Medicare is using AI to approve claims. The result has been ‘alarmingly high denial rates.’
The promise of artificial intelligence in health insurance is speed: fewer backlogs, faster approvals, and lower administrative costs. In the Medicare ecosystem, algorithms now screen, score, and in some cases help decide whether care gets authorized or paid. But as automated tools have moved from piloting to practice—especially in Medicare Advantage—the story has gotten more complicated. Clinicians and beneficiaries report care cut short, necessary services delayed, and denials reversed only after long appeals. Federal watchdogs have flagged patterns of inappropriate denials. And regulators have warned plans not to let algorithms override Medicare’s coverage rules or clinicians’ judgment.
The result is a paradox at the heart of AI-driven utilization management: the tools built to approve claims more efficiently are also producing denial rates that many find unacceptably high.
What “Medicare” means in this debate
– Original Medicare (fee-for-service) is administered by the federal government and its contractors. It uses automated systems and analytics to screen claims for errors and fraud and applies some prior authorization in limited areas (for example, certain durable medical equipment).
– Medicare Advantage (MA) is run by private insurers paid by Medicare to cover beneficiaries. MA plans apply prior authorization much more broadly and often use proprietary algorithms to decide when care is covered, how long it’s covered (for post-acute stays, for example), and when payment is warranted.
Most of the current controversy sits inside Medicare Advantage, which now enrolls roughly half of all beneficiaries. That scale means algorithmic policies affect millions.
What the evidence says about denials
Well before AI became a buzzword, federal oversight found patterns of inappropriate denials in Medicare Advantage:
– A 2018 report by the Department of Health and Human Services Office of Inspector General (HHS-OIG) found that MA plans overturned about three-quarters of their own denials when beneficiaries appealed—an indicator that many initial denials were erroneous.
– In 2022, HHS-OIG reviewed a sample of MA decisions and concluded that 13% of prior authorization denials and 18% of payment denials actually met Medicare coverage rules and should have been approved.
As AI and predictive models spread in claims review and prior authorization, lawsuits and investigative reporting described how tools were being used in practice—especially in post-acute care. One high-profile example focused on algorithms that set “expected” lengths of stay in skilled nursing or rehabilitation; once a patient hit the predicted day, coverage was often cut off unless providers cleared additional hurdles. Clinicians argued those predictions ignored individual recovery trajectories; families found themselves appealing to restart coverage that Medicare’s own rules would otherwise allow. While many denials were ultimately overturned, the delays and administrative burden were costly in their own right.
Why algorithms can drive high denial rates
– One-size-fits-most medicine. Predictive tools are trained on historical averages. They can nudge decisions toward the “typical” case and away from outliers, even when outliers are clinically justified.
– Proxy incentives. Even if models are labeled “clinical,” plans still optimize for utilization and cost. If a threshold is set to minimize “unnecessary” days, the model can overstep and curtail necessary care unless guardrails force patient-specific exceptions.
– Opaque criteria. Proprietary scoring and black-box thresholds make it hard for clinicians to know what documentation the algorithm expects, leading to avoidable denials.
– Automation bias. When reviewers are asked to “add a human check” on an algorithm’s recommendation, they may default to the machine’s output unless they have time, authority, and complete clinical context to differ.
– Data gaps. Claims data miss clinical nuance. If social needs, complications, or atypical recovery courses aren’t captured, the model’s “medical necessity” judgment can be wrong.
Regulators are responding
The Centers for Medicare & Medicaid Services (CMS) has issued new rules and guidance to curb overreach and restore due process:
– Follow Medicare’s coverage rules. MA plans must apply the same clinical coverage criteria used in Original Medicare (national and local coverage determinations) when they exist. Plans can’t substitute proprietary rules that are more restrictive.
– Algorithms can’t decide alone. CMS has told plans they may use algorithms or predictive models to support utilization management, but they cannot use them as the sole basis for denying care. Individual circumstances must be considered, and exceptions must be allowed.
– Explain denials. Plans must provide a specific, patient-level rationale for denials, not just cite generic policies or model outputs.
– Faster decisions and transparency. A 2024 CMS rule tightens prior authorization timeframes (e.g., 72 hours for urgent requests, seven days for standard in many programs) and requires plans to publicly report prior authorization metrics.
These steps don’t ban AI; they set conditions for its use. The burden now falls on plans to prove their tools are accurate, fair, and consistent with Medicare law—and on CMS to audit and enforce.
The human and system costs of getting it wrong
– Delayed care and worse outcomes. Appeals can take weeks. In post-acute settings, a gap of even a few days can stall recovery.
– Administrative overload. Clinicians spend hours on documentation, peer-to-peer calls, and appeals, time diverted from patient care.
– Financial strain on families. When coverage is cut off early, beneficiaries may face unexpected bills or discharge to settings that can’t meet their needs.
– Erosion of trust. When denials are frequently reversed, patients and providers see the process as arbitrary; when they aren’t, they view it as punitive.
How to make AI work for Medicare beneficiaries
If AI is here to stay in claims and prior authorization, a safer architecture is possible:
– Start with Medicare’s rules, not model outputs. Models should flag cases for review, not impose coverage limits that deviate from national or local coverage determinations.
– Human-in-the-loop, not human-on-the-loop. Require documented, accountable clinical judgment on every adverse decision, with authority to override algorithms.
– Patient-specific reasoning. Denial notices must cite the individual’s clinical facts, the applicable Medicare rule, and what additional documentation could change the outcome.
– Publish performance. Plans should report denial rates, overturn rates, average time to decision, and reasons for denials by service line. Outliers should trigger audits.
– Independent validation. Third-party evaluation of algorithms for accuracy, bias, and safety, including stress-testing on atypical cases and vulnerable populations.
– Audit trails. Maintain and provide logs showing what the model recommended, what data it used, and who made the final decision.
– Align incentives. Tie plan star ratings and payments to appropriate approval rates, timely decisions, and beneficiary outcomes—not just cost containment.
What beneficiaries and clinicians can do now
For beneficiaries (and families):
– Ask for the rule. When a service is denied, request the exact Medicare coverage policy (NCD/LCD) and the patient-specific reason cited.
– Use expedited appeals. If a denial delays care and health could worsen, request an expedited review. In Medicare Advantage, many organization determinations must be made within 72 hours if urgent.
– For post-acute discharge disputes. If you’re told coverage is ending at a skilled nursing facility or rehab, you have a right to a fast appeal through the Beneficiary and Family Centered Care Quality Improvement Organization (BFCC-QIO). File before the deadline on the notice to keep coverage during the review.
– Keep records. Save letters, dates, names, and notes from calls. If an algorithm was referenced, ask for the clinical basis used.
– Get help. Contact your State Health Insurance Assistance Program (SHIP) for free counseling, and consider assistance from your clinician’s office or a patient advocate.
For clinicians:
– Document to Medicare’s standard. Anchor notes to the relevant NCD/LCD and medical necessity criteria; explicitly describe why the patient deviates from “typical” recovery curves.
– Peer-to-peer strategy. Prepare with specific clinical milestones, risks of delay, and why alternative settings are inappropriate.
– Cite CMS guidance. Remind plans that algorithms cannot be the sole basis for denial and that MA coverage must match Original Medicare unless no policy exists.
– Track patterns. Escalate repeated denials by service line to medical staff leadership and professional societies; patterns support complaints to CMS or state regulators.
The bigger picture: AI as infrastructure, not arbiter
Used well, automation can clear backlogs, reduce paperwork, and get routine approvals out the door faster. But the boundary between “workflow support” and “automated denial engines” is thin. In Medicare—where coverage rules are public, and the stakes are literally life and death—AI should be infrastructure: a triage tool to surface what needs human judgment, not a gate that quietly narrows access.
The current controversies are less a referendum on AI than on governance. When oversight is weak, black-box tools drift toward cost control at the expense of clinical nuance. When rules are clear, audits are real, and beneficiaries have fast, fair appeals, algorithms can help rather than harm.
Medicare’s challenge is to capture the speed AI promises without reproducing the denial patterns oversight bodies have already warned about. That means demanding transparency, centering patient-specific decisions, and measuring what matters: not how quickly a claim is denied, but how reliably necessary care gets to the people Medicare is designed to protect.
