When the Algorithm Says No: A Working Guide to Appealing AI-Driven Insurance Denials

The denial came back in ninety-two seconds. The skilled nursing facility had submitted the request at 11:46 p.m. The carrier's portal time-stamped the denial at 11:47 p.m. and change. The patient was a 68-year-old retired letter carrier in Denver finishing his seventeenth day of rehabilitation after an ischemic stroke that had left his right side weak and his speech halting; his physical therapist had charted a thirty-day plan with measurable goals at days twenty-one and twenty-eight; his Medicare Advantage plan had pre-authorized the stay in five-day increments up to this point. The notice he received in early April contained no physician name and cited a generic length-of-stay threshold as its criterion. The discharge planner pulled it out of the portal the next morning. The patient's wife asked the only question that mattered: who actually read his chart? The discharge planner did not know. Ninety-two seconds is not enough time for a clinician to read a chart. It is roughly the time it takes a piece of software to score a length-of-stay forecast against a cutoff and emit a structured response.

That question is now the operative one in a growing share of American insurance denials. Predictive algorithms, claim-batching software, and length-of-stay forecasting tools have been embedded into utilization management at scale since roughly 2020, and federal regulators, state legislatures, and class-action plaintiffs' firms have all moved on the issue inside the last twenty-four months. The appeal landscape changed materially in 2024 and 2025, and the patient who knows what changed can write a counter-paragraph that lands.

What an "AI-driven denial" actually means in 2026

Three categories of tool matter for an appeal.

The first is the predictive algorithm that forecasts a clinical trajectory and is then used to set a coverage boundary. The best-documented example is the nH Predict tool UnitedHealthcare's naviHealth subsidiary deployed in post-acute care decisioning. Plaintiffs in Estate of Lokken v. UnitedHealth Group, Case No. 0:23-cv-03514 (D. Minn.), allege the model produced length-of-stay estimates that functioned as a de facto cap on continued-stay authorizations. The carrier disputes the characterization. Litigation is active and unresolved.

The second is the claim-batching or rules-engine tool that processes post-service claims against carrier-defined criteria at machine speed. The best-documented example is the PXDX tool described in Kisting-Leung v. Cigna Corp., Case No. 2:23-cv-01477 (E.D. Cal.). ProPublica's March 2023 investigation by Rucker, Miller, and Armstrong reported the system enabled review of large batches with limited individualized clinician examination. Cigna disputes the characterization and states the tool matches codes to coverage policies rather than making medical-necessity determinations. The case has cleared early motions.

The third is the upstream decision-support tool that recommends an approval or denial to a human reviewer who then issues the determination. These are the most common and least visible. The denial names a human reviewer; the recommendation was generated by software. Whether the reviewer meaningfully exercised independent clinical judgment, or rubber-stamped the algorithm, is the contested question driving most current AI-denial appeals.

The regulatory landscape that changed in 2024 and 2025

Three regulatory developments reshaped what patients can demand on appeal.

The first is the CMS guidance issued February 6, 2024, in a memorandum to MA organizations on the use of algorithmic tools in coverage decisions. It reiterated 42 CFR 422.101(c) (MA plans must furnish medically necessary Part A and B benefits) and stated that an algorithm or software tool cannot be the sole basis for denying a Medicare-covered service on medical-necessity grounds. It did not prohibit AI in claims processing; it framed the rule as a procedural safeguard requiring an individualized determination by a qualified clinician, informed by the patient's specific record, behind any adverse medical-necessity determination.

The second is California Senate Bill 1120, signed September 28, 2024, effective January 1, 2025. It amends Cal. Health & Safety Code Section 1367.01 and Cal. Insurance Code Section 10123.135 to require that UM decisions on medical necessity be made by a licensed physician or other appropriately licensed professional, that any algorithm or AI tool be fairly and equitably applied, that no such tool be the sole basis for a denial, modification, or delay, and that the carrier disclose its use on request. It applies to plans regulated by the California DMHC and to insurers regulated by the California DOI.

The third is the state activity that followed (see Exhibit 1). Outside MA, the federal hook is 29 CFR 2560.503-1, whose subsection (h)(3)(iv) requires review by a named fiduciary who is not the original decisionmaker and whose (m)(8) requires identification of medical experts whose advice was obtained.

How to identify an AI-driven denial, and the four questions to ask

The four diagnostic signals are summarized in Exhibit 2 below: speed, formulaic language, absence of case-specific reasoning, and a missing or implausible named-reviewer line. When three or four are present together, the denial is likely algorithmic.

Before drafting, run four questions against the denial and the carrier's records response. One. Who is the named medical reviewer, what are her credentials, and what is her board certification? If the denial does not name one, request the name under 29 CFR 2560.503-1(m)(8) for ERISA plans, under California SB 1120 for California-regulated plans, or under the parallel state-code provision. Two. What specific criteria were applied, and from what medical-policy document? Patients have a right to the specific clinical criteria under 29 CFR 2560.503-1(h)(2)(iii) for ERISA plans and 45 CFR 147.136(b)(2)(ii)(C) for ACA plans; a generic policy that does not identify which criterion the patient failed is procedurally vulnerable. Three. Was the record actually read? Request written confirmation that the named reviewer accessed the records, with date and time. Four. What is the carrier's published policy on algorithmic tools in this category? A gap between the published policy and actual conduct is itself a basis for regulatory complaint.

The procedural weight a self-prepared appeal carries

AI-denial appeals turn on disclosure rights and procedural arguments that read straightforward on paper and require very specific drafting in practice. The 29 CFR 2560.503-1(h)(3)(iv) and (m)(8) reviewer-disclosure framework, the CMS February 2024 MA memorandum, and the California SB 1120 framework each cite different statutory language and produce different counter-paragraphs. The patient who copies a generic AI-denial template off the internet often misses the state-specific cite that makes the demand enforceable, or names the wrong CFR subsection, and the carrier routes the appeal back as procedurally insufficient.

The mapped library Apellica has catalogued (more than two hundred carrier-by-denial-type cells, indexed at the bulletin level) each carry their own algorithmic-tool footprint. nH Predict at UnitedHealthcare for post-acute care, PXDX at Cigna for batched commercial claims, the Humana decision-support tool at issue in Barrows, and the unnamed upstream tools at every other major carrier each leave a different audit-trail fingerprint. The 30-day document-request right that compels reviewer-credential disclosure requires a demand letter with the correct CFR cite.

Procedural exhaustion missteps foreclose the class-action coordination and the regulatory-complaint path that often produces the fastest reversal. The carrier's UM department deploys these tools at scale. The patient is appealing one denial.

Ninety-two seconds is not enough time for a clinician to read a chart. The denial is a forecast, not a finding.

What separates a desk-prepared appeal from a self-prepared one

Apellica's reviewers work from a mapped library of over two hundred carrier-by-denial-type cells that tracks the algorithmic-tool footprint at every major carrier and Medicare Advantage organization. The desk tracks the named-reviewer disclosure patterns, the typical response-time thresholds that indicate algorithmic processing, and the state statutes (California SB 1120 and the follow-on bills in Colorado, New York, Texas, Illinois, Washington) that produce the strongest disclosure demand.

Same-day reviewer-credential demand letters go out with the correct CFR and state-statute cite. Apellica's senior reviewers build the four-part evidence stack, plan-language citation, clinical facts, peer-reviewed evidence, regulatory hook with the AI-specific disclosure framework, for every case. Parallel state DOI or DMHC complaints are filed where the state framework supports them. A senior reviewer reads every appeal before it goes out.

Initial review is free. There is no upfront fee. Patients are not asked to pay anything until the carrier reverses the denial.

How individual appeals connect to the class-action landscape

The three class actions share a structural theory: each alleges the carrier deployed an algorithmic tool whose output drove adverse determinations without meaningful individualized clinician review, and that the carrier knew or should have known the tool produced higher error rates than human reviewers. None has produced a verdict. Class certification is not a verdict.

The connection to an individual appeal is twofold. First, the public filings give patients a vocabulary and documentary record for the named-reviewer and algorithm-disclosure questions; citing them raises the institutional cost of defending the denial. Second, the patient's denial may itself fall within a putative class; preserve correspondence and consult counsel. Individual appeal and class participation are not mutually exclusive.

Where to ask for help

Two regulatory channels matter most. The state Department of Insurance, or in California the DMHC for Knox-Keene plans, runs consumer-complaint processes that move faster than litigation and that carriers track as regulatory exposure. A complaint documenting an AI-driven denial with the records response attached creates a record that influences how the carrier handles the underlying appeal. The Department of Labor's EBSA handles ERISA disputes, particularly where the carrier has failed to meet the disclosure obligations under 29 CFR 2560.503-1.

Two additional channels: the CFPB accepts complaints involving medical-debt collection arising out of denied claims; and ProPublica maintains a tipline for the reporting team behind the original PXDX investigation.

Exhibit 1: State AI-denial-review statutes effective in 2026

Leading enacted and pending measures as of mid-2026; confirm current statutory text before citing.

| State | Statute or measure | Status as of mid-2026 | |---|---|---| | California | SB 1120 (Health & Safety Code 1367.01, Insurance Code 10123.135) | Enacted September 2024, effective January 1, 2025 | | Colorado | SB 23-169 and Division of Insurance rulemaking | Enacted 2023, health-insurance rulemaking in progress | | New York | Assembly Bill A6188 | Reintroduced 2025-2026 session, in committee | | Texas | HB 2727 (introduced 2025) | Pending | | Illinois | HB 3801 (introduced 2025) | Pending | | Washington | SB 5575 (introduced 2025) | Pending | | Federal Medicare Advantage | CMS memorandum, February 6, 2024 | In effect as guidance, not a rule |

Action title for designer: "California led. The federal guidance reaches Medicare Advantage and stops there. The patient appealing an AI-driven denial has different tools depending on which state she lives in and which plan she carries."

Exhibit 2: The AI tell diagnostic

The four signals that together mark a denial as likely algorithmic.

| Signal | What to look for | Why it matters | |---|---|---| | Speed of response | Decision returned within minutes of submission, especially outside business hours | Genuine individualized review takes time; rapid turnaround suggests automated processing | | Formulaic language | Stock paragraphs that recur across patients, conditions, and procedures | Indicates template-driven generation rather than case-specific clinician reasoning | | Absence of case facts | No reference to the patient's actual labs, imaging, functional assessments, or treatment history | Suggests the criterion was applied without reading the record | | Named reviewer line | Missing name, mismatched specialty, non-clinician credential, or implausible decision volume | Disclosure failure is itself a procedural ground for appeal |

Action title for designer: "No single signal is dispositive. Three or four together are the fingerprint of an algorithmic decision and the basis for the appeal that demands disclosure."

Exhibit 3: Major AI-denial class actions

Confirm against current docket activity before citing.

| Case | Court | Filed | Tool at issue | Status mid-2026 | |---|---|---|---|---| | Estate of Lokken v. UnitedHealth Group, 0:23-cv-03514 | D. Minn. | November 2023 | nH Predict (naviHealth) | Past initial dispositive motions, in active discovery | | Kisting-Leung v. Cigna Corp., 2:23-cv-01477 | E.D. Cal. | July 2023 | PXDX | Past early motions, in active proceedings | | Barrows v. Humana Inc., 3:23-cv-00654 | W.D. Ky. | December 2023 | Post-acute care decision tool | Pending |

Action title for designer: "Three carriers, three federal districts, one structural theory. None has reached a verdict. The appeal that cites the public filings is citing allegations under active litigation, not findings."

What to do if you have an AI-driven denial right now

The denial that came back in ninety seconds is not entitled to the same procedural weight as one issued after a clinician read the chart. The carrier knows that. The clock starts when the carrier dated the letter; most patients calendar the wrong day.

Most patients leave coverage on the table because the AI-denial appeal is more procedural work than they can take on.

The Denver letter carrier's appeal demanded the name and credentials of the human reviewer who signed the ninety-two-second denial. The carrier could not produce one. The thirty-day rehabilitation plan was reauthorized through day twenty-eight.

The Apellica model, briefly

Apellica prepares the evidence-based appeal letter for AI-driven denials across every line of business, with the counter-paragraph framework adapted to the patient's state, plan type, and the specific tool implicated. The patient reviews and approves every word before submission and authorizes carrier communications under a HIPAA-compliant Assignment of Benefits. We are not a law firm, medical provider, or insurance carrier. We are an independent administrative service that turns a denied claim into a properly documented appeal letter. $0 upfront, flat fee on successful recovery. Coverage in all 50 states. A senior reviewer reads every case.

About the author

About the author. Mark Henderson is a senior reviewer at Apellica, an independent appeal-preparation service for denied health-insurance claims. The office is at One World Trade Center, Suite 8500, New York, NY 10007. Apellica covers all fifty states. Apellica does not provide legal advice and is not a law firm. For questions: press@apellica.com, +1 (888) 777-6120, apellica.com.

References

• 42 CFR 422.101(c), 422.566, 422.582, 422.629. Medicare Advantage.
• 45 CFR 147.136. ACA Claims and Appeals.
• 29 CFR 2560.503-1. ERISA Claims Procedure, including (h)(2)(iii), (h)(3)(iii), (h)(3)(iv), (m)(8).
• CMS, Memorandum to MA Organizations on Use of Algorithmic Tools in Coverage Decisions, February 6, 2024.
• California SB 1120 (2024), amending Cal. Health & Safety Code 1367.01 and Cal. Ins. Code 10123.135. Effective January 1, 2025.
• Colorado SB 23-169 (2023). New York AB A6188 (2025-2026 session).
• Estate of Lokken v. UnitedHealth Group, No. 0:23-cv-03514 (D. Minn.).
• Kisting-Leung v. Cigna Corp., No. 2:23-cv-01477 (E.D. Cal.).
• Barrows v. Humana Inc., No. 3:23-cv-00654 (W.D. Ky.).
• Rucker, Miller, and Armstrong, "How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them," ProPublica, March 25, 2023.
• KFF, analyses of MA prior-authorization data, 2023-2025 reporting years.
• HHS OIG, OEI-09-18-00260.
• California DMHC guidance on SB 1120 implementation, 2025.
• U.S. DOL, EBSA, claims-procedure enforcement guidance.