Health Insurance AI Claims Processing 2026 Deep-Dive — Cigna, UnitedHealth Optum, Humana, Elevance, CVS Aetna, Oscar + Korean P&C + Japanese Life Insurers

Prologue — The Era of AI Denying Your Claim

In November 2023, U.S. STAT News published an investigation that shook the entire U.S. health insurance industry. UnitedHealth Groups NaviHealth nH Predict algorithm reportedly overturned the clinical judgment of treating physicians 90.5% of the time and forced premature discharge of patients from skilled nursing facilities. The same year, a class action was filed in Minnesota federal court. In 2024 came California class actions against Cignas PXDX auto-denial system. In 2025, Humana faced its own algorithmic denial lawsuit. A 2024 U.S. Senate Permanent Subcommittee on Investigations (PSI) report concluded that the Big Three Medicare Advantage insurers denied prior authorization at roughly 16x the rate of traditional Medicare.

The exact same technology operates in the opposite direction too. The U.S. NHCAA (National Health Care Anti-Fraud Association) estimates conservative annual losses to healthcare fraud at $100B and broader estimates at$300B. You cannot catch that without AI. Optum, Cotiviti, SAS, Shift Technology, and ClarisHealth run LLMs and GBMs (gradient boosting machines) over ICD-10/CPT patterns, claim text, prescription sequences, and provider behavior graphs to flag fraud.

Korea and Japan sit on the same axis. Koreas October 2023 amendment to the Insurance Business Act enabled actual-loss insurance claim simplification, which rolled out in October 2024 (Stage 1: hospitals 30+ beds), October 2025 (Stage 2: clinics and pharmacies), and entered full operational mode in 2026. Samsung Fire & Marine, DB Insurance, Hyundai Marine, KB Insurance, and Meritz all operate proprietary AI auto-adjudication engines. Japan rides on top of MHLW (厚労省)s online qualification verification (オンライン資格確認) and the Mynanumber Health Card integration, where SOMPO, Dai-ichi Life, and Nichii are deep in claim automation plus healthcare data utilization.

This article lays out the entire terrain: automated claim adjudication, prior authorization automation, fraud detection, drug interactions, EMR/EHR linkage, and the social conflicts those create.

1. Anatomy of a Health Insurance Claim — The Stages

A single U.S. commercial health insurance claim typically passes through these stages.

The key insight: between Stage 5 and Stage 6 is where AI auto-denial bites hardest. And fewer than 1% of denials are appealed, so flipping even a fraction of that 1% leaves automation statistically in the black.

2. ICD-10-CM, CPT, HCPCS — The Alphabet of Claim Coding

To understand the AI, you must know the codes.

• ICD-10-CM (International Classification of Diseases, 10th, Clinical Modification): Diagnosis codes. The U.S. transitioned from ICD-9 in 2015. Around 70,000 codes. Example: E11.9 = Type 2 diabetes without complications.
• CPT (Current Procedural Terminology): Procedure codes maintained by the AMA. About 10,000. Example: 99213 = office visit, established patient, level 3.
• HCPCS Level II: CMS-maintained supplementary codes (DME, drugs, non-physician services). Example: J0696 = Injection, ceftriaxone sodium, per 250 mg.
• NDC (National Drug Code): FDA drug identifier, 11 digits.
• LOINC: Lab result identifiers. Example: 2345-7 = Glucose mass/volume in serum or plasma.
• SNOMED CT: Clinical terminology ontology, more granular than ICD-10.

In 2026 the U.S. uses ICD-10-CM 2026 with ongoing debate on ICD-11 adoption. WHO ICD-11 became effective January 2022, but CMS has not finalized a U.S. transition schedule. Korea uses KCD-8 (2021 edition). Japan adopts ICD-10 2013 edition via MHLW.

3. CAC (Computer-Assisted Coding) — Coding Goes LLM

Pulling ICD-10/CPT codes from charts (clinical notes, lab results, radiology reads) used to be done by certified coders. Computer-Assisted Coding (CAC) automated parts of this, and since the late 2020s LLMs have rapidly displaced traditional CAC.

# CAC LLM pseudo-code (generalizes Optum LifeCode, 3M 360 Encompass, etc.)
from langchain.prompts import PromptTemplate
from langchain_anthropic import ChatAnthropic

TEMPLATE = """
You are a certified medical coder (CPC-A) using ICD-10-CM 2026 and CPT 2026.
Analyze the following clinical note and extract ALL applicable codes.

Clinical Note:
{note}

Output JSON:
{{
  "diagnosis_codes": [{{"icd10": "...", "rationale": "..."}}],
  "procedure_codes": [{{"cpt": "...", "modifiers": [...], "rationale": "..."}}],
  "hcpcs": [...],
  "evaluation_management": {{"level": 1|2|3|4|5, "rationale": "..."}}
}}

CRITICAL: Do NOT upcode. If ambiguous, choose lower specificity and flag for human review.
"""

def code_encounter(note: str) -> dict:
    chain = PromptTemplate.from_template(TEMPLATE) | ChatAnthropic(model="claude-opus-4")
    return chain.invoke({"note": note})

The crucial constraint: prevent upcoding. If AI inadvertently inflates codes, hospitals can be hit with massive False Claims Act settlements (in 2023 DOJ recovered $2.7B from Medicare fraud alone). So payers in turn run downcoding AI on incoming claims — which is the next section.

4. Cigna PXDX — The 1.2-Second Auto-Denial

Cignas PXDX (Procedure-to-Diagnosis) system, exposed by ProPublica in 2023, was the ignition point for the U.S. health insurance AI debate. Internal documents obtained by ProPublica showed Cigna medical directors denied claims at an average pace of 1.2 seconds — without reading charts. PXDX automatically recommended denial if a claims diagnosis-procedure code pair was not on Cignas pre-defined "allowed matrix," with doctors batch-signing the denials.

# Simplified illustration of PXDX-style matrices (actual matrices are confidential)
Diagnosis ICD-10     | Allowed CPT                | Denied CPT
---------------------|----------------------------|-----------------------------
M54.5 (back pain)    | 99213, 97110, 97140        | 72148 (MRI lumbar) — requires 6wk conservative care
J02.9 (sore throat)  | 99213, 87880               | 87070 (culture) — denied if Centor < 3
F32.9 (depression)   | 90834, 90837               | 90791 (initial eval) — denied if more than 2/yr

The California class action (Kisting-Leung v. Cigna, July 2023) is in active litigation and entered the class certification stage in 2025. Cigna defends PXDX as "merely a sanity check on code pairs, not a medical judgment," while plaintiffs argue that "denying claims without medical judgment is itself a California Insurance Code violation."

5. UnitedHealth NaviHealth nH Predict — Shortening SNF Stays

The 2023 STAT News series by Casey Ross and Bob Herman exposed UnitedHealths nH Predict algorithm. nH Predict produces predicted length-of-stay numbers for Medicare Advantage members moving from acute hospitals to skilled nursing facilities (SNFs). Two problems emerged.

1. Model accuracy: Internal data obtained by STAT showed that when the model predictions clashed with treating physicians and appeals were filed, 90.5% of appeals were sustained — meaning the model was wrong 90.5% of the time.
2. How it was used: UnitedHealth case managers were instructed (per STAT reporting) to recommend SNF discharge within 1% of the model-predicted date.

A class action (Estate of Lokken v. UnitedHealthcare) was filed in Minnesota federal court in November 2023. The 2024 Senate PSI report concluded that UnitedHealth, Humana, and CVS/Aetna denied prior authorization at 16x the rate of traditional Medicare. As of 2026 NaviHealth has been reorganized into Optum Health Solutions and nH Predict has been repositioned as "decision support tooling."

6. Optum AI — UnitedHealth Groups Brain

UnitedHealth Group splits into the insurance arm (UnitedHealthcare) and the healthcare services arm (Optum). By 2026 Optums revenue is more than half of total group revenue, and essentially all AI and data capabilities are concentrated inside Optum.

Optums LLM infrastructure in 2026 is reportedly a hybrid of self-trained clinical-domain models plus external GPT-4/Claude API calls. After the December 2024 fatal shooting of UnitedHealthcare CEO Brian Thompson, the company stood up an AI governance committee and announced certain policy changes including mandatory human review of denial rationales.

7. Humana — The Big Three Medicare Advantage Player

Humana is the #2 Medicare Advantage carrier (around 17% share in 2026), heavily concentrated in MA. In December 2025 Humana also faced a class action (Barrows v. Humana) alleging similar algorithmic SNF-shortening.

Humana announced expansion of its in-house AI coding subsidiary CenterWell and the MORE Health AI infrastructure in September 2023. By 2026 Humanas claim auto-adjudication runs a three-stage pipeline:

[Inbound 835/837 EDI]
    ↓
[Stage 1] Rules engine — Medicare/CMS policy + Humana policy
    ↓ (denial candidates extracted)
[Stage 2] GBM classifier — fraud probability score
    ↓ (score > threshold → human review)
[Stage 3] LLM chart review — pulls chart via Epic/Cerner FHIR, assesses medical necessity
    ↓
[Pay / Deny / Request documents]

Humana does not publish AI denial rates, but BCBS Association estimates an industry average around 16-20% of claims initially denied (with some recovered via appeal/resubmission).

8. Elevance Health — Why Anthem Rebranded

Anthem rebranded as Elevance Health in June 2022. The reason was simple — "we want to look like a healthcare company, not an insurer." Elevance owns Anthem Blue Cross Blue Shield, Carelon (health services, pharmacy, behavioral), Wellpoint (Medicaid), and others.

Elevances AI strategy centers on Carelon.

• Carelon Insights — claim analytics, predictive modeling, cost forecasting
• Carelon Health — primary care clinics and home health with AI triage
• Carelon Rx — PBM with AI prescription review
• Carelon Behavioral Health — AI screening (PHQ-9, GAD-7 auto-scoring)

The June 2025 Senate PSI report alleged that Anthem automated prior authorization denials and artificially inflated denial rates. Elevance pushed back, arguing "AI is constrained to decision support only."

9. CVS Health / Aetna — Vertical Integration of Pharmacy + Insurer + Clinic

CVS acquired Aetna in 2018 for $69B, then added Oak Street Health (primary care) and Signify Health (home care) in 2023. By 2026 CVS Health is essentially the only U.S. company that owns pharmacy, PBM, insurer, and primary care under one roof.

CVS AI operates on four axes.

In 2024 CVS deployed its in-house LLM assistant "Connect" to brief call-center agents on per-member recommendations. In 2025 Aetna opened a prior authorization automation API to providers — "if AI recommends denial, the appeal pathway is presented in the same call" — a bidirectional model now in pilot.

10. Oscar Health — Differentiating via AI Navigator

Oscar Health, founded in 2012, is a digital-first health insurer with 2026 footprint across 18 states in the ACA marketplace. It is small compared to the Big Five (about 1M members) but its AI strategy goes the opposite direction — member navigation, not denial.

Oscars flagship product is the Oscar Care Navigator. When a member tells the chatbot "I have had a headache for a week," the AI handles (1) triage on ER vs urgent care vs PCP, (2) in-network provider recommendations, (3) immediate telehealth booking, and (4) automatic prior-auth pre-check. Oscar announced an OpenAI GPT-4 based assistant in 2024 and as of 2026 also uses Anthropic Claude in select workflows.

// Oscar Care Navigator pseudo-code
interface PatientQuery {
  member_id: string
  symptoms: string  // natural language
  history?: string[]  // recent ICD-10 codes
}

async function navigate(q: PatientQuery): Promise<NavigationResult> {
  // 1. Triage LLM
  const triage = await llm.chat({
    system: "You are a triage nurse. Output: ER | URGENT_CARE | PCP | TELEHEALTH | SELF_CARE.",
    user: `Symptoms: ${q.symptoms}. History: ${q.history?.join(", ")}`,
  })

  // 2. Network match
  const providers = await provider_directory.search({
    member_id: q.member_id,
    specialty: triage.recommended_specialty,
    in_network: true,
    sort_by: "distance",
  })

  // 3. Prior auth pre-check
  const pa_required = await prior_auth.check({
    member_id: q.member_id,
    suspected_codes: triage.likely_cpt_codes,
  })

  return { triage, providers, pa_required }
}

Oscars bet is clear — instead of cutting costs by denying claims, route members to the right care from the start. Their medical loss ratio (MLR) is higher than the Big Five but NPS in the 50+ range crushes the industry average (NPS around 8).

11. Prior Authorization Automation — The Core of AI Denials

Prior authorization is the payers way of asking "do you really need that?" before a procedure. In the U.S. roughly 18% of services require prior auth (KFF 2024), and the rate is higher in Medicare Advantage.

The CMS 2024 final rule (CMS-0057-F) requires every Medicare Advantage/Medicaid plan to expose a FHIR Prior Authorization API, and from January 2027 to issue PA decisions within 7 days for standard and 72 hours for urgent. That is the accelerator pedal for AI automation.

POST /Bundle HTTP/1.1
Host: payer.example.com
Content-Type: application/fhir+json
Authorization: Bearer eyJhbGc...

{
  "resourceType": "Bundle",
  "type": "transaction",
  "entry": [
    {
      "resource": {
        "resourceType": "Claim",
        "status": "active",
        "type": { "coding": [{ "system": "http://terminology.hl7.org/CodeSystem/claim-type", "code": "professional" }]},
        "use": "preauthorization",
        "patient": { "reference": "Patient/12345" },
        "diagnosis": [
          { "diagnosisCodeableConcept": { "coding": [{ "system": "http://hl7.org/fhir/sid/icd-10-cm", "code": "M54.5" }]} }
        ],
        "item": [
          {
            "sequence": 1,
            "productOrService": { "coding": [{ "system": "http://www.ama-assn.org/go/cpt", "code": "72148" }]}
          }
        ]
      }
    }
  ]
}

That request is met with an AI auto-response — rules plus GBM plus, if needed, LLM chart pull.

12. Medical Necessity — AI Mimicking Clinical Judgment

"Medical necessity" is the central clause in U.S. insurance contracts. If a service is not medically necessary, the payer has no obligation to pay. The fight is over who decides medical necessity.

Traditionally InterQual (Change Healthcare) and MCG (Milliman Care Guidelines) are the two dominant U.S. clinical criteria. By 2026 payers stack LLMs on top — "does the chart satisfy InterQual criteria?"

# Medical necessity auto-assessment pseudo-code
def assess_medical_necessity(claim, ehr_chart):
    # 1. Look up InterQual criteria
    criteria = interqual.lookup(
        cpt=claim.procedure_codes[0],
        icd10=claim.diagnosis_codes[0],
    )

    # 2. Extract evidence from EHR chart
    evidence = llm.extract(
        prompt=f"For criteria {criteria.text}, extract supporting evidence from chart.",
        chart=ehr_chart,
    )

    # 3. Evaluate satisfaction
    assessment = llm.evaluate(
        criteria=criteria.text,
        evidence=evidence,
        prompt="Does the evidence satisfy ALL criteria? Output: YES | NO | UNCLEAR. Cite each criterion.",
    )

    return assessment

The problem: InterQual/MCG are paid licenses with per-payer customizations. Providers do not know which criteria they are being judged against, and AI mediation deepens that opacity — which is exactly the central allegation in PXDX/nH Predict lawsuits.

13. Healthcare Fraud Detection — Behind the NHCAA $300B Number

NHCAA estimates 3-10% of U.S. healthcare spend leaks to fraud annually. With 2026 U.S. healthcare spending around $5T, that pegs fraud at$150B-$500B. The AI catching it lives in payer SIUs (Special Investigations Units) plus vendor stacks (Cotiviti, SAS Fraud Framework, Shift Technology, ClarisHealth, Optum Investigative Services).

Major patterns.

# GBM-based fraud detection pseudo-code
import lightgbm as lgb

features = [
    "claim_amount", "provider_avg_claim", "provider_claim_count_30d",
    "diagnosis_procedure_match_score", "patient_provider_distance_km",
    "provider_em_distribution_kl", "ndc_unusual_for_diagnosis_flag",
    "patient_concurrent_claims_count", "billing_time_outlier_zscore",
]

model = lgb.Booster(model_file="fraud_v2026_05.txt")
score = model.predict(claim.features_vector(features))[0]
# score > 0.92 → SIU priority queue
# 0.7 < score <= 0.92 → human review queue
# score <= 0.7 → auto-pay

Beyond GBMs, graph neural networks (GNNs) learning provider-patient-pharmacy-DME co-occurrence graphs are the 2026 frontier. Optum Investigative Services is reportedly a leader and recovers billions annually (per public disclosures).

14. Drug Interaction Review — RxNorm + FHIR + LLM

Prescription drug claims are handled by PBMs (Pharmacy Benefit Managers). The U.S. Big Three are CVS Caremark, Express Scripts (under Cigna), and OptumRx (under UnitedHealth). PBM AI does two things.

1. DUR (Drug Utilization Review): Does a new prescription conflict with existing meds?
2. Substitution: Is there a cheaper or more effective equivalent?

Data standards.

• RxNorm (NLM): Drug naming standard. Example: rxcui 197361 = Acetaminophen 325 MG Oral Tablet.
• NDC: FDA drug code, 11 digits.
• First Databank, Wolters Kluwer, IBM Micromedex: Commercial drug databases (interactions, contraindications, pregnancy, pediatric).

# Drug interaction check pseudo-code
def check_interactions(new_rx_rxcui, member_id):
    # 1. Pull active prescriptions (FHIR R5 MedicationRequest)
    active = fhir.search("MedicationRequest", patient=member_id, status="active")

    # 2. Normalize to RxNorm
    active_rxcuis = [normalize_to_rxcui(m) for m in active]

    # 3. Look up drug interactions (Wolters Kluwer/FDB)
    interactions = []
    for rxcui in active_rxcuis:
        x = fdb.interaction(new_rx_rxcui, rxcui)
        if x and x.severity in ["major", "contraindicated"]:
            interactions.append(x)

    # 4. LLM clinical recommendation
    if interactions:
        recommendation = llm.chat(
            system="You are a pharmacist. Given interactions, suggest action.",
            user=f"Interactions: {interactions}. New Rx: {new_rx_rxcui}.",
        )
        return {"action": "PHARMACIST_REVIEW", "details": recommendation}
    return {"action": "AUTO_APPROVE"}

When DUR works, adverse drug events causing hospitalization drop, so PBMs save not only on drug cost but also on downstream admissions. In 2024 the FDAs Sentinel Initiative began using PBM data for adverse-event signal monitoring.

15. FHIR R5 Claim Resource — The Standard Claim Data Model

HL7 FHIR (Fast Healthcare Interoperability Resources) reached R5 in 2024. CMS and ONC mandate FHIR R4 with phased R5 adoption. The Claim resource standard layout:

resourceType: Claim
id: claim-2026-12345
status: active
type:
  coding:
    - system: http://terminology.hl7.org/CodeSystem/claim-type
      code: professional
use: claim
patient:
  reference: Patient/12345
created: '2026-05-25T10:30:00Z'
insurer:
  reference: Organization/cigna
provider:
  reference: Practitioner/dr-jones
priority:
  coding:
    - system: http://terminology.hl7.org/CodeSystem/processpriority
      code: normal
diagnosis:
  - sequence: 1
    diagnosisCodeableConcept:
      coding:
        - system: http://hl7.org/fhir/sid/icd-10-cm
          code: E11.9
          display: Type 2 diabetes mellitus without complications
    type:
      - coding:
          - system: http://terminology.hl7.org/CodeSystem/ex-diagnosistype
            code: principal
item:
  - sequence: 1
    productOrService:
      coding:
        - system: http://www.ama-assn.org/go/cpt
          code: '99213'
    servicedDate: '2026-05-25'
    unitPrice:
      value: 165.00
      currency: USD
    net:
      value: 165.00
      currency: USD
total:
  value: 165.00
  currency: USD

CMS-0057-F mandates that prior authorization run on the FHIR-based PAS (Prior Authorization Support) IG. By 2026 Epic, Cerner (Oracle Health), and Athenahealth all support PAS IG.

16. EMR/EHR Integration — Epic, Cerner, Athenahealth

Over 80% of U.S. hospitals run Epic or Cerner (now Oracle Health) EHRs. Payers needing to adjudicate claims need charts, and to pull charts they need EHR connectivity. Three paths exist.

# SMART on FHIR pulling chart from Epic pseudo-code
from fhirclient import client

settings = {
    "app_id": "payer-app",
    "api_base": "https://fhir.epic.com/interconnect-fhir-oauth/api/FHIR/R4/",
}
smart = client.FHIRClient(settings=settings)
smart.authorize()  # OAuth2 backend services flow

# Pull patient
from fhirclient.models.patient import Patient
patient = Patient.read("12345", smart.server)

# Observations (labs)
from fhirclient.models.observation import Observation
search = Observation.where(struct={"patient": "12345", "category": "laboratory"})
obs = search.perform_resources(smart.server)

# Clinical notes
from fhirclient.models.documentreference import DocumentReference
docs = DocumentReference.where(struct={"patient": "12345"}).perform_resources(smart.server)

The 2024 ONC USCDI v4 and HTI-1 rules require EHRs to expose FHIR R4 APIs. By 2026 there is a standard legal pathway for payer AI to pull charts.

17. Korea Actual-Loss Claim Simplification — Fourth-Generation Goes Mainstream

Korea amended the Insurance Business Act in October 2023 to establish a legal basis for actual-loss insurance claim simplification, with a staged rollout.

• October 1, 2024: Stage 1 for hospitals with 30+ beds (about 4,200 facilities)
• October 25, 2025: Stage 2 expansion to clinics and pharmacies (about 70,000 facilities)
• 2026: Full operational phase — claim simplification effectively becomes the default channel

Subscribers, after care, just say "please file my actual-loss claim." The providers EMR sends claim data through the Korea Insurance Development Institute (KIDI)s clearinghouse to each insurer, which then auto-adjudicates and pays the next day. For this to work, AI auto-adjudication on the payer side is essential.

[Hospital/Clinic EMR]
    ↓ (KIDI standard interface)
[Clearinghouse — KIDI]
    ↓
[Each Insurer's Claim System]
    ↓
[AI Auto-Adjudication — rules + GBM + LLM chart review]
    ↓
[Payment T+1 — avg claim around KRW 80,000]

Fourth-generation actual-loss insurance (launched July 2021) raised non-covered self-pay to 30%, separated premium into covered/non-covered tracks, and put caps on over-utilized services like manual therapy and MRI. AI adjudication had to get more sophisticated under 4G — it must track split billing of non-covered items.

18. Samsung Fire & Marine — Korea P&C #1s AI

Samsung Fire & Marine led Korean P&C with around 29% share in 2026. Auto, long-term personal, health, comprehensive, and general lines all run AI auto-adjudication.

Samsung Fire announced an in-house claim AI platform in 2024 and as of 2025 runs OCR + LLM automatic interpretation of medical bills and diagnostic statements. Medical opinions are handled by external bodies (e.g. Korea Medical Opinion Support Center), with AI in a supporting role.

# Korean actual-loss AI auto-adjudication pseudo-code
def adjudicate_kr_health_claim(claim):
    # 1. Basic validation
    if not validate_kcd_code(claim.diagnosis_kcd):
        return reject("Invalid KCD code")

    # 2. Coverage mapping
    coverage = match_coverage(
        product=claim.product_code,
        diagnosis=claim.diagnosis_kcd,
        treatment_type=claim.treatment_type,  # inpatient/outpatient/visit
    )

    # 3. Self-pay calculation (varies by generation 1-4)
    deductible = calc_deductible(
        generation=claim.generation,  # 4G = since 2021.07
        coverage_type=coverage.type,
        bill_total=claim.bill_total,
        non_covered_amount=claim.non_covered,  # non-covered
    )

    # 4. Fraud/overuse score (GBM)
    fraud_score = gbm.predict(claim.features())

    # 5. Routing
    payable = claim.bill_total - deductible
    if fraud_score < 0.3 and payable <= 100_000:
        return auto_approve(payable)  # T+1 payment
    elif fraud_score > 0.7:
        return route_to_siu()
    else:
        return route_to_human_reviewer()

19. Hyundai Marine GoodAndGoodHi, KB Insurance, Meritz AI

The Big Five Korean P&C AI lineup.

Hyundai Marines GoodAndGoodHi differentiates premiums via health checkup results (NHIS data linkage with consent) — good results yield a discount. AI handles checkup data classification and risk scoring.

KB Insurances KB Health Plus app lets members snap photos of prescriptions/diagnostic statements, OCR them to text, and have an LLM auto-draft the claim. Since 2025 it runs in parallel with the simplification channel.

Meritz markets "payment within 24 hours," achievable only with 70%+ auto-adjudication.

20. Japan SOMPO, Dai-ichi Life himawari, Nichii

Japans average patient cost-share is about 30% (general under-75) — lower than Korea — so private medical insurance (医療保険) plays a supplementary role. AI claim processing and healthcare data usage are spreading fast nonetheless.

• SOMPO Japan: In 2024 consolidated healthcare subsidiaries into SOMPO Health Support. AI care management and elder-care data utilization.
• Dai-ichi Life — himawari Life Insurance: Since 2017 the himawari chatbot has automated claim filing and contract changes. In 2025 the LLM assistant was upgraded (GPT-4o + in-house domain tuning).
• Tokio Marine & Nichido — Medical Kick AI: OCR + auto-adjudication of medical receipts.
• Dai-ichi Life — Kenshin (health checkup) data linkage: Risk assessment from checkup results plus Nichii Gakkans home-care visit data.
• Nichii Gakkan: Japans largest medical-administrative and elder-care provider. Aggressive AI/OCR deployment in claim BPO.

In 2024 MHLW launched the My Number health insurance card in full operation, and online qualification verification adoption exceeded 90% of medical institutions. On top of that infrastructure, private insurer claim automation is accelerating.

21. NHS England AI — Public Insurance AI

NHS England is a completely different model from U.S. private insurance, but AI adoption is active. The 2025 NHS AI Lab runs four programs.

1. NHS AI Diagnostic Fund — £23M for imaging AI (chest X-ray, mammography, retina, etc.).
2. NHS Federated Data Platform (FDP) — Palantir won a £330M contract in November 2023 to unify patient flow, inventory, and operations data. AI analytics is central.
3. AI Stroke Pathway — Brainomix e-Stroke and similar imaging AI deployed across nearly all English emergency departments.
4. Doctor AI Co-Pilot Pilots — LLM assistants for GPs at select trusts.

NHS AI does not exist to deny claims — it exists for resource allocation and diagnostic support, because the U.K. is a single-payer system. A nice example of the same technology used differently under different incentives.

22. The Politics of AI Denial — U.S. Senate Hearings and Regulation

In July 2024 the U.S. Senate PSI held "Medicare Advantage and Patient Care: Algorithmic Denials" and put UnitedHealth, Humana, and CVS/Aetna executives under oath. Core findings.

• Big Three MA insurers denied PA at 16x traditional Medicares rate.
• 80%+ of denials were overturned on appeal — meaning 80% of initial denials were inappropriate.
• Some insurers misrepresented their use of AI systems in disclosures.

The January 2024 CMS Final Rule (CMS-0057-F) required, for prior authorization decisions:

• Auto-denial must be on medical grounds, and algorithms cannot decide alone.
• PA decisions must be communicated within standard timelines (7 days standard, 72h urgent).
• Clear explanation of denial rationale is mandatory.

The December 2024 killing of the UnitedHealthcare CEO marked the peak of public anger. The incident itself was tragic, but the social backlash had been accumulating for years. In 2025 insurers responded with AI governance committees, expanded human review requirements, and broader denial-rationale disclosure.

23. The Providers Counter-Attack — DocAI, PA Helper

Providers are not standing still. Opposite the prior-auth automation wave, a parallel wave of "AI for drafting and appealing prior authorizations" has appeared.

• Cohere Health — provider-side prior auth workflow platform.
• Notable Health — EHR-embedded prior-auth automation.
• AKASA — full RCM (Revenue Cycle Management) automation.
• Doximity GPT — physician LLM built into Doximitys social network.
• Glass Health, Hippocratic AI — clinical LLMs.

# Provider-side prior auth auto-draft pseudo-code
def auto_draft_prior_auth(patient_id, requested_cpt):
    # 1. Pull clinical context from EHR
    chart = epic.fetch_chart(patient_id)

    # 2. Look up payer policy (insurer-published criteria or InterQual mapping)
    payer_policy = payer_policies.lookup(patient.payer, requested_cpt)

    # 3. LLM drafts prior-auth justification
    draft = llm.chat(
        system=f"You are a clinical writer. Draft prior auth justification per {payer_policy.criteria}.",
        user=f"Patient chart: {chart}. Requested CPT: {requested_cpt}.",
    )

    # 4. Submit after physician review
    return draft

This is an arms race. Payers auto-deny via AI, providers auto-draft stronger filings via AI, and payers then train stronger denial models on the new evidence — and so on.

24. Korea/Japan/U.S. Comparison — The AI Claim Automation Map

25. 2026 What Comes Next — Whats Coming

• Mandatory prior auth API: When CMS-0057-F goes live in January 2027, the PAS FHIR IG becomes standard and payer AI will get faster and more transparent (or it must).
• Mandatory algorithmic audits: NIST AI RMF, EU AI Act, and state DOI rules may converge to force third-party audits of adjudication AI.
• Patient-side appeal AI: As free appeal-drafting tools (e.g. Counterforce Health) proliferate, the cost structure of deny-appeal-redeny shifts.
• Health data marketplaces: Koreas MyData, Japans My Number, and the U.S. HIPAA Right of Access combined may let patients share their own data with non-insurer AIs for better underwriting.
• EHR-embedded AI assistants: Epics ED Brain and Cerner (Oracle Health)s Clinical AI Assistant generalize and coding/claims become nearly complete at the point of care.
• Price transparency: The U.S. No Surprises Act and Hospital Price Transparency rules make charges visible, enabling patient-side AI for cost estimation.

The central question: whose side is AI on — payers or patients? The 2026 answer is undeniably "payers, so far." But the same LLM tech is diffusing rapidly to patients, providers, and regulators. The next five years hinge on rebalancing this asymmetry.

26. References

• ProPublica, "How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them" (Patrick Rucker, David Armstrong, Doris Burke, 2023): https://www.propublica.org/article/cigna-pxdx-medical-health-insurance-rejection-claims
• STAT News, "UnitedHealth used secret rules to restrict rehab care for seriously ill Medicare Advantage patients" (Casey Ross, Bob Herman, 2023): https://www.statnews.com/2023/11/14/unitedhealth-medicare-advantage-investigation-prior-authorization-naviHealth/
• U.S. Senate Permanent Subcommittee on Investigations (PSI), "Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care" (2024): https://www.hsgac.senate.gov/subcommittees/investigations/
• CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F): https://www.cms.gov/newsroom/fact-sheets/cms-interoperability-and-prior-authorization-final-rule-cms-0057-f
• HHS Office of Inspector General, "Medicare Advantage Compliance Audits" (2024): https://oig.hhs.gov/oei/medicare-advantage/
• HL7 FHIR R5 Claim Resource: https://hl7.org/fhir/R5/claim.html
• DaVinci PAS (Prior Authorization Support) Implementation Guide: https://hl7.org/fhir/us/davinci-pas/
• NHCAA (National Health Care Anti-Fraud Association): https://www.nhcaa.org/
• Kaiser Family Foundation, "Use of Prior Authorization in Medicare Advantage" (2024): https://www.kff.org/medicare/issue-brief/use-of-prior-authorization-in-medicare-advantage-exceeded-46-million-requests-in-2022/
• Korea Insurance Development Institute (KIDI) Actual-Loss Claim Simplification Guide: https://www.kidi.or.kr/
• Korea Financial Supervisory Service (FSS), "Actual-Loss Claim Simplification Implementation Guide": https://www.fss.or.kr/
• Korea Insurance Research Institute (KIRI) auto-adjudication reports: https://www.kiri.or.kr/
• MHLW (Japan) Online Qualification Verification: https://www.mhlw.go.jp/stf/seisakunitsuite/bunya/0000172019.html
• MHLW (Japan) My Number Health Insurance Card: https://www.mhlw.go.jp/stf/index_16745.html
• General Insurance Rating Organization of Japan (GIROJ): https://www.giroj.or.jp/
• General Insurance Association of Japan: https://www.sonpo.or.jp/
• Dai-ichi Life himawari chatbot: https://www.dai-ichi-life.co.jp/
• Nichii Gakkan: https://www.nichiigakkan.co.jp/
• NHS England AI Diagnostic Fund: https://www.england.nhs.uk/aac/what-we-do/innovation-for-healthcare-inequalities-programme/ai-diagnostic-fund/
• NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
• Cohere Health Prior Authorization Platform: https://coherehealth.com/
• Cotiviti Fraud, Waste, and Abuse Solutions: https://www.cotiviti.com/
• Optum Insight: https://www.optum.com/business/insights.html
• AMA CPT Code Set: https://www.ama-assn.org/practice-management/cpt
• CDC ICD-10-CM: https://www.cdc.gov/nchs/icd/icd-10-cm.htm
• WHO ICD-11: https://icd.who.int/en
• NLM RxNorm: https://www.nlm.nih.gov/research/umls/rxnorm/
• Epic FHIR Documentation: https://fhir.epic.com/
• Oracle Health (Cerner) FHIR APIs: https://docs.oracle.com/en/industries/health/millennium-platform-apis/index.html