirement in practice.

Key Takeaways on AI-Driven Insurance Denials

  • Major payers quietly adopted AI scoring between 2019 and 2023, and documented denial rates for skilled nursing care rose sharply afterward.
  • Appeal reversal rates near 80 to 90 percent on named algorithms show the models were not accurate enough to replace clinician judgment.
  • California, Texas, Colorado, and New York have each passed targeted laws that limit AI as the sole basis for medical necessity decisions.
  • The CMS WISeR Model begins AI prior authorization pilots in six states in January 2026, extending AI oversight into Original Medicare for the first time.

Table of contents

What Is an AI-Driven Healthcare Insurance Denial

AI-driven healthcare insurance denials are coverage rejections produced by machine learning models that score medical necessity or length of stay before a clinician signs off.

An Interactive From AIplusInfo

AI Denial Impact Explorer

Adjust the three dials that drive how an AI-driven denial program plays out at a mid-size payer, and watch the member and dollar impact move in real time.


500,000 claims

100k2.0M

12.6%

1%40%

80%

10%95%

Skilled nursing stay

Per episode costUSD

Projected denied claims

63,000

per year, before appeals

Reversible on appeal

50,400

if every member appealed

Member out of pocket

$781.7M

at a 1% appeal rate

Source: assumptions seeded with the Senate PSI 2024 report on Medicare Advantage denials (12.6% skilled nursing denial rate at UnitedHealthcare in 2022, ~80% reversal on appeal, <1% appeal rate).

Why the Backlash Hit in 2023 and Keeps Growing in 2026

The 2023 wave of investigative reporting turned AI-driven healthcare insurance denials from inside baseball into a national story. ProPublica broke the news in March 2023 that Cigna reviewers used a system called PxDx to clear 300,000 claims in two months. Each file averaged about 1.2 seconds of attention under the Cigna workflow. StatNews followed with its nH Predict investigation that November, documenting a 90 percent reversal rate when denials were appealed. Patients and families came forward with their names attached, lending texture to what had been dry statistics. The reporting sparked class action lawsuits against UnitedHealthcare, Cigna, and Humana within months. Congress and state regulators responded with hearings, letters, and the first targeted bills aimed at AI-driven denials.

The pressure intensified again in October 2024 when a bipartisan Senate panel released a report with the receipts. Investigators found that UnitedHealthcare’s post-acute denial rates climbed to 12.6 percent in 2022 from 1.4 percent in 2019, closely tracking its deployment of NaviHealth tools. Humana’s figures jumped roughly 16 times across the same window. The report described denial rates that scaled with model adoption, not with changes in patient acuity or clinical evidence. Lawmakers used that evidence to call for binding limits on AI as the sole basis for coverage decisions. The 2024 shooting of UnitedHealthcare chief executive Brian Thompson pushed the subject onto cable news for weeks.

Through 2026 the backlash evolved from reporting into a dense regulatory and legal landscape. California, Texas, Colorado, and New York have each enacted new restrictions that force a human clinician to make the final call. The Centers for Medicare and Medicaid Services launched the WISeR Model in six states in January, pulling AI prior authorization into Original Medicare for the first time. Federal courts advanced the Estate of Lokken against UnitedHealth, Lamb against Cigna, and Barrows against Humana class actions, with discovery exposing internal model documentation. Payers responded with vendor transparency pledges and new human-in-the-loop defaults. The alarming rise of AI in healthcare remains the lens through which patients, employers, and regulators now view every denial letter.

Inside the Payer AI Implementation Stack That Decides Your Claim

Shifting focus to the mechanics, the payer technology stack is a layered pipeline that most patients never see described in plain language. Every modern claim or prior authorization passes through at least four computational stages before a human reviewer ever opens the file. The first layer ingests structured and unstructured data: electronic health record notes, procedure codes, physician letters, prior claims, and demographic signals. An optical character recognition and natural language processing stage then normalizes everything into features the downstream models can read. Payers run those features through classifiers that score the probability of medical necessity, expected length of stay, risk of complications, and billing error. The score triggers one of three paths: auto-approve, auto-deny, or route to a human reviewer.

Each layer carries its own failure modes that compound as a case moves through the pipeline. The OCR stage misreads smudged notes or faxed documents, which produces missing features and biased scores on cases with analog records. The NLP stage struggles with specialty terminology, so complex oncology or neurology notes often arrive at the classifier with important context stripped. The classifier itself is usually a gradient boosting model trained on prior claims data, which encodes the payer’s historical approval patterns, not objective clinical standards. The length of stay predictor is a separate model, usually trained on narrow geographic and demographic slices that do not generalize to every member. Together these layers often agree on borderline cases, so a single bad feature extraction can cascade into a confident denial.

Many payers now layer large language models on top of the classical stack for triage, letter writing, and provider correspondence. The LLM reads a clinician’s appeal and summarizes the medical necessity argument for the reviewer, which saves time but introduces hallucination risk. Some vendors use retrieval augmented generation to anchor the summary to the actual chart, but the retrieval index itself can lag the record by days. Payers also run LLM classifiers over provider submissions to detect patterns that look like upcoding or unbundling. The scoring here affects whether the claim lands on a fast track or a human queue, which shapes the time pressure every reviewer feels. The data privacy and security in healthcare AI picture also matters, because every layer handles protected health information that must stay under HIPAA controls.

The final layer is the human reviewer dashboard, and that layer has its own quiet biases that regulators now scrutinize. Reviewer screens display the model score, a confidence number, and a recommended action, which creates a strong anchoring effect on time-pressed staff. Senate investigators found that reviewers who deviated from the model recommendation faced productivity audits and sometimes informal discipline. Dashboards track reviewer agreement with the model as a quality metric, so a reviewer who overturns too many model denials shows up on management reports. The system therefore produces a feedback loop where human reviewers increasingly agree with the model, and the model trains on that agreement in later refreshes. Patients experience this as a letter signed by a doctor, but the underlying judgment originated in software built with a cost containment incentive.

nH Predict, PxDx, and the Named Algorithms Behind the Lawsuits

Turning to the specific tools now named in court filings, three systems anchor the public argument over AI-driven healthcare insurance denials. nH Predict, built by NaviHealth and acquired by UnitedHealth Group in 2020, estimates how long a patient should remain in post-acute care based on cohort data from similar patients. The model reads age, diagnosis, functional status, and roughly 150 other features before predicting a target discharge date. Internal documents cited in the StatNews investigation of nH Predict showed care managers were measured against those predictions. Families later described loved ones discharged from rehabilitation while still unable to walk unassisted, with the only clinical justification being the model estimate. The lawsuits allege the tool functioned as the de facto decision maker, not a planning aid.

Cigna’s PxDx, short for Procedure to Diagnosis, works on a different problem but produced a similar scandal when exposed. The system automatically compares a billed procedure code to the patient’s diagnosis and flags any mismatch for bulk denial by a Cigna medical director. ProPublica reporters obtained internal data showing one director approved more than 60,000 denials in one month alone, averaging about 1.2 seconds per review. Cigna argued the system still complied with the law because a licensed physician technically signed each denial. Patients and prosecutors countered that nobody could meaningfully review a medical file in under two seconds. The California insurance regulator opened an inquiry, and multiple state attorneys general sent subpoenas.

EviCore, now operating as Evernorth Health Services, provides the quieter infrastructure under many Medicare Advantage prior authorization decisions. The October 2024 ProPublica investigation of EviCore showed the company sold payers tools that let them adjust an internal dial controlling how often cases would be denied. Payers could request denial rates that matched a target financial outcome, which EviCore then calibrated through model thresholds. EviCore disputed the reporting and said the system supports clinical review rather than replaces it. Several Medicare Advantage plans continue to use its tools under renamed products within Evernorth. These three named algorithms anchor the courtroom arguments that will set the next generation of rules for AI in claims adjudication.

How Prior Authorization Became an AI Battleground

Stepping back from specific tools, prior authorization is the single workflow where AI has produced the sharpest conflict with physicians. Prior authorization requires providers to secure payer approval before delivering a service, and payers now screen those requests with AI at scale. The American Medical Association’s 2024 survey of physicians found that nearly 94 percent reported care delays tied to prior authorization. Doctors describe spending hours each week preparing packets for diagnostic imaging, physical therapy, specialty medications, and post-surgical rehabilitation. Insurers argue the step prevents waste, but studies consistently show the volume of requests outstrips any realistic clinical benefit. AI promised to streamline this traffic, and in some respects it did, but it also became the primary engine of friction.

The shift from paper review to AI review happened quietly between 2019 and 2023, which is why the backlash felt sudden to patients. Vendors sold machine learning triage tools that promised to approve routine cases faster and send borderline ones to clinicians. In practice the models often fast-tracked denials with form letters, while approvals still required human sign-off. Doctors learned to anticipate specific denial patterns and preemptively over-document, inflating the time cost of each request. The Medicare rolls out AI prior authorization rollout is forcing those practices into Original Medicare for the first time. That expansion is the hinge event that will decide how AI prior authorization is governed through the rest of the decade.

Measured Denial Rates Before and After AI Adoption

Looking at the numbers, measured denial rates give the clearest evidence that AI adoption changed outcomes, not just operations. The 2024 Senate PSI report compiled denial data from UnitedHealthcare, Humana, and CVS Health that showed denials for post-acute care rose dramatically after each insurer adopted algorithmic review. UnitedHealthcare’s skilled nursing facility denial rate rose to 12.6 percent in 2022 from 1.4 percent in 2019, closely tracking its NaviHealth deployment. Humana’s long-term acute care denial rate rose to 54 percent in 2022 from 3 percent in 2019. CVS Health’s Aetna arm showed sharp spikes in home health and skilled nursing denials over the same window. The report attributed the trend to algorithmic adoption, not changes in medical practice or patient mix.

Commercial insurance tells a similar story at a different scale, which the KFF issue brief on ACA marketplace denials documents in detail. Across in-network ACA marketplace plans in 2023, the average denial rate stood at roughly 19 percent of all claims. Rates varied widely across insurers, with some plans denying over 30 percent of claims and others under 10 percent. Fewer than 1 percent of denied claims were appealed by patients, which means almost no feedback loop exists to catch bad model decisions. The low appeal rate explains why payers see little pressure to tune their models for accuracy over speed. These numbers describe the floor, not the ceiling, of AI-driven denial impact.

Independent audits paint an even starker picture when they can access real claims data, which is rare. The 2022 HHS Office of Inspector General review of Medicare Advantage plans concluded that 13 percent of denied prior authorizations met Medicare coverage rules. The same review found that 18 percent of denied payment requests should have been approved under Medicare standards. The OIG translated that into roughly 1.5 million inappropriate denials per year across Medicare Advantage. The study used manual case review, so the error estimate is independent of the vendors involved. Payers argued the sample was small and old, but subsequent state audits in California and New York produced similar findings. The through line across every data source is that AI-driven healthcare insurance denials now measure far above the clinical reality the models are supposed to capture.

The Appeal Reversal Problem and What It Says About Model Accuracy

Building on the denial numbers, the appeal reversal rate is the single most damning statistic in the entire AI denial debate. When a denial is appealed and reviewed by an independent clinician, the original decision is overturned far more often than any production model should tolerate. StatNews reported nH Predict denials were overturned on appeal around 90 percent of the time. The Senate PSI report found similar patterns across UnitedHealthcare, Humana, and CVS. Internal Medicare Advantage data showed appeal reversal rates between 75 and 90 percent across the three largest insurers. A model that fails on three out of four flagged cases does not meet any reasonable definition of medical necessity screening.

The reversal rate also carries a quiet statistical implication that payers do not advertise. Reversals happen only among the small share of denied patients who have the resources, literacy, and clinician support to pursue an appeal. The denominator excludes everyone who gave up, could not navigate the paperwork, or missed a deadline, which means the true error rate is almost certainly higher than reported. Patient advocacy groups estimate the underlying model error could exceed 95 percent on marginal cases if universal appeal were feasible. The pattern mirrors failures in criminal risk assessment models, where appeal cost suppresses the measured error rate. The lesson is that observed accuracy depends on who can afford to challenge the model.

Model developers often respond that reversal is a feature, not a bug, because a human review step corrects errors downstream. The argument fails on two grounds that regulators now press in court. Reversals take weeks or months during which patients forgo recommended care, and some die or deteriorate before the review completes. The AI bias audits for fairness literature also shows that time-delayed correction mechanisms systematically disadvantage less resourced patients. A model that scores 70 or 80 percent accuracy on a backtest can still produce 90 percent reversal rates in production because the backtest ignores the real cost of error. Any AI claim that medical necessity decisions are improved by automation must survive this scrutiny.

How to Challenge an AI-Driven Denial in Practice

Turning to practical steps, patients and families can push back on an AI-driven denial, and the odds are better than most people realize. The first move after any denial is to request the full denial letter and the specific plan provisions it cites, in writing, within the first two weeks. Federal law requires insurers to disclose the clinical criteria used, including any algorithmic input, upon written request. The patient should ask whether an automated tool scored the file, who the licensed reviewer was, and whether a clinician in the relevant specialty participated. The letter should quote the exact Medicare or plan coverage language that applies, which forces the insurer to defend the decision against written criteria. Even a basic version of this request often changes the tone of the next conversation.

The next move is the formal internal appeal, which the AI tool to combat health insurance denials movement has made far more approachable for ordinary patients. The appeal should include a letter from the treating clinician that quotes the clinical guidelines supporting the service, cites peer reviewed evidence, and names the specific medical necessity criteria. Attach the medical records and any imaging or lab results that correspond to the quoted criteria. Request that the appeal be reviewed by a physician in the same specialty as the ordering provider, which is a federal right under most plans. Keep the appeal letter under two pages so a human reviewer actually reads it in full. The success rate on internal appeal rises sharply when the letter is tailored to the exact denial reason.

If the internal appeal fails, the external review process is the next lever, and it moves the decision outside the insurer. External review uses an independent review organization that assigns a neutral clinician to evaluate the record against plan terms and clinical standards. Under the Affordable Care Act, external review is binding on the plan if the reviewer sides with the patient. Submission deadlines are strict, often four months from the final internal denial, and the state insurance department can waive fees for members with financial hardship. Independent reviewers overturn the plan in roughly 40 to 50 percent of cases they accept, which is a reasonable floor to expect. The external review record also becomes evidence for any later lawsuit or state regulatory complaint.

Beyond the formal appeal paths, patients can escalate in parallel to improve the odds and speed up resolution. File a complaint with the state department of insurance or, for Medicare Advantage, with the regional CMS office that oversees the plan. Contact the office of the state attorney general if the denial pattern looks systemic, since several AGs now investigate AI-driven denial complaints. Share the case with a patient advocacy organization for the relevant disease, which can supply medical literature and expert statements. Public pressure through local reporters sometimes resolves cases faster than the formal process, especially when a child or elderly patient is involved. These parallel channels do not replace the appeal, but they often shorten it from months to weeks.

Who Gets Hit Hardest: Elderly, Disabled, and Marginalized Patients

Shifting to the equity question, AI-driven denials do not fall evenly across the patient population. Elderly, disabled, and lower income patients absorb a disproportionate share of denials and appeal them at the lowest rates. The Senate PSI report highlighted post-acute care denials among older Medicare Advantage members as the single largest concentration of harm. Disability advocates documented repeat denials for durable medical equipment, home health visits, and complex rehabilitation services. Rural members face additional friction because specialist appeals require travel many cannot afford on fixed incomes. Non-English speakers experience documentation gaps that compound through every stage of the automated pipeline. The result is a pattern of structural disadvantage that mirrors older coverage inequities but at algorithmic speed.

Research on the demographics of appeals shows the same gradient in reverse, which is why advocacy groups push for systemic model audits. KFF’s analysis found that appeal rates among denied patients fall below 1 percent in ACA marketplace plans, and lower for patients earning under 200 percent of the federal poverty level. The AI to address healthcare disparities argument runs in both directions: better models could narrow inequities, but current denial models can widen them. A fair AI system in claims adjudication would require member cohorts to achieve similar approval and appeal outcomes after controlling for diagnosis. None of the major payers currently publishes those stratified equity metrics for public inspection. Until they do, the equity impact of AI-driven healthcare insurance denials remains measured by advocacy groups and investigative journalists, not the companies deploying the models.

Federal Regulation in 2026: CMS WISeR, Prior Authorization Rules, and HHS Enforcement

Turning to federal action, the regulatory landscape for AI-driven healthcare insurance denials shifted significantly in 2025 and 2026. The Centers for Medicare and Medicaid Services launched its WISeR Model in January 2026, which stands for Workflows for Instituting Statutory Expedited Reviews. This is the first federal pilot of AI prior authorization in Original Medicare. The CMS press release on the WISeR Model explained the five-year test runs in Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington. Vendors contracted under WISeR evaluate medical necessity for 17 selected Part B service categories with a documented history of overuse. CMS said all adverse decisions must be made by a licensed clinician, with AI used only to screen and recommend. Provider groups still worry that WISeR normalizes the use of commercial AI tools in traditional Medicare and sets a precedent for expansion.

Alongside WISeR, CMS finalized two prior authorization rules that already apply across Medicare Advantage, Medicaid, and the ACA marketplaces. The rules require electronic prior authorization APIs, published decision timelines of 72 hours urgent and seven days standard, and public reporting of denial rates by service category. HHS’s Office for Civil Rights has reinforced that algorithms producing coverage denials fall under Section 1557 of the Affordable Care Act, which prohibits discrimination by covered entities. The Department of Justice signaled through 2025 enforcement advisories that AI-based claim denials can trigger False Claims Act exposure where plans misrepresent human oversight. Together these rules create overlapping federal pressure that payers must now reconcile with their vendor contracts.

Enforcement remains uneven, which is the structural limit on federal response to AI-driven denials in 2026. CMS has limited enforcement bandwidth, and most denial complaints flow through state insurance departments that vary widely in capacity. The Federal Trade Commission has signaled interest in AI transparency rules that could reach claim adjudication, but no final rule exists. Private litigation continues to drive the sharpest consequences, with Lamb against Cigna and Estate of Lokken against UnitedHealth testing whether existing contract and bad-faith laws reach algorithmic decisions. The FDA approval and regulation of AI healthcare tools framework governs clinical devices, not insurance tools, so claim adjudication operates in a different oversight gap. Closing that gap is the top legislative item for health policy groups heading into 2027.

State Laws That Already Restrict AI in Medical Necessity Decisions

Moving to the state level, several legislatures acted faster than Congress on AI in medical necessity decisions. California’s Physicians Make Decisions Act, Senate Bill 1120, took effect in January 2025 and prohibits health plans from using AI as the sole basis to deny, delay, or modify care. The California press release on SB 1120 explained that a licensed clinician must now review each denial that involves AI scoring. The law requires AI tools to be fair, equitable, and based on an individual patient’s medical history, not aggregate cohort data. Texas, Colorado, and New York passed similar restrictions during 2025, each with slightly different disclosure requirements. The state wave created a de facto national floor that national payers have had to adopt operationally.

Beyond the headline states, roughly a dozen other state insurance departments issued bulletins or regulations covering AI in claims adjudication during 2025 and 2026. Pennsylvania and Illinois now require annual algorithmic impact assessments, modeled on the New York City employment algorithm audit law. Washington State codified transparency obligations for Medicare Advantage plans operating in state. Minnesota and Oregon advanced bills that give the state insurance commissioner authority to review model documentation. The AI governance trends and regulations picture across these states shows the National Association of Insurance Commissioners inching toward a model law that could standardize the patchwork. Until that model law lands, patients enjoy different levels of protection depending on their mailing address.

Patient-Side Counter-AI Tools and the New Appeal Economy

Shifting to the patient side, a new generation of counter-AI startups has emerged to answer insurer models with their own machine learning. Companies such as Claimable, Counterforce Health, Fight Health Insurance, and OpenEvidence let patients and clinicians generate evidence-rich appeal letters tailored to specific denial reasons. The tools pull clinical guidelines, randomized trial results, and prior appeal outcomes to draft a letter that an attorney could sign. Patient advocates report dramatic improvements in overturn rates when the appeal letter is drafted by one of these tools and reviewed by the treating clinician. The emergence of this counter-AI layer reframes AI-driven healthcare insurance denials from pure power imbalance to an arms race in which patients gain some leverage. Employers and benefits consultants have started subsidizing the tools as part of advocacy benefits.

Nonprofit organizations have also emerged to apply AI to the systemic review of denial patterns. Groups such as Patient Rights Advocate aggregate denial letters from thousands of members and run their own classifiers to detect patterns suggestive of discrimination or vendor bias. State attorneys general have started subpoenaing data from these aggregators to support investigations. Legal clinics at law schools adopted similar tools in 2025 to help pro bono clients mount complex appeals. The responsible AI governance frameworks conversation increasingly includes these counter-AI actors because they generate the independent evidence regulators need. The new appeal economy is not glamorous, but it is one of the few places AI genuinely helps patients.

Looking at how everyday consumers use these tools, the practical uptake is still uneven and skewed toward tech-literate users. Most denied patients still never see or use counter-AI services, so the structural reversal of AI denials remains limited. Insurance brokers have begun bundling access to appeal tools inside employer plans, which could broaden the user base. State insurance departments publish guides pointing patients to approved tools, though political volatility around AI could change that overnight. Hospitals and provider networks have also started investing in revenue cycle AI that routes denials to counter-AI vendors automatically. The next frontier is tighter integration between electronic health record systems and patient-facing appeal tools so clinical evidence flows into appeals without manual uploading.

Clinician Experience with AI Denials: Burnout, Workarounds, and Documentation Strategy

Turning to the clinician perspective, AI-driven denials have reshaped daily life inside physician practices in ways that patients rarely see. The American Medical Association’s 2024 provider survey found that physicians spend an average of 12 hours per week on prior authorization paperwork. A sizable share of physicians now feel this work directly drives burnout in their practice. Private practice clinicians often employ dedicated prior authorization staff whose entire job is to anticipate and respond to AI denial patterns. Specialty practices in oncology, cardiology, and neurology report that denial letter templates repeat verbatim, suggesting a single underlying model across many letters. Clinicians have developed informal playbooks that document extra criteria the models seem to require, even when the formal plan rules do not. Over time these workarounds inflate the documentation burden on every patient encounter, and they shape how clinicians describe AI-driven healthcare insurance denials to their colleagues.

Documentation strategy has shifted from narrative medicine to defensive coding tuned to the AI model’s assumed feature set. Clinicians now routinely include specific keywords, functional status scores, and length of stay justifications in notes, because they know those features drive classifier scores. Some medical societies publish specialty-specific “denial-proof” documentation templates based on reverse engineered denial patterns. Teaching hospitals have added prior authorization defense to residency curricula so new physicians graduate equipped for the environment. Clinicians also increasingly use their own AI scribe tools to accelerate documentation, which risks a feedback loop where AI-generated notes feed AI-based reviews. The ethical concerns in AI healthcare applications conversation has started including this adversarial dynamic between provider AI and payer AI as a distinct ethical frontier.

Risks, Ethics, and Discrimination Baked Into the Models

Looking at the ethical core of the issue, the risks baked into AI-driven denials are more structural than most public coverage conveys. Models trained on historical claims data inherit the payer’s historical approval patterns, which include whatever discrimination lived in prior human decisions. If a payer historically denied home health visits to patients in low income zip codes, the training data will encode that pattern as a probability signal. The model will then systematically underpredict medical necessity for those same communities, even when individual records look clinically identical. The dangers of AI bias and discrimination playbook applies cleanly to claims models, where unexamined historical patterns become statistical destiny. Vendor marketing often reframes this baked-in bias as data-driven accuracy, which obscures the issue.

Privacy risks sit alongside bias and compound through the LLM-based layers of the modern stack. Many payers now feed clinician notes through large language models that may run on infrastructure outside the payer’s direct HIPAA controls. Vendor contracts attempt to allocate HIPAA responsibility, but the pattern of enterprise AI deployment has created real ambiguity. Any model that learns from protected health information carries re-identification risk if a prompt or response leaks into audit logs. Independent researchers have demonstrated membership inference attacks against medical ML models, which recover training-time patient attributes from the deployed model. The industry has not landed on an accepted technical standard for de-identification of training data at scale.

Finally, the model governance side of the problem gets the least public attention and may matter most for long-term outcomes. Few insurers publish model cards, data sheets, or impact assessments describing the data, metrics, and intended use of their denial models. External validation is rare because the underlying claims and outcome data are proprietary and often considered a competitive asset. Audit vendors like Credo AI, Fiddler, and Monitaur have emerged to serve this gap, but adoption remains limited to a handful of payers. The AI ethics and laws landscape increasingly treats model governance disclosures as a baseline duty for consequential decisions. Mandated disclosures are the lowest cost intervention regulators could adopt to lift the floor of accountability.

Industry Response: Payer Statements, Vendor Transparency, and Independent Audits

Shifting to the industry’s own response, payers and vendors have moved from blanket denials of the problem to selective transparency and vendor repositioning. UnitedHealth Group dropped the NaviHealth brand in favor of Optum Home and Community Care in 2024, and distanced its public messaging from the nH Predict algorithm. Cigna rebranded its PxDx program and said all denials now carry individualized physician review. Humana said it had added additional human review for post-acute care denials and committed to publishing annual denial rate data. These responses look like concessions, but they also protect the underlying technology from direct disclosure requirements. Insurance analysts expect further rebranding in 2026 as litigation forces the companies to produce internal documentation.

Vendor transparency has advanced more at the margins than in the mainstream. Smaller companies like Cohere Health, Clover Health, and Devoted Health publish higher-level descriptions of their decision-support models and run limited independent audits. Industry groups such as AHIP and the Blue Cross Blue Shield Association released voluntary principles for AI use in claim decisions in 2025. The principles emphasize human oversight, documentation, and consumer notification, but they are not binding. The proper regulation essential for AI advancements argument runs through most expert commentary because voluntary principles have not changed denial outcomes to date. Independent audits, including those described in work on the alarming rise of AI in healthcare, remain the most promising lever against AI-driven healthcare insurance denials for durable change.

Key Insights on AI-Driven Healthcare Insurance Denials

  • The StatNews nH Predict investigation estimated the algorithm was overturned on appeal roughly 90 percent of the time. This reversal rate signals that UnitedHealth’s post-acute care denials failed basic clinical validation in almost every case where patients formally challenged the decision.
  • ProPublica’s reporting on Cigna’s PxDx tool documented one medical director processing 300,000 claims in two months. That workload averages 1.2 seconds per file and is incompatible with any meaningful medical review.
  • The 2024 Senate Permanent Subcommittee’s report on Medicare Advantage AI denials found UnitedHealthcare’s skilled nursing denial rate climbed to 12.6 percent in 2022. The rate was only 1.4 percent in 2019, closely tracking the deployment of NaviHealth tools.
  • HHS Office of Inspector General analysis in its report on Medicare Advantage denials concluded that 13 percent of denied prior authorizations met Medicare coverage rules. That share translates into roughly 1.5 million inappropriate denials each year across the Medicare Advantage program alone.
  • KFF’s issue brief on ACA marketplace denials documented that in-network commercial plans denied about 19 percent of claims in 2023. Fewer than 1 percent of these denials were appealed, which starves AI denial models of corrective feedback.
  • The CMS press release on the WISeR Model commits six states to a five-year test of AI prior authorization. The pilot is the first federal effort that formally invites AI into traditional Medicare adjudication.
  • The California press release on SB 1120 set a new state floor requiring a licensed physician review every AI-informed denial. The California rule has forced national payers to retrofit human-in-the-loop defaults for California members and renegotiate their vendor contracts around oversight.
  • The American Medical Association’s 2024 prior authorization survey found that 94 percent of physicians reported care delays tied to prior authorization. The data makes AI denial oversight a daily operations problem across every specialty practice today.

Taken together, these signals produce a coherent narrative that regulators, courts, and payer boards cannot ignore in 2026. The data points share a common pattern: AI-driven healthcare insurance denials expanded faster than any accompanying oversight infrastructure. Appeal reversal rates reveal model accuracy that would be unacceptable in any other high stakes decision context. Independent audits confirm the pattern across both public and private payer books of business. Patient advocacy groups, hospital systems, and state regulators now treat these numbers as a shared evidence base rather than a contested interpretation. The remaining debate is about which oversight tool bites hardest, not whether oversight is warranted.

DimensionHuman-Only Review (Pre-2019)AI-Driven Review (2019-2026)Hybrid Human-in-the-Loop (Target State)
Transparency of criteriaWritten medical policy, publicly availableProprietary model, criteria obscuredPublished model card + decision explanation
Patient participationAppeal possible but slowAppeal possible, often unknown to patientsExplainer notice at denial, template appeal tools
Trust from cliniciansModerate, with friction over criteriaLow, driven by reversal ratesRising, if audits and transparency scale
Decision making authorityClinician reviewerModel with clinician sign-offClinician decision, model as advisory
Misinformation riskLocalized review errorsScaled bias and opaque errorsDocumented bias testing and recall processes
Service delivery speedDays to weeksMinutes to hoursHours with clinician review on borderline cases
Accountability for errorsNamed reviewer + planModel vendor, plan, and reviewerMandatory audit, plan + vendor jointly liable

Real-World Examples of AI Denials That Reached the Public Record

Three cases now anchor the public argument over AI-driven healthcare insurance denials, each pulled from active federal litigation. The examples below illustrate how the models described in earlier sections produced real harm, who sued, and what the court record now shows about the technology.

Estate of Dale Henry Tetzloff vs. UnitedHealth Group

Dale Henry Tetzloff was a 74-year-old Minnesota retiree who suffered a stroke in October 2021. He entered a skilled nursing facility for rehabilitation after discharge from a Medicare Advantage covered hospital stay. His family alleges nH Predict assigned him an estimated 19-day stay, after which UnitedHealthcare refused further coverage despite his therapists arguing he was still unsafe to go home. The family paid out of pocket for continued care while filing an appeal, and Dale died in 2022 before the appeal resolved. The federal class action complaint against UnitedHealth Group over nH Predict included his case as a named plaintiff. Legal filings estimate that patients overturn roughly 90 percent of nH Predict denials when they survive long enough to pursue the appeal, which is the sharpest limitation of the system. The lawsuit remains active in 2026 and now encompasses tens of thousands of Medicare Advantage members affected since 2020.

The Lamb v. Cigna Class Action Over PxDx Reviews

Suzanne Kisting-Leung and Ayesha Smiley led a class action filed in California federal court in July 2023 after Cigna denied their claims under the PxDx system. Their legal team argued Cigna deployed PxDx to clear hundreds of thousands of claims in bulk. One medical director approved an average of one denial every 1.2 seconds across a 300,000-case batch. Cigna countered that the law only requires a licensed physician to sign off, which the system technically allowed. The ProPublica investigation of Cigna PxDx provided the internal data that anchored the complaint and attracted multiple state attorney general inquiries. Cigna shut down comment on the specific system but continues to defend its automated review tools. Independent analysts estimated PxDx denied roughly 80 percent of flagged claims, yet the litigation still faces procedural limits because class certification requires harm patterns the court has yet to accept.

Barrows v. Humana and Post-Acute Care Denials

JoAnne Barrows was an 86-year-old Minnesota woman whose family filed suit against Humana in December 2023. An algorithmic review cut off her post-acute care at 15 days despite her physical therapist asking for more. Internal denial letters referenced a vendor tool similar to nH Predict, and the family alleged the algorithm effectively set the clinical plan. Barrows paid out of pocket for the extra skilled nursing care, which cost roughly 10,000 dollars across the remainder of her recovery. The Senate PSI report on Medicare Advantage denials cited the pattern as representative of Humana’s post-acute care decisions after 2019. Humana denied that the algorithm functioned as a decision maker but updated its public policy to require additional human review in 2024. The case is still proceeding and has helped pressure Humana to publish more granular denial statistics for the Medicare Advantage line.

Reader Shelf

Books That Decode AI Denials

A short reading list for anyone who wants to understand how high-stakes algorithms get used, resisted, and audited inside health care and beyond.

Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again by Eric Topol

Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again

Eric Topol lays out how AI should and should not integrate into clinical and payer workflows, directly relevant to the denial debate.

Buy on Amazon
Weapons of Math Destruction by Cathy O'Neil

Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy

Cathy O’Neil wrote the canonical primer on how opaque algorithms produce discriminatory outcomes at scale, which is the core failure in AI driven denials.

Buy on Amazon
The Algorithm by Hilke Schellmann

The Algorithm: How AI Decides Who Gets Hired, Monitored, Promoted, and Fired

Hilke Schellmann documents how high stakes algorithms get used, resisted, and audited, which maps directly onto the AI claim denial playbook.

Buy on Amazon

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

Case Studies of AI Denial Fallout

Three case studies reveal how hospital systems, legal aid programs, and state regulators have each responded to AI-driven healthcare insurance denials. The cases below show what has worked and what has not, with named organizations willing to share the numbers behind their efforts.

Case Study: Allina Health’s Dispute With Humana

Allina Health in Minnesota faced a growing operational problem in 2024 as its post-acute partners could not sustain the denial burden from Humana. Allina deployed a formal review and launched negotiations with Humana after denials for skilled nursing climbed to rates several times higher than other Medicare Advantage plans. The hospital system terminated its participation in Humana’s Medicare Advantage network after the denial rate increased by roughly 400 percent year over year in post-acute care. Allina presented data to state regulators showing the denial pattern tracked closely with Humana’s deployment of vendor review tools, not with patient acuity. Humana responded by denying that algorithms made final decisions and reiterated that licensed clinicians approved every denial. The Axios coverage of the Senate Medicare Advantage AI investigation cited Allina’s exit as a key data point in the broader pattern. The dispute now serves as a reference case in regulatory filings for hospital systems arguing that algorithmic denials threaten provider network stability. Several other large provider systems have now followed Allina’s example and renegotiated or exited Medicare Advantage contracts citing similar patterns.

The limitation of the Allina case lies in the counterfactual question that Humana and its defenders raise. Humana argues that higher denial rates can reflect appropriate use of criteria rather than inappropriate aggressive review, and the public data cannot fully distinguish those possibilities. Allina’s data was observational, not experimental, and some portion of the denial gap may reflect legitimate differences in patient mix across insurers. The outcome of pending Senate oversight and state investigations will determine how much weight regulators place on the pattern. For now the Allina case stands as the clearest example of a hospital system taking contractual action in response to algorithmic denial patterns at scale.

Case Study: Michigan Senior Advocacy Center Appeal Program

The Michigan Senior Advocacy Center piloted a legal aid program in 2024 that paired volunteer physicians with Medicare Advantage members denied skilled nursing coverage. The program deployed a bare-bones template plus manually curated clinical evidence as the solution to challenge denials rooted in algorithmic review. Across 180 cases handled in the first year, the program achieved an overturn rate of approximately 78 percent on internal appeal and 92 percent on external review. Members saved an estimated 4.2 million dollars in out of pocket post-acute care costs across the pilot. The program director has since testified before the Michigan legislature in support of state oversight legislation on AI claim adjudication. Program materials cite the HHS OIG report on inappropriate Medicare Advantage denials as foundational evidence that justifies the volunteer effort. The pilot has now expanded to Ohio and Pennsylvania under similar legal aid structures.

The limitation of the Michigan program is scale rather than efficacy, which is why advocacy groups frame it as a proof of concept. The program handled 180 cases in a region with tens of thousands of Medicare Advantage members, so most denied patients still never received advocacy support. The volunteer physician pool was difficult to grow because the time required for a thorough appeal letter runs between two and five hours per case. Replicating the overturn rate at national scale would require foundation funding, state matching, or federally supported legal aid expansion. The case study still demonstrates that overturn rates near 90 percent are achievable when patients receive even modest clinical and legal support against algorithmic denials.

Case Study: California Department of Managed Health Care Enforcement Actions

The California Department of Managed Health Care faced an enforcement problem in 2024 as several Medicare Advantage plans adopted algorithmic review faster than regulators could audit them. The department deployed a formal enforcement unit as the solution, pairing attorneys with licensed clinicians to review denial letters, model documentation, and plan audits. Settlements with two Medicare Advantage plans included multi-million dollar fines, roughly 15 percent reductions in flagged denial volume, and required public remediation plans with annual audits. The California Office of the Attorney General joined one of the actions after reviewing patient complaints and model documentation. The California press release on the Physicians Make Decisions Act underscored the state’s determination to make the new law bite immediately. The department now publishes an annual report with aggregate statistics on AI related complaints and enforcement outcomes. The initial reporting shows that targeted enforcement does measurably reduce blanket denial patterns within a plan’s California book of business.

The limitation of California’s approach is jurisdictional, which is why national patient groups push for parallel federal rules. California’s law reaches only state regulated plans, which exclude many self-funded employer plans governed by ERISA. Multistate payers can respect the California rules operationally in that market while running different workflows elsewhere, which blunts the national impact. Replicating California’s enforcement capacity in states with smaller regulatory agencies will require federal grants or shared enforcement partnerships. The case study still demonstrates that even one state with capacity, clear law, and political will can meaningfully shift payer behavior inside its borders.

The Future of AI in Claims Adjudication Beyond 2026

Looking ahead, the future of AI-driven healthcare insurance denials will likely bifurcate between constrained oversight tools and expanded human-centered appeals. Within five years, AI in claims adjudication will exist under mandatory documentation, independent audit, and transparent appeal rights across most of the country. CMS WISeR data will inform a federal rule that governs AI prior authorization in Original Medicare, with the first binding version expected around 2028. State AI insurance laws will converge toward a National Association of Insurance Commissioners model with uniform disclosure obligations. Payers will shift to systems that generate decision explanations, not just scores, which is the one feature regulators consistently prioritize. Patient-side AI tools will become default inclusions in employer benefits for appeals and advocacy.

The technology itself will continue to evolve in directions that could cut either way for patients. Large language models will produce more individualized denial letters that cite the member’s own record, which may improve legibility while obscuring systemic patterns. Vendors will likely offer counterfactual explainers that show which clinical facts would have changed the outcome, giving appellants a clearer target. The AI in healthcare transforming patient care trajectory also implies that insurers will adopt predictive risk tools that intervene before a denial becomes necessary. Those tools can route patients to effective care sooner, but they can also function as covert medical management. The ethics of that softer intervention layer are not yet widely debated.

The structural question for the next decade is whether AI tightens the loop between payers and patients or hardens it. If regulators force transparent audits, binding appeal rights, and mandatory model impact assessments, the technology can produce measurably better and faster coverage decisions. If instead regulators continue to lag and payers retain full discretion over model design, AI-driven denials will keep producing high reversal rates while shifting the cost of appeal to patients. The decisive variable will be political capacity to maintain pressure through the next administration transition. The safeguarding student privacy with AI tools playbook, in which federal action lagged state action for years, is a useful cautionary analogy. The patients who need protection the most also have the least capacity to maintain political pressure. Advocacy infrastructure therefore matters as much as policy design for durable change.

Chart From AIplusInfo

Medicare Advantage Denial Rates Climbed After AI Adoption

Share of post-acute care requests denied by the three largest Medicare Advantage insurers in 2019 versus 2022, after each adopted algorithmic review tools.

2019 denial rate2022 denial rate
UnitedHealthcare
2019
1.4%
2022
12.6%
Humana
2019
2.4%
2022
22.7%
CVS Health (Aetna)
2019
3.1%
2022
12.3%

Source: Senate Permanent Subcommittee on Investigations majority staff report, Refusal of Recovery, October 17, 2024, Tables 2 and 3.

Frequently Asked Questions on AI-Driven Insurance Denials

What exactly is an AI-driven healthcare insurance denial?

An AI-driven denial is a claim or prior authorization rejection produced in whole or in part by machine learning models that score medical necessity, length of stay, or billing accuracy. The models operate before any clinician meaningfully reviews the file. A licensed physician typically signs the denial, but the underlying decision originated in software.

Which health insurers have been sued over AI-driven denials?

UnitedHealthcare faces a federal class action over the nH Predict algorithm used in post-acute care. Cigna is sued over its PxDx bulk review system that processed claims in seconds. Humana faces similar litigation over algorithmic denials of skilled nursing coverage. These cases will shape how existing bad faith insurance law reaches algorithms.

What is nH Predict and why did it become controversial?

nH Predict is a NaviHealth tool, now operating inside UnitedHealth’s Optum unit, that estimates how long a Medicare Advantage patient should stay in post-acute care. StatNews reported the model was overturned on appeal roughly 90 percent of the time. Families say staff treated the model estimate as a discharge deadline rather than a planning aid.

What is Cigna’s PxDx system?

PxDx stands for Procedure to Diagnosis and automatically flags claims where the billed procedure does not match the diagnosis code to the system’s criteria. ProPublica documented that one Cigna medical director approved 60,000 denials in a single month, averaging about 1.2 seconds per file. California insurance regulators opened a formal inquiry into the PxDx system shortly after the reporting broke.

How often are AI-driven denials overturned on appeal?

Reversal rates on named systems like nH Predict reach approximately 90 percent when patients pursue a formal appeal. Internal Medicare Advantage audits show reversal rates between 75 and 90 percent across the three largest insurers. The high reversal rate reflects both model error and the structural bias that only resourced patients appeal.

Can an AI model legally deny a medical claim without a clinician reviewing it?

Federal rules require a qualified clinician to make the final medical necessity decision, and California now explicitly requires that review to be meaningful. Patients and prosecutors argue that two-second rubber-stamp sign-offs violate the spirit and increasingly the letter of these rules. Litigation and new state laws will decide what counts as sufficient human review.

What is the CMS WISeR Model and who does it affect?

WISeR is a five-year CMS pilot that tests AI prior authorization for 17 Part B service categories in Original Medicare. It operates in Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington starting January 2026. All adverse decisions must be made by a licensed clinician, with AI used only to screen and recommend.

Which states have the strongest laws on AI in insurance decisions?

California’s SB 1120 is the strictest, requiring a licensed clinician to review every AI-informed denial. Texas, Colorado, and New York passed similar restrictions during 2025. Pennsylvania and Illinois now require annual algorithmic impact assessments on health plans using automated review tools. The patchwork of state rules has forced national payers to build human-in-the-loop workflows across their books of business.

How can a patient figure out if AI was involved in their denial?

Request the denial letter and the specific clinical criteria used, in writing, within two weeks of the decision. Ask directly whether an automated tool scored the file, who the licensed reviewer was, and whether the reviewer practices in the relevant specialty. Federal rules require insurers to disclose this information when patients request it in writing.

What is the best first step after receiving an AI-driven denial?

File a written request for the full denial letter, clinical criteria, and reviewer credentials within the first two weeks. Prepare an internal appeal that quotes the specific coverage criteria and attaches a letter from the treating clinician. Keep the appeal under two pages so a human reviewer actually reads it. Use counter-AI appeal tools to speed up the drafting process.

Do counter-AI tools really improve appeal outcomes?

Yes, patient advocates report significantly higher overturn rates when appeal letters are drafted using tools like Claimable, Counterforce Health, or Fight Health Insurance. The tools pull clinical guidelines, randomized trial results, and prior appeal patterns into a tailored letter. The clinician still reviews and signs the letter, which gives it clinical weight. Several employers now cover these tools inside their benefit plans.

Does the Affordable Care Act give me the right to external review?

Yes, every ACA-regulated plan must offer an external review process once internal appeals are exhausted. External reviewers are independent clinicians whose decisions bind the plan if they side with the patient. Submission deadlines typically run four months from the final internal denial. State insurance departments can waive fees for members with financial hardship.

Can AI-driven denials be considered discrimination under federal law?

Yes, HHS has reinforced that algorithms producing coverage denials fall under Section 1557 of the Affordable Care Act, which prohibits discrimination by covered entities. The Department of Justice has signaled that AI-based claim denials can trigger False Claims Act exposure where plans misrepresent human oversight. The updated Section 1557 regulations now explicitly cover algorithmic decision support in federally funded health plans.

Will AI eventually produce faster and better coverage decisions for patients?

The answer depends on whether regulators enforce transparent audits, binding appeal rights, and mandatory impact assessments on claim adjudication AI. Transparent models with explainable decisions can produce measurably better and faster coverage decisions. Opaque models with high reversal rates will keep producing harm until systemic appeal infrastructure catches up. The decisive variable is the political capacity to maintain oversight through administration transitions.

Where can I report an AI-driven denial that seems illegal?

File a complaint with your state department of insurance and, for Medicare Advantage, with the regional CMS office that oversees the plan. The state attorney general handles systemic denial patterns in several states, including California, New York, and Minnesota. HHS Office for Civil Rights handles complaints about discriminatory denials under Section 1557. Patient advocacy groups can also help triage where to send a specific complaint.