Your claim came back denied within seconds. No phone call, no records request, nothing that looked like a person reading your file. If that felt less like a medical judgment and more like a machine saying no, you weren't imagining it.
Three
federal class actions filed against major health insurers now allege exactly
that — that predictive software, not individualized human review, decided
whether patients kept their coverage. None of these cases has produced a final
ruling on liability. But one of them just cleared a real hurdle: in March 2026,
a federal court ordered one insurer to open up its AI system to scrutiny for
the first time.
This
guide explains what's actually happened in court, what hasn't, how to tell
whether AI touched your own denial, and what to do next.
How Insurers Use AI To Process And Deny Claims
Software
touches nearly every claim an insurer processes. Most of that is unremarkable —
routing, duplicate checks, fraud flags. The dispute that's ended up in federal
court is narrower: whether software is making, or effectively making, the
decision to deny coverage.
Two tools sit at the center of the current litigation. nH Predict, built by naviHealth (a UnitedHealth Group subsidiary later folded into Optum), projects how many days of skilled nursing or rehab care a Medicare Advantage patient "should" need, based on comparisons to similar past patients. Plaintiffs in three separate cases allege insurers used that projected number as the actual cutoff date for coverage — regardless of what the treating physician recommended.
A naviHealth spokesperson has described the tool
differently, telling CBS News it functions as "a guide to help inform
providers" rather than a coverage determination on its own.
Why
This Matters to You: if your rehab or home-health coverage ended on a specific
day that didn't match your doctor's recommendation, and nobody could explain
why that day was chosen, that gap between "projected" and
"medically necessary" is worth documenting.
Cigna's PxDx ("procedure-to-diagnosis") process works differently — it batch-matches procedure codes against diagnosis codes for review. ProPublica's March 2023 investigation, based on internal company documents and interviews with former Cigna medical directors, reported that Cigna doctors denied more than 300,000 requests over two months in 2022, spending an average of 1.2 seconds per case.
One former
physician described the process bluntly: "We literally click and
submit". That's investigative reporting Cigna disputes, not a
court finding — but it's the fact pattern that triggered the Kisting-Leung lawsuit
discussed below.
Volume
is where the numbers get hard to ignore. In 2024, Medicare Advantage insurers
made roughly 53 million prior authorization determinations. About 4.1 million
were denied in whole or in part — but only a small share of denials were ever
appealed, and most of those appeals succeeded.
Medicare Advantage, 2024
| Metric (Medicare Advantage, 2024) | Figure |
|---|---|
| Total prior authorization determinations | ~52.8 million |
| Determinations denied in whole or in part | ~4.1 million (7.7%) |
| Denials that were appealed | 11.5% |
| Appeals that were overturned in whole or in part | 80.7% |
| Traditional Medicare prior-auth reviews (comparison) | 625,000 |
| Traditional Medicare denials (comparison) | 143,705 |
Federal auditors have reached the same conclusion from inside the data. In a 2026 review of 19 Medicare Advantage organizations, the HHS Office of Inspector General found that insurers overturned 95% of appealed denials of skilled nursing facility admission, calling the rate "extremely high" and saying it "raises concerns about denials that were not appealed."
The same review singled out one
contractor by name: naviHealth processed half of all SNF admission requests in
the study period and denied 14% of them, above both in-house reviewers (11%)
and other contractors (9%) — and 97% of naviHealth's denials were overturned
when appealed. OIG's earlier 2022 audit had already found that 13% of denied prior authorization requests actually met
Medicare coverage rules.
Read
those last two rows together: four out of five challenged denials get reversed,
but almost nobody challenges one. That gap is the real story — a screening
process that's either badly calibrated or working exactly as designed for
reasons that have nothing to do with medical necessity. Physicians see the same
pattern from the other side. In the 2025 AMA Prior Authorization Physician Survey of 1,000 practicing doctors, 60% said they're concerned that AI is increasing prior
authorization denials, and 26% reported that a prior-auth delay led to a
serious adverse event for a patient.
The reversal rate isn't a Medicare Advantage
quirk, either. Across marketplace and employer coverage the same lopsided
pattern shows up — denied claims that actually get appealed win far more often than almost anyone appeals them — which points the problem at the initial screen rather than at the underlying
medical facts.
An
important distinction: "algorithm" isn't the same as "AI,"
and neither one automatically means no human looked at your file. But a human
rubber-stamping a machine's output in 1.2 seconds isn't meaningful review
either. What matters legally isn't the technology label — it's whether your
claim got the individualized evaluation your policy and the law require.
The Major Lawsuits: Unitedhealth, Cigna, And Humana
Three
federal class actions anchor this area of law. All three have survived a motion
to dismiss, at least in part. None has reached a liability finding, a class
certification decision, or a settlement as of this writing.
Barrows v. Humana, Inc., No. 3:23-cv-00654-RGJ(W.D.Ky.) followed a similar pattern to Lokken. In a Memorandum Opinion and Order dated August 14, 2025, Judge Rebecca Grady Jennings allowed breach of contract, breach of the implied covenant, unjust enrichment, and common-law fraud to proceed, while dismissing the North Carolina unfair-claims, unfair-competition, insurance bad-faith, and unfair-and-deceptive-practices counts with prejudice.
Notably, the court reached the merits at all only because it waived the
Medicare Act's administrative exhaustion requirement, finding that exhaustion would
have been futile where plaintiffs alleged repeat denials issued immediately
after successful appeals.
Lokken is the case worth watching most closely. Gene Lokken, the 91-year-old lead plaintiff, broke his leg and ankle; his family says they paid thousands of dollars a month out of pocket after his coverage was cut off on a schedule the complaint says came from nH Predict, not his doctors.
In Estate of Gene B. Lokken v. UnitedHealth Group, Inc., 766 F.Supp. 3d 835 (D. Minn. Feb. 13,2025) — No. 0:23-cv-03514, decided by Judge John R. Tunheim — the court let Counts 1 and 2 proceed while dismissing Counts 3 through 7 with prejudice, including the pleaded insurance bad-faith count, as preempted by the Medicare Act. That split is a signal that a plan's own contractual promises of individualized review can survive even where state bad-faith statutes can't.
Kisting-Leung v. Cigna Corp., No.2:23-cv-01477-DAD-CSK (E.D.Cal.) turned partly on standing. In an order issued March 30, 2025, Judge Dale A. Drozd granted in part and denied in part Cigna's motion to dismiss, finding that several named plaintiffs couldn't tie PxDx to their own specific denials while allowing others to proceed — a reminder that in these cases, proving that an automated system touched your particular claim is often the harder legal question, not whether AI use is wrong in the abstract.
Case Citations And Court Records.
Estate of Gene B. Lokken v. UnitedHealth Group, Inc.
Citation: 766 F. Supp. 3d 835 (D. Minn. 2025) · Docket: No. 0:23-cv-03514 · Court: U.S. District Court for the District of Minnesota (8th Cir.) · Judge: John R. Tunheim · Filed: November 14, 2023 · Decided: February 13, 2025 · Parties: Estate of Gene B. Lokken, Estate of Dale Henry Tetzloff, and additional named plaintiffs v. UnitedHealth Group, Inc., UnitedHealthcare, Inc., and naviHealth, Inc. ·
Disposition: Motion to dismiss
granted in part, denied in part — Counts 1 (breach of contract) and 2 (breach
of implied covenant) proceed; Counts 3–7 dismissed with prejudice · Governing
statute: Medicare Act, 42 U.S.C. § 1395w-26(b)(3) (express preemption); 42
C.F.R. § 422.101(c) · Court record: govinfo.gov, United States Courts Opinions collection
Kisting-Leung v. Cigna Corporation
Docket: No. 2:23-cv-01477-DAD-CSK · Court: U.S.
District Court for the Eastern District of California (9th Cir.) · Judge: Dale
A. Drozd · Filed: July 24, 2023 · Order on motion to dismiss: March 30, 2025 ·
Parties: Suzanne Kisting-Leung, Samantha Dababneh, Randall Rentsch, Cristina
Thornhill, Amanda Bredlow v. Cigna Corporation and Cigna Health and Life
Insurance Company · Disposition: Granted in part, denied in part, with leave to
amend · Court record: Georgetown Law Health Care Litigation Tracker
Barrows v. Humana, Inc.
Docket: No. 3:23-cv-00654-RGJ · Court: U.S.
District Court for the Western District of Kentucky, Louisville Division (6th
Cir.) · Judge: Rebecca Grady Jennings · Filed: December 12, 2023 · Decided:
August 14, 2025 · Parties: Estate of JoAnne Barrows, Susan Hagood, Sharon
Merkley, Lorraine Kohl, Dolly Balani v. Humana, Inc. · Disposition: Motion to
dismiss granted in part, denied in part — Counts 1, 2, 3, and 8 proceed; Counts
4, 5, 6, and 7 dismissed with prejudice · Governing statutes: Medicare Act, 42
U.S.C. §§ 405(g), (h), 1395w-26(b)(3) · Court record: Georgetown Law HealthCare Litigation Tracker
Can an AI Denial Be Evidence of Bad Faith?
Not on its own — not yet, as a matter of settled law. No published ruling has held that using AI to screen or deny a claim is, by itself, bad faith. What the cases so far actually turn on is more specific: did the insurer's plan documents promise individualized review, and did the automated process deliver something less than that?
Federal regulators have already drawn a line that plaintiffs' lawyers are now borrowing. In a February 6, 2024 memo to all Medicare Advantage organizations, CMS told plans that an algorithm "can be used to assist MA plans in predicting a potential length of stay, but that prediction alone cannot be used as the basis to terminate post-acute care services" — the patient "must no longer meet the level of care requirements" at the time of termination, a judgment that can only come from re-assessing the individual.
That obligation isn't guidance alone; it sits in the regulation
itself. 42 C.F.R. §422.101(c) requires
medical necessity determinations to be based on the enrollee's own medical
history, physician recommendations, and clinical notes.
For employer-sponsored coverage the parallel duty runs through ERISA. The Department of Labor's claims-procedure regulation, 29 C.F.R. § 2560.503-1, requires an adverse benefit determination to identify the specific plan provisions relied on and, on request, to disclose the internal rules, guidelines, protocols, or criteria used — which is the legal hook behind the practical advice later in this guide about asking, in writing, what criteria decided your claim.
Courts were already drawing that line long before
AI entered the picture — most visibly in Woods v. Progressive American Insurance Co.,
where a Florida policyholder won a jury verdict nearly double her policy limit
and still lost her bad-faith case, because what the law measured was the
insurer's conduct while the claim was open, not how the claim eventually turned
out.
Note:
This is Woods v. Progressive American Insurance Co. (11th Cir. No. 23-13407), a
first-party underinsured-motorist bad-faith case. It is unrelated to the
similarly named Wood v. Progressive Select Insurance Co. (11th Cir. No.
24-13479), a third-party bad-faith case over a $50,000 policy discussed here.
Practical Takeaway for Policyholders: the legal theory isn't "AI is illegal." It's "you promised me a person would look at this, and the record suggests nobody really did." That's why documentation — what your denial letter says, and doesn't say, about who or what reviewed your claim — matters more than the word "AI" itself.
State Laws & Compliance Deadlines
| Law | Signed | Statute effective | Operational compliance deadline |
|---|---|---|---|
| California SB 1120 (Physicians Make Decisions Act) |
Sept. 2024 | Jan. 1, 2025 | Jan. 1, 2025 |
| Texas SB 815 (automated decision systems) |
Jun. 2025 | Sept. 1, 2025 | Jan. 1, 2026 (plans issued/renewed on or after) |
A cure period — the statutory window in which an insurer can fix the conduct you complained
about and extinguish a bad-faith suit before it's filed — is why timing governs
these disputes as much as substance does, and an AI-specific statute doesn't
displace it. If an automated denial gets quietly reversed inside that window,
the reversal may be the whole remedy.
Outside
California and Texas, most states haven't passed AI-specific insurance
statutes, which means a policyholder's strongest argument usually still runs
through ordinary contract and bad-faith law — proving the insurer didn't do
what it promised — rather than a new AI-specific rule.
The rights that attach the moment a denial letter arrives don't turn on whether a person or a model produced it — a written explanation,
disclosure of the criteria relied on, and a defined appeal window are owed
either way, which is precisely why the "was it AI?" question is worth
less to a policyholder than the "did you do what you promised?"
question.
State insurance regulators are moving through a parallel channel. The National Association of Insurance Commissioners adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers on December 4, 2023, setting out state regulators' expectations that insurers maintain a documented AI governance program and remain accountable for outcomes produced by third-party models.
It isn't a statute, and it binds nothing on its own — but it's the document a state insurance department is most likely to reach for when a complaint lands on its desk.
On the oversight side, the Government Accountability Office reported in 2025 that CMS's own monitoring of Medicare Advantage prior authorization has gaps, which is part of why so much of this is being worked out in litigation rather than by regulators.
Courts Allowing Discovery Into Insurer AI Systems
That's
a meaningfully different thing from a ruling on the merits, and it's worth
being precise about the difference. The court didn't decide UnitedHealth did
anything wrong. It decided the plaintiffs are entitled to see how the tool
actually works before that question gets answered — records the company had
previously kept out of plaintiffs' hands. One insurance-coverage attorney following
the case told Bloomberg Law the ruling leaves the bigger question "The courts will hopefully determine whether or
not it's acceptable to use this kind of technology in this
context without disclosing it.
Why
This Matters to You: discovery orders like this are how the public eventually
learns things insurers don't volunteer — error rates, whether the tool was
designed to override physician judgment, how often internal reviewers actually
overrode the algorithm's output. If Lokken produces internal documents showing
the tool operated the way plaintiffs allege, that evidence could shape not just
this case but how every future AI-denial claim gets litigated. If it doesn't,
that matters too. Either way, nothing has been decided yet — a policyholder
relying on this article next year should check whether the underlying documents
surfaced anything, since this is exactly the kind of fact pattern that changes
fast.
How to Tell If AI Was Involved In Your Claim Denial
This is where the law moved furthest and fastest in 2026. On March 9, 2026, a federal magistrate judge in the Lokken case ordered UnitedHealth to produce a broad set of internal records: policies and training materials for post-acute care claims, internal analysis of how nH Predict performs, and business records tied to the acquisition of naviHealth.
The court granted or partially granted the plaintiffs' requests across six of
seven document categories, and rejected the insurer's argument that records
predating nH Predict's 2019 rollout were irrelevant.
The court's reasoning on that last point leaned
on a congressional investigation. The Senate Permanent Subcommittee on
Investigations' October 2024 majority staff report, Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care,
found — on the basis of more than 280,000 pages of internal documents
subpoenaed from three insurers — that UnitedHealthcare's post-acute prior
authorization denial rate rose from 10.9% in 2020 to 16.3% in 2021 to 22.7% in
2022 as the company automated more of the process. A before-and-after baseline,
the court concluded, was exactly what the earlier records could establish.
You
often can't tell from the denial letter alone, but a few patterns are worth
checking.
Original Editorial Insight: A pattern worth watching in these cases: the strongest tell isn't a mention of 'AI' — insurers rarely use that word in the letter itself. It's a denial that cites a specific number (a day count, a length-of-stay figure) with no accompanying clinical explanation of why that number applied to your case specifically.
Signs
worth checking:
- Turnaround time. A denial issued in minutes, with no records request in between, is a signal — not proof — that no one opened your file.
- A number instead of a reason. "Exceeds expected length of stay for this diagnosis" is a projection, not an individualized medical judgment. Ask what individualized judgment was applied on top of it.
- No named reviewer. ERISA plans in particular are required to identify the criteria relied on; ask for the name and credentials of whoever made the decision.
- A form letter that doesn't mention your specific treatment. Denials built from templates sometimes leave in language that doesn't match your actual diagnosis or facility — a strong sign the letter was generated in batch.
- A named tool, guideline, or "screening criteria" you don't recognize. Denial letters sometimes cite a proprietary system or criteria set by name instead of a plain clinical explanation. You're entitled to ask what that name refers to and how it was applied to your case.
What you send back matters as much as what you
spot. One claimant learned that the hard way after assuming the treating
physician's chart would speak for itself — the medical evidence that actually moves a denied claim turned out to be dated, specific, function-focused documentation nobody had
thought to request, and by the time that gap was obvious the file had been
closed twice.
Steps To Take If You Suspect An Automated Denial
What To Do If Your Claim Was Denied
| Step | What to do | Why it matters |
|---|---|---|
| 1 | Request your complete claim file — ask in writing for reviewer names, credentials, and time-stamped notes | Gaps or refusals are themselves meaningful evidence |
| 2 | Ask directly whether AI or an algorithm was used — put the question in writing, not just on a call | Creates a dated record of the request and any evasive answer |
| 3 | File an internal appeal before any deadline — check your denial letter for the appeal window | Roughly a third of internal appeals succeed — this is the highest-leverage step most people skip |
| 4 | Request external review if the internal appeal fails — ask your insurer or state regulator how to escalate | Only about 34% of eligible marketplace enrollees know this option exists, let alone use it |
| 5 | File a regulatory complaint stating you believe an automated system made the decision without individualized review | Regulators can compel answers you can't get on your own |
| 6 | Talk to a lawyer about which framework applies to you — ERISA, state bad-faith law, and Medicare appeals are different tracks with different deadlines | Most bad-faith attorneys review denials for free |
| 7 | Preserve everything — keep denial letters, portal screenshots, call logs, and your own correspondence | A clean, factual record is more persuasive than an emotional one |
Step 5 carries more weight in some states than
others: the specific-violation requirement behind a formal bad-faith notice is what most defective filings fail on, and a complaint that says only "a
computer denied my claim" — without naming the conduct and the provision
it violated — gives a regulator almost nothing to act on.
Core Lessons: the biggest mistake isn't missing a legal argument — it's missing the deadline. Internal appeal windows and external review windows are unforgiving, and once they close, even a strong case can be foreclosed permanently.
Deadlines aren't the only door that closes. A
Florida claimant once found this out after signing a settlement release too quickly — the language waived far more than the single disputed claim it was meant to
close, and by the time that became clear there was nothing left to argue about.
Editorial Disclaimer
This article summarizes ongoing litigation
(Lokken, Kisting-Leung, Barrows), regulatory guidance, and investigative
reporting on AI-driven claim denials. It is for general informational purposes
only, not legal advice.
None of the lawsuits discussed has reached a
liability finding, class certification, or settlement — each has only survived
a motion to dismiss in part, and outcomes described here (including the March
2026 discovery order) may change as these cases proceed. Confirm the current
status of any case before relying on it.
Whether AI was involved in a specific denial, and
whether that matters legally, depends on your plan's own terms (ERISA, Medicare
Advantage, or state law each apply differently) and the facts of your claim —
not on the general patterns described here. The "signs of automated
denial" listed are indicators to check, not proof of anything.
State AI-insurance laws are limited and evolving;
the California and Texas statutes referenced may not apply to your policy, and
other states may adopt new rules after publication.
For a specific denial, consult your denial
letter, your plan documents, or a licensed attorney familiar with insurance
bad-faith or ERISA claims.
Post a Comment