On October 17, 2024, the US Senate Permanent Subcommittee on Investigations (PSI) released Refusal of Recovery, a majority staff report built from over 280,000 internal pages across UnitedHealthcare, Humana, and CVS Health (US Senate PSI, Oct 17, 2024). As UnitedHealthcare expanded automated post-acute utilization tools including naviHealth's nH Predict, its post-acute prior-authorization denial rate climbed from 10.9% in the 2020 report series to 22.7% in the 2022 report series, while Humana's post-acute denial rate exceeded 16 times its overall medical denial rate. Eighteen months earlier, ProPublica and The Capitol Forum revealed that Cigna medical directors used batch software called PXDX to reject over 300,000 claims in two months, averaging 1.2 seconds per claim without opening patient files (ProPublica, Mar 25, 2023). At a skilled nursing facility desk on Friday afternoon, that machinery arrives as a two-page termination notice giving a stroke patient forty-eight hours to pack.
An AI medical claim denial is vulnerable on appeal whenever an insurer lets predictive length-of-stay targets or batch code filters override individual patient facts. Federal Medicare Advantage rules (42 CFR § 422.101(c) and February 2024 CMS FAQs) and state statutes like California SB 1120 require individualized clinical review by a qualified human clinician. Demand the utilization-review file, document patient-specific barriers, and send our appeal letter and checklist below.
Insurers profit from friction rather than unbeatable legal merits. In the HHS Office of Inspector General's landmark audit (OEI-09-18-00260, n = 250 sampled prior-authorization denials), 13% of the denied requests in the sample already met Medicare coverage rules (HHS OIG, Apr 2022). Across federal Medicare Advantage reporting analyzed by KFF, roughly 82% of appealed prior-authorization denials in CMS program records are overturned in whole or in part, yet only about 11% of denied requests in those records are ever appealed (KFF Medicare Advantage Analysis).
This playbook serves clinic medical billers, skilled nursing facility (SNF) and home-health case managers, patient navigators, and adult children advocating for hospitalized relatives. ed.cafe at No Worker Left Behind is an adult AI-literacy nonprofit, not a law firm, medical provider, or public insurance adjuster; this guide offers general educational information rather than legal or clinical advice. Never miss an expedited appeal clock while gathering extra paperwork. See our wider map in the two meanings of AI safety at work.
How predictive length-of-stay models and batch claim filters generate denials
Algorithm-assisted medical denials flow from three distinct software architectures: post-acute length-of-stay predictors (nH Predict), outpatient code-mismatch batch queues (PXDX), and Medicare's 2026-2031 WISeR prior-authorization pilot.
Knowing which machine touched the claim tells a biller or family which record to demand:
1. Post-acute predictors (Estate of Lokken and Barrows v. Humana)
When an older adult enters a Skilled Nursing Facility (SNF), Inpatient Rehabilitation Facility (IRF), or home health care after a stroke or fracture, vendors run tools like naviHealth's nH Predict or CareCentrix models. Those models match diagnosis codes, age, and admission scores against millions of historical stays to output a target stay, such as 14.0 days.
Class actions in Estate of Gene B. Lokken v. UnitedHealth Group (D. Minn.) and Barrows v. Humana (W.D. Ky.) allege reviewers faced targets keeping stays within 1% of the nH Predict projection, cutting off care even when treating doctors documented patients could not stand or climb home stairs. In 2025, a federal court in Lokken ordered broad discovery into UnitedHealth's AI review boards, staff pay, and nH Predict audits (ArentFox Schiff).
2. Batch claim-rejection queues (Cigna PXDX)
Cigna's PXDX ("procedure-to-diagnosis") tool compared CPT procedure codes against pre-mapped ICD-10 diagnosis lists. Missing pairs queued 50 claims at once so a medical director rejected all 50 in 1.2 seconds per claim without opening charts (ProPublica, Mar 25, 2023); ERISA fiduciary claims later survived Cigna's dismissal motion (InsuranceNewsNet).
3. Original Medicare's 2026-2031 WISeR pilot
From January 1, 2026 through December 31, 2031, CMS runs the Wasteful and Inappropriate Service Reduction (WISeR) pilot in six states: Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington (CMS WISeR). Vendors pair AI/ML screening with human clinicians on targeted Original Medicare Part B outpatient procedures (select spinal, nerve-stimulator, and skin-substitute services).
Picture a bakery manager who pulls every loaf out at 22 minutes because last year's spreadsheet averaged 22 minutes, refusing to open the oven when a dense sourdough loaf is still cold batter inside. That analogy breaks down on who pays for the mistake: a raw loaf hits the bakery trash, whereas an 82-year-old discharged early absorbs the fall risk at home while the insurer keeps unappealed savings. Compare commercial documentation rules in our Massachusetts DOI guide for claims adjusters.
What CMS 42 CFR 422.101(c) and state laws like California SB 1120 require
Federal Medicare regulations and state utilization-management statutes establish one bright legal rule: software may flag or organize a file, but an algorithm may never serve as the sole basis to deny care or override a patient's individual clinical facts.
Put the exact statute matching the patient's plan into paragraph one of your appeal:
- Medicare Advantage (
42 CFR § 422.101(c)& Feb 6, 2024 CMS FAQs): Plans cannot apply proprietary rules stricter than Traditional MedicareNCDsorLCDs(§ 422.101(b)) (eCFR § 422.101). CMS's Feb 2024 FAQs bar algorithms from deciding medical necessity alone or overriding individualized patient facts, and42 CFR § 422.566(d)requires review by a clinician with relevant expertise. - California
SB 1120("Physicians Make Decisions Act", Jan 1, 2025): Health & Safety Code§ 1367.01and Insurance Code§ 10123.135forbid AI from supplanting provider judgment (CDI SB 1120 PDF). Only a licensed, competent human clinician who reviews the clinical record may deny medical necessity. - Commercial ERISA Plans (
29 CFR § 2560.503-1): Entitles enrollees to free copies of all records, internal rules, guidelines, protocols, and rationales generated or relied upon in the denial (eCFR § 2560.503-1). - Massachusetts (
M.G.L. c. 176O& pendingS.2632): Commercial plans owe written clinical criteria and Office of Patient Protection (OPP) external review (c. 176O §§ 12-14), while billS.2632on healthcare AI remained in Senate Ways and Means as of October 29, 2026 (S.2632).
| Coverage Type | Rule Against Algorithmic Cutoff | File & Criteria Demand Right | Urgent Appeal Clock |
|---|---|---|---|
| Medicare Advantage | 42 CFR § 422.101(c), § 422.566(d) + Feb 2024 CMS FAQs: patient facts beat cohort scores | Coverage criteria (§ 422.101(b)) & appeal case file | SNF/HH NOMNC: Noon of day before cutoff (BFCC-QIO). Pre-service: 72 hr |
Original Medicare (WISeR) | WISeR (2026-2031 in AZ, NJ, OH, OK, TX, WA) cannot narrow Part B NCD/LCD rules | MAC medical-review & peer rationales | 3-business-day target; unlimited resubmission with added facts |
| Commercial ERISA | 29 CFR § 2560.503-1: full and fair review of submitted medical records | Free copy of all records, rules, & protocols relied upon | Urgent: 72 hr; Standard: 180 calendar days |
| California Plans | SB 1120 (Jan 1, 2025): only a licensed, competent clinician may deny necessity | DMHC IMR & CDI utilization-file rules | Urgent IMR 72 hr; standard grievance 30 days |
| Massachusetts Plans | M.G.L. c. 176O + DOI Bulletin 2024-02: human clinical review required | Written rationale + OPP external review | Expedited OPP 72 hr; standard 4 months |
Four clues an algorithm shaped the denial and how to build the clinical counter-file
Insurers rarely print "Denied by nH Predict v4.2" on a letter, yet clinic billers and SNF case managers can spot software-driven denials from four recurring tells and counter them with five individualized facts algorithms cannot model.
Watch for four tells in the denial file:
- Benchmark-day SNF cutoffs:
NOMNCforms arriving on round days (Day 7, 10, 14, or 20) despite active PT gains or daily IV/wound care. - Peer-cohort phrasing: Generic lines stating "achieved expected recovery for diagnostic group" while ignoring yesterday's vitals or home stairs.
- Instant batch turnaround: Sub-minute rejections on multi-page claims (
PXDX-style code flags) where attached notes went unopened. - Unnamed or out-of-specialty reviewer: Letters lacking a clinician's license or signed outside the relevant specialty.
First, audit your own clinic's AI-drafted notes. UC Davis Health (n = 356 records) found ambient scribes omitted spoken details in 18.0% of the records (JMIR Med Inform, 2026). If your clinic's scribe dropped six weeks of failed PT or inserted a phantom "gait normal" line, fix the chart first via our 90-second clinician scribe check and HIPAA 45 CFR 164.526 amendment steps.
Algorithms model averages; appeals win on variance from the average:
| Individualized Clinical Dimension | What a Cohort Algorithm Sees | What Your Appeal Packet Must Document |
|---|---|---|
| 1. Functional Gap | I63.9 (stroke) + Day 14 | "Moderate 2-person assist for bed-to-chair transfer; walks 15 ft vs. 150-ft baseline." |
| 2. Home Barriers | ZIP code & marital status | "14 steep stairs, no elevator, narrow bathroom blocking walker; 83-year-old ill spouse." |
| 3. Daily Skilled Need | Standard med list | "Orthostatic BP drops to 84/50 mmHg in PT; daily IV antibiotics or wound vac (42 CFR § 409.31)." |
| 4. Trajectory / Jimmo | Flat statistical curve | Weekly distance gains OR skilled maintenance therapy needed under Jimmo v. Sebelius to prevent decline. |
| 5. Risk Attestation | Routine discharge assumed | Signed MD/PT line: "Discharge on [Date] poses immediate fall and 30-day readmission risk." |
Run a 48-hour SNF cutoff timeline (assumptions stated): At 3:30 PM Thursday, an 81-year-old hip-fracture patient in a Brockton SNF gets a NOMNC ending coverage Saturday. Thursday 4:00 PM: Call the BFCC-QIO number on the notice before noon Friday to freeze financial liability during independent physician review. Thursday 4:30 PM: Fax a demand for NCD/LCD criteria, reviewer credentials, and any nH Predict score. Friday 9:00 AM: Fax a 1-page signed PT/MD variance sheet listing transfer assist, home stairs, and readmission risk under 42 CFR § 422.101(c).
The ed.cafe AI-Assisted Medical Denial Appeal Letter & Record-Request Checklist
The ed.cafe AI-Assisted Medical Denial Appeal Letter & Record-Request Checklist provides an ungated clock triage, file demand, and fill-in appeal letter. Never paste unredacted patient charts into public ChatGPT tabs; fill in this template locally inside your HIPAA-compliant workspace (see what not to paste into an AI tool at work).
PART 1: DEADLINE TRIAGE (CHECK CLOCK FIRST - DO NOT WAIT FOR RECORDS!)
[ ] A. IN SNF, REHAB, OR HOME HEALTH ("NOMNC" Cutoff Notice):
Call BFCC-QIO on notice BY NOON OF DAY BEFORE care ends. QIO #: [_______]
[ ] B. URGENT PRE-SERVICE / PRIOR AUTH (Delay risks health or severe pain):
Request EXPEDITED 72-HOUR APPEAL by phone + fax with physician sign-off.
[ ] C. POST-SERVICE / OUTPATIENT CLAIM DENIAL (Batch Code Rejection):
Calendar plan deadline (60 days Medicare Advantage; 180 days ERISA).
PART 2: 6-ITEM DENIAL-FILE & ALGORITHM DISCLOSURE CHECKLIST
Demand free copies in writing (29 CFR 2560.503-1 / 42 CFR Part 422):
[ ] 1. Name, board specialty, and license # of reviewing human clinician
(42 CFR 422.566(d); CA SB 1120).
[ ] 2. Exact NCD/LCD or public coverage policy applied (42 CFR 422.101(b)).
[ ] 3. Predictive length-of-stay or batch software outputs generated or used
(e.g., nH Predict, PXDX, Milliman, InterQual).
[ ] 4. Timestamped audit log showing whether attached clinical notes were opened.
[ ] 5. Complete utilization-review nurse/physician notes from the claim file.
[ ] 6. Instructions and deadline for Independent External Review / QIO / ALJ.
PART 3: COPY-READY APPEAL LETTER (BILLERS, CASE MANAGERS & FAMILIES)
Date: [MM/DD/YYYY] Sent via: [ ] Fax (Keep Receipt) [ ] Portal [ ] Cert. Mail
To: Appeals & Grievances, [Insurer], Fax/Address: [___________________________]
Patient: [Full Name] DOB: [MM/DD/YYYY] Member ID: [___________]
Claim / Auth #: [_______________] Date(s) of Service: [___________]
Type: [ ] EXPEDITED 72-HOUR APPEAL [ ] STANDARD INTERNAL APPEAL
Re: Appeal of Adverse Determination & Demand for Individualized Clinical Review
To Whom It May Concern:
I am appealing the denial/termination of [Service / SNF Stay / CPT Code] for
[Patient Name]. The denial relies on generalized cohort or code-matching rules
that ignore [Patient Name]'s individualized condition in the medical record.
1. GOVERNING LEGAL STANDARD AGAINST ALGORITHMIC OVERRIDE
- [Medicare Advantage]: Under 42 CFR 422.101(c), 42 CFR 422.566(d), and CMS's
Feb 6, 2024 Part C FAQs, a plan may not use an algorithm or length-of-stay
predictor to override Medicare criteria or individualized patient facts, and
denials require review by a clinician with relevant expertise.
- [Commercial / ERISA / State Plan]: Under 29 CFR 2560.503-1 [and California
SB 1120 (Health & Safety Code 1367.01) / M.G.L. c. 176O], this claim
requires individualized clinical review by a licensed, competent healthcare
professional rather than an automated code filter or cohort benchmark.
2. INDIVIDUALIZED CLINICAL FACTS MISSED BY INITIAL REVIEW
Treating clinician [Dr./PT Name] documents on pages [X-Y] attached:
- Current Status: [e.g., Moderate 2-person transfer assist; walks 15 ft vs.
150-ft baseline; unstable BP / daily IV therapy]
- Home Barriers: [e.g., 14 entry stairs, narrow bath, no safe solo transfer]
- Skilled Need: [Failed conservative care OR daily skilled nursing/PT under
42 CFR 409.31, including maintenance therapy per Jimmo v. Sebelius]
- Harm Risk: Premature cutoff/denial creates immediate risk of [fall /
clinical decompensation / 30-day hospital readmission].
3. FORMAL DEMAND FOR CLAIM FILE & SOFTWARE RECORDS
If not overturned in full, provide free of charge: (a) reviewing clinician's
name, license, and specialty; (b) all internal rules, guidelines, and
protocols used; and (c) any algorithmic, batch-screening, or length-of-stay
outputs (nH Predict, PXDX, etc.) generated on this claim, plus external
review / QIO / ALJ instructions.
Sincerely,
[Signature, Printed Name, Role/Relationship & Callback Phone]
Attachments: [ ] Signed Treating Clinician/PT Note [ ] Auth Representative FormFor navigators helping older adults or low-income households handle automated benefit flags alongside medical denials, pair The ed.cafe AI-Assisted Medical Denial Appeal Letter & Record-Request Checklist with our guide to state AI flags on SNAP cases and the ed.cafe Staff Room.
Where this stops being true
Appeal templates hit legal and clinical boundaries where attorneys, ombudsmen, or physicians must step in:
- Not legal or medical advice: General educational information as of October 29, 2026. Never delay emergency care, and consult Massachusetts SHINE, the Center for Medicare Advocacy, or ERISA counsel on complex denials.
- Never paste charts into public chatbots: Pasting denial files into personal ChatGPT tabs exposes SSNs, Medicare IDs, and HIPAA-protected charts without a BAA (
45 CFR § 164.502(e)). Use local templates only. - Self-funded ERISA preemption: State laws like California
SB 1120and MassachusettsM.G.L. c. 176Obind state-regulated plans; self-funded employer plans generally fall under federal ERISA (29 U.S.C. § 1144). - Statutory exclusions:
42 CFR § 422.101(c)stops algorithmic cutoffs on covered care, not non-covered cosmetic items or custodial care lacking skilled need.
Questions people ask
Can health insurance companies legally use AI to deny my medical claim?
Insurers may use software to organize files or approve requests. However, Medicare Advantage rules (42 CFR § 422.101(c) and Feb 2024 CMS FAQs) and state laws like California SB 1120 forbid using an AI algorithm alone to deny medical necessity without individualized human clinical review.
How do I appeal a Medicare Advantage skilled nursing rehab cutoff driven by nH Predict?
Call the BFCC-QIO number on your NOMNC notice by noon of the day before coverage ends to freeze financial liability and trigger independent physician review. Then fax daily PT/OT transfer scores, home stair counts, and a treating-doctor risk note citing 42 CFR § 422.101(c).
How often do Medicare Advantage prior-authorization appeals succeed?
Federal CMS records analyzed by KFF show roughly 82% of appealed Medicare Advantage prior-authorization denials are overturned partially or fully, and HHS OIG (OEI-09-18-00260, n = 250 denials) found 13% of sampled denials already met Medicare rules. Yet only about 11% of denials are appealed.
Can I force my insurer to tell me if an algorithm reviewed my claim?
Under ERISA rules (29 CFR § 2560.503-1) and Medicare Advantage appeal rights, you can demand free copies of all records, internal rules, guidelines, protocols, and clinical rationales relied upon or generated during your denial, plus the reviewing clinician's specialty credentials.
What is CMS's WISeR model for Original Medicare prior authorization?
From January 1, 2026 through December 31, 2031 in Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington, CMS's WISeR pilot pairs AI screening with human clinicians for prior authorization on select Original Medicare Part B outpatient procedures without narrowing underlying Medicare coverage rules.
Should I upload my hospital records to ChatGPT to write an insurance appeal letter?
No. Uploading hospital records, Medicare IDs, or denial letters into a personal chatbot exposes sensitive health data outside HIPAA protections and risks hallucinated citations. Fill in a local appeal template using exact functional quotes from your clinician's signed notes.
Key terms
42 CFR § 422.101(c): Federal Medicare Advantage rule requiring individualized clinical determinations rather than blanket algorithmic predictions.nH Predict: Post-acute length-of-stay software challenged in Estate of Lokken and examined in the October 2024 US Senate PSI report.- California
SB 1120: California law (Jan 1, 2025) requiring a licensed, competent human clinician, never AI alone, to decide medical-necessity denials. - Notice of Medicare Non-Coverage (
NOMNC): SNF, rehab, or home-health cutoff notice carrying a noon prior-day deadline for fast-trackBFCC-QIOreview. - ERISA Claim-File Rule (
29 CFR § 2560.503-1): Federal rule giving employer-plan enrollees free copies of all records, rules, and protocols used in a denial. - CMS
WISeRModel: Six-state CMS pilot (2026 to 2031) pairing AI screening and human clinical review on select Original Medicare Part B procedures.
Sources
- US Senate PSI, "Refusal of Recovery" (Oct 17, 2024): US Senate PSI
- ProPublica, "How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them" (Mar 25, 2023): ProPublica
- eCFR, "42 CFR § 422.101: Medicare Advantage Basic Benefits" (current): eCFR § 422.101
- HHS OIG, "MA Prior Auth Denials (
OEI-09-18-00260,n = 250)" (Apr 2022): HHS OIG - KFF, "MA Prior Authorization Appeals Analysis" (CMS data): KFF
- CA Dept. of Insurance, "Guidance on SB 1120" (Jan 1, 2025): CDI PDF
- CMS Innovation Center, "WISeR Model (2026-2031)" (current): CMS WISeR
- ArentFox Schiff, "Discovery in UHC AI Lawsuit (Estate of Lokken)" (2025): ArentFox Schiff
- InsuranceNewsNet, "Cigna ERISA Claim Survives Dismissal" (2024): InsuranceNewsNet
- eCFR, "29 CFR § 2560.503-1: ERISA Claims Procedure" (current): eCFR § 2560.503-1
- MA General Court, "Bill S.2632: Healthcare AI" (2025-2026): S.2632
- UC Davis Health / JMIR Med Inform, "Ambient AI Notes (
n = 356)" (2026): JMIR PDF
