When the Algorithm Becomes the Adjuster: Senate Scrutiny of AI in Medicare Advantage Denials

By Elena Pak, Credentialing Department, WCH

Two senators just gave three of the country’s largest health insurers two weeks to explain their machines.

The Letters

On July 14, Sens. Richard Blumenthal (D-Conn.) and Josh Hawley (R-Mo.) sent formal document requests to UnitedHealthcare, Humana, and CVS Health, demanding a full accounting of how each company uses algorithms and AI in post-acute care coverage decisions. The senators want to know whether it is still each company’s stated policy that a machine cannot issue a final denial for a Medicare Advantage member — and they want a complete inventory of every predictive tool currently in use. The deadline for a response: July 28.

The pairing of Blumenthal and Hawley is itself notable. One is a longtime consumer-protection Democrat; the other is a populist Republican who has made tech accountability a signature issue. When senators this far apart on everything else agree on a target, it tends to mean the underlying data is hard to spin.

And the data here is unflattering. The letters cite two OIG reports released June 8 that examined prior authorization decisions across 19 Medicare Advantage insurers in June 2024. One found that UnitedHealth, Humana, and CVS denied requests for long-term acute care and inpatient rehabilitation at rates above their peers. The other found the same three companies denied 12% of skilled nursing facility requests — and that nearly every one of those denials, when appealed, got overturned. Both reports singled out naviHealth, the utilization-management vendor UnitedHealth operates through Optum, as a recurring factor.

That last detail is the one worth sitting with. A denial that gets reversed on appeal almost every time isn’t really a coverage decision — it’s a delay tactic with a due-process escape hatch. Whether that pattern comes from the algorithm, the humans supervising it, or both is precisely what the senators are trying to pin down.

Money Is the Subtext

The senators didn’t bury the financial argument — they led with it. Their letter notes that the federal government will spend an estimated $76 billion more this year covering Medicare Advantage enrollees than it would if those same people were on Traditional Medicare instead. Set against that backdrop, rising denial rates in exactly the care categories where AI-assisted review has expanded reads, to the senators, less like efficiency and more like a business model.

That’s the political frame. It’s also, separately, a fair empirical question: does utilization-management technology reduce unnecessary care, or does it just reduce paid claims? Those are not the same thing, and MA plans have strong financial reasons to prefer the second while describing it as the first.

This Isn’t a New Fight — It’s an Escalation

None of this started with the July letters. It builds on:

An October 2024 Senate subcommittee report — based on more than 280,000 pages of internal documents — that found denial rates for post-acute care rose specifically as these insurers adopted automated review.
A follow-up Blumenthal letter last October pressing the same three companies on how their AI tools interact with human clinical judgment.
A separate, still-active federal class action in Minnesota, where families of deceased MA members allege that a naviHealth AI product drove post-acute care denials. A magistrate judge already ordered UnitedHealth to turn over broad discovery on the tool and its internal policies back in March.
A parallel, unrelated Senate track — Warren and Wyden expanding their own probe into UnitedHealth’s nursing home practices in January, following Guardian reporting on residents who allegedly died after Optum delayed or denied hospital transfers. UnitedHealth has denied those allegations and is suing the outlet for defamation.

Lay these side by side and a pattern emerges that’s bigger than any single letter: courts, regulators, and Congress are all independently arriving at the same insurer, the same vendor, and the same clinical category — post-acute care — as the place where AI-assisted denial is most aggressive.

What the Companies Are Saying

The public responses so far split into three postures. CVS confirmed receipt and said it would respond to the senators — a holding statement, nothing more. Humana declined to comment entirely.

UnitedHealthcare pushed back hardest, arguing the OIG reports take a narrow view of a process that gets the vast majority of requests approved, often instantly, and that reversals on appeal simply reflect a system built to re-evaluate cases as new information comes in. The industry trade group AHIP made a similar case at the sector level, arguing the OIG analysis leaves out administrative denial reasons and ignores legitimate variation in post-acute care costs and quality.

Both of those defenses are worth taking seriously — appeal-reversal rates genuinely can reflect legitimate re-review, not just bad first decisions. But that argument gets harder to make at scale. A process built for occasional re-evaluation should not produce reversal rates near total, across an entire denial category, at three of the largest plans in the market simultaneously. At some point “we welcome appeals” stops being reassurance and starts looking like the actual quality-control mechanism, with the initial review functioning as a filter that assumes most people won’t push back.

The Real Compliance Question Underneath This

Strip away the politics and there’s a genuine governance question every MA plan, and honestly every payer running utilization management software, needs an answer to right now: can you actually document, case by case, that a human clinician — not the model — made the final call on a denial?

That’s not a rhetorical standard. CMS already requires that medical necessity determinations for Medicare Advantage be made based on individualized assessment, not algorithm output alone, and this inquiry is effectively testing whether that requirement is being met in practice or just in policy documents. If a plan’s answer to “show me the human decision” is a rubber-stamped sign-off on a system-generated recommendation, that’s not going to hold up under subpoena, in court, or in front of a Senate committee — regardless of what the vendor contract says about “human-in-the-loop.”

For compliance and UM leadership at any Medicare Advantage plan, the practical takeaway from this episode isn’t “wait and see.” It’s an audit checklist:

Can you produce, for any denial, the specific clinical rationale a human reviewer added — distinct from the algorithm’s output? If the answer is “the reviewer approved the system recommendation,” that’s a gap, not a safeguard.
Do you track reversal rates by review pathway (algorithm-flagged vs. standard clinical review) separately? If you don’t already know whether your AI-flagged denials get overturned at a different rate than your other denials, you don’t actually know what the tool is doing to your members.
Is your predictive-technology inventory current and complete — including vendor tools like naviHealth-style platforms that sit a layer removed from your own IT stack? Regulators are now explicitly asking for this list; plans that can’t produce one quickly have a bigger problem than a slow response to a letter.
Does your appeal data get fed back into model or policy review, or does it just close the case file? A near-total reversal rate on appeal that never changes the upstream process is the single data point most likely to attract exactly this kind of scrutiny.

Where This Goes Next

July 28 is not a deadline with teeth by itself — it’s a Senate request, not a subpoena. But the surrounding context is what makes it matter: an active federal lawsuit already compelling discovery on the same technology, an OIG finding that’s hard to explain away, and two senators from opposite ends of the political spectrum who now have a reason to keep pulling on this thread regardless of what comes back in the responses. Whatever these three companies submit will very likely surface in the Minnesota litigation, in a future hearing, or both — which means the responses due July 28 are the first draft of a record that’s going to follow these companies for a while.

Sources