💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Prologue — Five Years After Lemonades IPO, and the K-ICS Era

When Lemonade listed on NYSE in July 2020, Wall Street had two takes: "Does an insurer deserve a SaaS multiple?" and "Insurance is now software." Six years later, in 2026, that question has tipped toward an answer. **The core logic of underwriting and claims has shifted to algorithms.**

In the US, Lemonade, Root, Hippo, Ladder, Bestow, and Ethos created their own category. Policygenius, The Zebra, Insurify, and Gabi own the comparison marketplace. Legacy carriers are not standing still — AIG, Allstate, State Farm, and Geico built proprietary ML pricing engines, while MetLife and Prudential use ML to accelerate medical underwriting.

In Korea, the K-ICS (Korean Insurance Capital Standard) went into effect in 2023, requiring insurers to mark assets and liabilities to market. Capital efficiency now depends on the speed and accuracy of underwriting, so Samsung Life, Kyobo Life, and Hanwha Life adopted AI underwriting. On the P&C side, KB Insurance and Hyundai Marine deploy AI fraud detection; Carrot Insurance runs telematics-based per-mile auto coverage.

Japan, after Solvency II compatibility work, is rolling out a Japanese version of ICS (Insurance Capital Standard) in stages between 2025 and 2027. Representative cases include Dai-ichi Lifes himawari AI, Nippon Lifes AI underwriting, LIFENETs internet-only term coverage, and SBI Insurances auto fraud detection.

This article draws that map: what underwriting is, why AI took over, which algorithms power it, how we validate, and how we regulate.

Chapter 1 · Traditional vs AI Underwriting

The traditional underwriting process was slow. A single term life policy typically took 4–6 weeks from application to issue, because the underwriter manually reviewed the application, paramedical exam, income evidence, credit report, MIB (Medical Information Bureau) record, and a prescription database such as Milliman IntelliScript.

AI underwriting changed the flow on two axes.

| Dimension | Traditional | AI |

| --- | --- | --- |

| Time | Term life 4–6 weeks, auto 1–3 days | Term life: instant to 72h, auto: instant |

| Data | Application + paramedical + MIB + interviews | Application + telematics, IoT, medical APIs, social signals |

| Algorithm | Rules + underwriter judgment | GBM/XGBoost, neural nets + rule guardrails |

| Claims | Manual adjusters | NLP + CV triage, partial instant pay |

| Pricing | Age/sex/risk-class averages | Individualized, dynamic |

| Fraud detection | Post-hoc adjusters + SIU | Real-time model scoring |

| Regulation | Rules in writing | NAIC AI Bulletin, EU AI Act, national guides |

The pivot is **the breadth and freshness of data**. Traditional carriers rely on static data at application time; AI insurers exploit time-series signals from telematics, IoT, and wearables. Pricing becomes not just sharper but a behavioral incentive — "safe driving lowers your premium" is now part of the pricing model.

Chapter 2 · Lemonade — Maya, Jim, and `<3초` Claim Settlements

Lemonade launched in New York in 2015. Two pillars: **chatbot underwriting** and a **fixed-fee P2P model**.

- Maya: the application chatbot. Goal: finish onboarding in 30 seconds. It asks about residence, asset value, prior coverage, and quotes instantly.

- Jim: the claims chatbot. Receives the loss notice, verifies photos, documents, and statements with AI, and pays automatically. The fastest reported settlement: `<3초` (under three seconds).

The P2P model keeps a fixed 25% of premium as fee; the rest funds claims, and what is left over goes to charity (Giveback). The structure **dampens the incentive to deny claims**.

The core models concentrate on fraud detection. AI Jim extracts 18 antifraud signals from the loss video and routes the case: pay instantly above threshold, escalate to a human adjuster below.

Lemonade-style claim triage decision model (conceptual)

from xgboost import XGBClassifier

Inputs: user signals + claim text/meta + device fingerprint

FEATURES = [

"policy_age_days", # days since policy bound

"claim_amount_usd", # claim amount

"prior_claims_12m", # claims in last 12 months

"device_risk_score", # device fingerprint risk

"geo_risk_score", # local incident rate

"voice_stress_score", # voice statement stress signal

"photo_metadata_anom", # EXIF tampering suspicion

"text_inconsistency", # statement consistency

]

model = XGBClassifier(

n_estimators=600,

max_depth=5,

learning_rate=0.05,

subsample=0.8,

colsample_bytree=0.8,

eval_metric="auc",

)

Predict and route

def triage(claim_features, model, auto_threshold=0.05, review_threshold=0.35):

"""Below auto_threshold: pay; above review_threshold: SIU."""

risk = float(model.predict_proba(np.array([claim_features]))[0][1])

if risk < auto_threshold:

return {"action": "auto_pay", "risk": risk}

if risk < review_threshold:

return {"action": "human_review", "risk": risk}

return {"action": "siu_investigation", "risk": risk}

The secret of `<3초` settlements is not a fast model — it is that **the model only automates the extreme tails of the probability distribution**. The gray zone always escalates to humans.

Chapter 3 · Root — Telematics Sets the Price

Root Insurance (2015, Ohio) is the most aggressive telematics carrier. Before binding a policy the applicant runs the Root app for 2–3 weeks; that data determines price. If the data is insufficient or the score is low, Root declines coverage altogether.

Signals captured:

- Acceleration and braking patterns

- Cornering g-force

- Steering variability

- Night and late-night driving share

- Phone use while driving (screen interactions)

- Commute distance and highway ratio

- GPS-based cumulative mileage

Roots model compresses these into a **driver_score** from 0 to 100, combines them with traditional signals (geo, vehicle, age, credit), and produces a price. The pricing engine is an ensemble of a generalized additive model (GAM) and gradient boosting.

Root-style telematics driver_score model (conceptual)

TELEMATICS_FEATURES = [

"hard_brake_per_100mi",

"hard_accel_per_100mi",

"cornering_g_p95",

"night_driving_pct",

"phone_usage_min_per_100mi",

"highway_pct",

"miles_per_week",

"trip_count_per_week",

]

Target: 6-month loss ratio

params = {

"objective": "regression",

"metric": "rmse",

"learning_rate": 0.03,

"num_leaves": 63,

"feature_fraction": 0.9,

"bagging_fraction": 0.8,

"bagging_freq": 5,

}

booster = lgb.train(params, train_set, num_boost_round=2000)

The regulatory friction is discrimination. The NAIC AI Bulletin worries about "proxy discrimination" — variables such as ZIP code becoming surrogates for race. Root settled with the California DOI in 2024 and re-weighted a subset of signals.

Chapter 4 · Hippo — Home IoT to Prevent Fires and Leaks

Hippo Insurance (2015, California) wedded home insurance with IoT. It ships free smart leak sensors, smoke detectors, and door sensors to policyholders and reads the data to prevent claims. The thesis: **preventing a loss is cheaper than paying one**.

Key data sources:

- Aerial and satellite imagery (roof condition, tree overhang) — Cape Analytics, ZestyAI APIs.

- Public fire, crime, and flood data.

- Smart device telemetry (with auto-shutoff valves on leak detection).

- Selfie/home photos at application time → CV models infer condition.

Hippo combines all of this into a **HomeProtect Score**. Below a threshold, prevention kicks in — free leak sensors, roof inspections, routine reminders.

On the regulatory side, the interesting twist is that **gradient boosting models impact on price must be explainable per the NAIC bulletin**. Hippo surfaces SHAP values as reason codes to the applicant.

Chapter 5 · Ladder, Bestow, Ethos — Automating Life Underwriting

Automating life underwriting is harder than auto or home. Medical data carries more weight, the coverage horizon is long, and the cost of mispricing accumulates.

- **Ladder** (2017, California): a "ladder" term policy where the user freely adjusts coverage. Underwriting in five minutes, with paramedical exams routed selectively.

- **Bestow** (2016, Texas): API-first infrastructure. Other carriers can call Bestows underwriting API to sell life under their own brand (BaaS).

- **Ethos** (2016, California): "no-blood-test life" — underwrites without a medical exam by leaning on medical databases, prescription data, and motor vehicle records.

Their common data pipeline:

- MIB (Medical Information Bureau).

- Milliman IntelliScript (prescription history).

- LexisNexis Risk Solutions (public records, credit).

- MVR (Motor Vehicle Records).

- Self-attested data with identifier matching.

-- Term-life data mart — Ladder/Bestow/Ethos style joined query

SELECT

a.applicant_id,

a.dob,

a.gender,

a.smoker_status,

a.bmi,

a.requested_coverage_usd,

m.rx_burden_score, -- IntelliScript prescription burden score

l.public_record_flags, -- LexisNexis public record flags

v.mvr_violations_24m, -- last 24-month MVR violations

mb.mib_alerts -- MIB alerts

FROM applications a

LEFT JOIN intelliscript m ON m.applicant_id = a.applicant_id

LEFT JOIN lexisnexis_pub l ON l.applicant_id = a.applicant_id

LEFT JOIN mvr_records v ON v.applicant_id = a.applicant_id

LEFT JOIN mib m_bureau mb ON mb.applicant_id = a.applicant_id

WHERE a.created_at >= '2026-05-01'

AND a.product_code = 'TERM_LIFE_20Y'

The risk model usually runs in stages.

1. Identifier matching (name, DOB, SSN tail, address) across sources.

2. Imputation and outlier clipping.

3. Candidate models: GBM (XGBoost/LightGBM/CatBoost), neural networks, rule guardrails — ensembled.

4. SHAP-based explanation.

5. Decision rules: auto-issue / auto-decline / paramedical / human review.

Chapter 6 · Policygenius, The Zebra, Insurify — Comparison Marketplaces

Comparison marketplaces do not underwrite themselves, but they matter because **they are the entry point for underwriting decisions** — they choose which carrier sees an applicant.

- **Policygenius** (2014, NY): life, health, home, auto, even pet. Acquired by Zinnia (Eldridge) in 2024. Strong in life matching.

- **The Zebra** (2012, TX): auto comparison. Aggregates quotes from 200+ carriers.

- **Insurify** (2013, MA): AI chatbot Evia. Auto, life, home.

- **Gabi** (2017, CA): upload your existing policy PDF for automated comparison. Acquired by Experian in 2021.

Marketplaces run on two algorithms. One is the **user-carrier matching model** (which carriers are likely to quote this user?), the other is the **ranking model** (how to sort the quotes returned). Ranking learns from click and conversion data.

The regulatory issue is **fairness**. If a marketplace systematically favors one carrier in display, it can fall under the NAIC Unfair Trade Practices Act. So SHAP/LIME explainers are used in internal audits as well.

Chapter 7 · The Algorithm Stack — GBM, XGBoost, Neural Nets

In 2026, the workhorse of insurance underwriting is still **gradient boosting**. Neural networks play supporting roles in medical imaging, OCR, and NLP.

Selection criteria:

- Tabular + medium-scale data (hundreds of thousands to hundreds of millions of rows): GBM nearly always beats a neural net in both speed and accuracy.

- Raw image/text/time-series: CNNs and transformers shine.

- Explainability is paramount: GLM/GAM are safe; pairing SHAP with GBM is the next best.

| --- | --- | --- | --- |

Chapter 8 · Telematics Data Processing

Telematics is a 100 Hz time series. One year of driving easily produces over 100 million raw rows. Storage, preprocessing, and aggregation are the work.

A typical pipeline:

1. **Ingest**: phone SDK or OBD-II dongle into Kafka.

2. **Sessionize**: split into trips by ignition.

3. **Quality filter**: drop GPS gaps and accelerometer noise.

4. **Event labeling**: hard braking, hard acceleration, cornering, stops.

5. **Windowing**: aggregate over 100-mile, 7-day, and 30-day windows.

6. **Feature store**: Feast/Tecton — identical features for training and inference.

Telematics feature pipeline (conceptual Feast feature view)

- name: telematics_driving_features_7d

entities: [driver_id]

ttl: 30d

source:

type: kafka

topic: telematics.events.v1

features:

- name: hard_brake_per_100mi

dtype: float

- name: hard_accel_per_100mi

dtype: float

- name: night_driving_pct

dtype: float

- name: phone_usage_min_per_100mi

dtype: float

- name: cornering_g_p95

dtype: float

online_store: redis

offline_store: snowflake

The operational concern is **data freshness**. To show an up-to-date driver_score at quote time, online store latency must be `<100ms` end-to-end. Carrot Insurance in Korea and SBI Insurance in Japan run similar architectures.

Chapter 9 · Home IoT Data — Hippo, Cape Analytics

Home IoT delivers less volume than telematics but more variety.

- **Leak sensors**: minute-level humidity, temperature, pressure.

- **Smoke and CO detectors**: alarm events and heartbeats.

- **Door and window sensors**: open/close events.

- **Smart locks**: lock state and access attempts.

- **Aerial/satellite imagery**: quarterly — roof, vegetation, pool, structures.

Combining these yields the **2026-era home risk score**. It does more than predict loss ratio; it triggers proactive outreach. Hippo nudges policyholders — "Your roof has moss growth. We recommend a professional inspection."

{

"policy_id": "HPO-2026-998877",

"home_risk_features": {

"roof_age_years": 12,

"roof_condition_score": 0.62,

"tree_overhang_score": 0.78,

"pool_present": true,

"fire_zone_risk": "moderate",

"neighborhood_burglary_index": 41,

"leak_sensors_installed": 4,

"smoke_detectors_active": 6

"predicted_loss_ratio_24m": 0.43,

"premium_discount_applied_pct": 12

}

Chapter 10 · Medical Data AI — The Core of Life Underwriting

In life underwriting, medical data decides accuracy. The US relies on a standard stack.

- **MIB**: industry-shared database of reported medical info.

- **Milliman IntelliScript**: 5–7 years of prescription history.

- **ExamOne / LabCorp**: paramedical results (blood, urine, body composition).

- **Human API / Health Gorilla**: applicant-consented EHR retrieval.

- **Apple Health / Google Fit**: wearable data (optional).

On the ML side, **NLP on EHR text** is the hardest part. Physician notes are full of abbreviations, typos, and shorthand. Insurers fine-tune BERT or clinical-BERT variants for this.

An interesting pattern: reinsurers like Munich Re and SCOR run **clinical NLP APIs** in-house. Their models train on data from many carriers, so they generalize better than any single carriers model.

Chapter 11 · Automated Claim Payouts — Lemonades `<3초` Secret

The reason AI Jim can settle in `<3초` is not a faster model — it is **narrow auto-pay scope**.

Conditions for auto-pay include:

- Claim amount under `<$500/mo` band.

- Policy aged at least 90 days (anti churn-and-claim).

- Simple claim type (damage, theft).

- Consistency between statement and evidence above threshold.

- Device, geo, and behavior fingerprints in the normal distribution.

- Less than `<3회` (3 claims) on this policy in the trailing 12 months.

If any fails, the claim routes to a human adjuster. **The point is not fast decisions but quickly deciding whether to auto-decide at all.**

Auto claim payout decision (conceptual)

from dataclasses import dataclass

@dataclass

class ClaimContext:

amount_usd: float

policy_age_days: int

claim_type: str

statement_consistency: float

device_score: float

geo_score: float

behavior_score: float

prior_claims_12m: int

def can_auto_pay(c: ClaimContext) -> bool:

if c.amount_usd >= 500:

return False

if c.policy_age_days < 90:

return False

if c.claim_type not in ("damage_simple", "theft_simple"):

return False

if c.statement_consistency < 0.85:

return False

if c.device_score < 0.7 or c.geo_score < 0.7 or c.behavior_score < 0.7:

return False

if c.prior_claims_12m >= 3:

return False

return True

Chapter 12 · Claim Fraud Detection — KB Insurance and SBI Insurance

Claim fraud is estimated to absorb 5–15% of P&C insurer losses worldwide. Fraud detection is therefore as important an ML domain as underwriting itself.

Typical fraud patterns:

- Repeat claims by the same driver or vehicle (soft fraud).

- Claims right after binding (churn-and-claim).

- Synchronized claims after one accident (staged accidents).

- Medical overbilling (provider fraud).

- Inflated claims.

Fraud detection: XGBoost + graph embedding hybrid (conceptual)

FRAUD_FEATURES = [

"policy_age_days",

"claim_amount_usd",

"prior_claims_12m_count",

"claim_to_premium_ratio",

"provider_risk_score", # concentration at provider/repair shop

"ring_score", # network embedding across applicants, repairers, providers

"device_anomaly_score",

"geo_anomaly_score",

"text_consistency_score",

]

dtrain = xgb.DMatrix(X_train, label=y_train, feature_names=FRAUD_FEATURES)

params = {

"objective": "binary:logistic",

"max_depth": 6,

"eta": 0.05,

"subsample": 0.8,

"colsample_bytree": 0.8,

"eval_metric": "auc",

"scale_pos_weight": 30, # class imbalance

}

booster = xgb.train(params, dtrain, num_boost_round=1500)

SHAP for reason codes — NAIC AI Bulletin compliance

explainer = shap.TreeExplainer(booster)

shap_values = explainer.shap_values(X_inference)

KB Insurance in Korea has run an AI fraud detection system since 2023. ML scores plug into the loss adjustment and SIU (Special Investigation Unit) workflow so investigators triage from the top of the queue. SBI Insurance, Tokio Marine, and MS&AD follow the same pattern in Japan.

Chapter 13 · Graph-Based Ring Fraud Detection

A meaningful share of auto and medical fraud happens in **rings** — recurring networks of repair shops, doctors, lawyers, and drivers. Tabular ML alone misses them; you need a graph.

A typical graph schema:

- Nodes: applicants, claims, vehicles, repair shops, physicians, attorneys, agents.

- Edges: applicant↔claim, vehicle↔accident, repair shop↔repair, physician↔diagnosis, attorney↔representation.

Run Node2Vec, GraphSAGE, or GAT over this graph to embed each applicant or shop. Embeddings combine with other signals as fraud-score features.

Node2Vec-based ring score (conceptual)

from node2vec import Node2Vec

G: multi-edge graph of applicants, claims, providers, attorneys

n2v = Node2Vec(G, dimensions=64, walk_length=30, num_walks=200, p=1.0, q=0.5)

model = n2v.fit(window=10, min_count=1, batch_words=4)

Per-applicant embeddings → features for the next model

def get_ring_embedding(node_id):

return model.wv[str(node_id)]

NAIC AI Bulletin still imposes explainability — embedding dimensions alone are not a reason code. So graph-based scores ship alongside natural-language reasons such as "This applicant shares 12 claims with repair shop X."

Chapter 14 · NAIC AI Bulletin — The US Insurance Standard

The NAIC (National Association of Insurance Commissioners) released the **Model Bulletin on the Use of Artificial Intelligence Systems by Insurers** in December 2023. By 2026, nearly every state has adopted or is adopting it.

Core obligations:

1. **AI Systems Governance Program (AISGP)**: governance at board and executive level.

2. **Data and model risk assessment**: provenance, bias, accuracy, stability.

3. **Vendor governance**: the insurer remains accountable for third-party AI.

4. **Testing, validation, monitoring**: pre-launch and operational.

5. **Consumer protections**: reason codes, human review path.

6. **Records retention**: model decisions remain traceable.

NAIC AI Bulletin requirements mapping (conceptual yaml)

ai_system: claim_triage_v3

deployment_state: production

governance:

owner: chief_risk_officer

review_cycle_days: 90

board_review_required: true

risk_assessment:

data_sources:

- claim_intake_form

- device_telemetry

- third_party_geo_risk

bias_metrics:

- demographic_parity

- equal_opportunity

fairness_threshold: 0.05

testing:

backtest_window_months: 12

shadow_test_required: true

champion_challenger_required: true

monitoring:

drift_alert_psi_threshold: 0.2

performance_alert_auc_drop: 0.03

consumer:

reason_code_required: true

adverse_action_letter_template: AAL-2026-V2

human_review_path: /claims/escalate

records_retention_years: 7

The bulletin is lighter than the EU AI Act but operationally sharper — state insurance departments can examine and fine immediately. California, Connecticut, and New York ran the first round of examinations in 2025.

Chapter 15 · EU AI Act — Insurance UW Is High-Risk

The EU AI Act came into force in 2024 and is in serious application by 2026. Two categories matter for insurance underwriting.

- **High-risk AI**: life and health underwriting, pricing. Listed in Annex III.

- **Limited risk**: chatbots and customer support — transparency duties only.

High-risk AI obligations:

- **Risk management system** (Article 9)

- **Data governance** (Article 10): quality and bias review of training/validation data.

- **Technical documentation** (Article 11)

- **Record-keeping** (Article 12): log retention.

- **Transparency** (Article 13): disclose AI use to users.

- **Human oversight** (Article 14): final decision authority.

- **Accuracy, robustness, cybersecurity** (Article 15)

- **Quality management system** (Article 17)

- **CE marking and registration** (Articles 49, 60)

Fines reach 7% of global turnover or €35M, whichever is larger. For an insurer thats effectively class-action scale.

EIOPA (European Insurance and Occupational Pensions Authority) issued an insurance-specific guideline in 2025. Korean FSS and Japanese FSA reference it for their own.

Chapter 16 · Explainability — SHAP, LIME, Counterfactuals

NAIC and the EU AI Act both demand explainability. The 2026 insurance ML stack defaults to SHAP, with LIME and counterfactuals as backups.

- **SHAP (SHapley Additive exPlanations)**: decomposes the prediction by feature contribution. Global and local. TreeSHAP is exact on GBM.

- **LIME (Local Interpretable Model-agnostic Explanations)**: fits a local linear model around a prediction. Model-agnostic.

- **Counterfactual**: "What would need to change to flip the decision?" — DiCE, Wachter, etc.

In practice, SHAP values translate into **reason codes** through a mapping table.

SHAP → reason code mapping (conceptual)

REASON_CODES = {

"claim_amount_usd": "Claim amount sits in the upper quantile vs. the mean.",

"policy_age_days": "Policy has been bound for only a short time.",

"prior_claims_12m_count": "High count of claims in the last 12 months.",

"provider_risk_score": "Connected provider/repair shop has high claim frequency.",

"ring_score": "Network analysis shows multiple connected claims.",

}

def shap_to_reason_codes(shap_values, feature_names, top_k=3):

pairs = sorted(zip(feature_names, shap_values), key=lambda p: -abs(p[1]))

codes = []

for name, val in pairs[:top_k]:

reason = REASON_CODES.get(name, name)

codes.append({"feature": name, "shap": float(val), "reason": reason})

return codes

Consumers see 3–5 natural-language reasons rather than raw SHAP. This also aligns with the US Fair Credit Reporting Act (FCRA) adverse-action notice requirement.

Chapter 17 · Model Monitoring and Drift

Once a model ships, monitoring matters. Distribution shift erodes accuracy, which becomes either loss-ratio deterioration or a discrimination signal.

Standard operational metrics in 2026:

- **PSI (Population Stability Index)**: shift between training and production feature distributions.

- **CSI (Characteristic Stability Index)**: shift in the model output distribution.

- **Performance metrics**: AUC, KS, calibration.

- **Fairness metrics**: demographic parity, equal opportunity, calibration by group.

- **Volume**: counts of claims, quotes, declines.

Tools like Evidently AI, Arize, Fiddler, and WhyLabs handle ML observability. Korean firms often run BrightICS (Samsung SDS), and Japan tends to roll its own.

Alert examples:

| Metric | Threshold | Action |

| --- | --- | --- |

| PSI | `>0.2` | Data review |

| PSI | `>0.25` | Retraining review |

| AUC drop | `>0.03` | Champion vs challenger |

| Decline rate shift | `>5%` | Business review |

| Group calibration gap | `>5%` | Bias investigation |

| Reason-code distribution change | `>10%` | Explainability regression check |

Chapter 18 · Reinsurers and AI Risk-Sharing — Munich Re, Swiss Re, SCOR

Reinsurers are insurance for insurers — they share large losses and spread capital strain. The 2026 wrinkle is reinsurers sharing the AI model itself.

- **Munich Re**: aiSure — reinsurance coverage for losses caused by AI model failure.

- **Swiss Re**: Magnum — digital underwriting platform that carriers call for life decisions.

- **SCOR**: Velogica / SCOR Velogica — similar underwriting automation.

Reinsurers hold global data, so their training sets are large. They generalize better than single-carrier models, and newer carriers often adopt the reinsurers underwriting model directly.

Regulationally, Munich Res aiSure softens part of the **vendor governance** burden under NAIC and the EU AI Act. The combination of insurer-side validation and reinsurer-side coverage spreads the risk.

Chapter 19 · K-ICS and Korean AI Underwriting — Samsung, Kyobo, Hanwha, KB, Hyundai

Koreas K-ICS solvency standard took effect on January 1, 2023. Insurer assets and liabilities are now marked to market in the RBC (Risk-Based Capital) calculation.

The conceptual K-ICS RBC ratio formula is:

`RBC ratio = (Capital - Retained Risk) / (Required Capital)`

Retained risk combines insurance, market, credit, operational, and concentration risk via simulation. If AI underwriting accuracy drops, insurance risk rises and required capital rises with it.

Representative cases:

- **Samsung Life**: AI underwriting deployed (vendor name undisclosed); expanding share of paramedical-free decisions.

- **Kyobo Life**: medical-data-driven underwriting automation.

- **Hanwha Life**: digital onboarding plus automated underwriting.

- **KB Insurance**: in-house AI fraud detection and auto-claim adjusting automation.

- **Hyundai Marine**: vehicle damage estimation by photo.

- **DB Insurance**: telematics auto product.

- **Samsung F&M**: ML pricing in the direct channel.

- **Carrot Insurance**: a Hanwha-SK joint venture. Per-mile auto built on telematics.

The FSS published the "Financial AI Guideline" in 2024. Carriers operating AI underwriting take on impact-assessment, record-keeping, and consumer-protection obligations.

Korean FSS AI guideline mapping (conceptual)

financial_ai_system: life_underwriting_v2

target_business: term_life_auto_uw

impact_assessment:

consumer_impact: high

fairness_check_required: true

governance:

model_governance_committee: true

review_cycle_days: 180

records:

retention_years: 5

consumer:

algorithm_explainability_required: true

human_review_path_url: https://example.kr/underwriting/escalate

incident_reporting:

to_fss_within_days: 5

Chapter 20 · Japan — Dai-ichi, Nippon Life, LIFENET, SBI

Japan is conservative but its market is shifting under depopulation and aging. AI underwriting trails the US and Korea by half a step but is unmistakably advancing.

Representative cases:

- **Dai-ichi Life**: himawari AI — claim photo analysis for partial auto-pay; expansion of paramedical-free life underwriting.

- **Nippon Life**: in-house AI underwriting in collaboration with NTT Data.

- **Meiji Yasuda**: ML pricing in group life and pension.

- **Sumitomo Life**: Vitality (with South Africas Discovery) — wearable data feeds into pricing.

- **LIFENET**: internet-only, fast underwriting in five minutes.

- **SBI Insurance**: AI fraud detection in auto.

- **Tokio Marine**: claim automation and enterprise insurance AI.

- **MS&AD**: vehicle damage adjustment AI.

The Japanese FSA published its "Insurance AI Governance Guideline" in 2025, drawing on EIOPA and NAIC. It is being phased in along with the Japan ICS adoption schedule.

Japans ICS is based on IAISs ICS 2.0, calibrated to local market structure. Phased introduction 2025–2027, full application targeted by 2030. Like K-ICS, underwriting accuracy directly affects capital strain.

Chapter 21 · Discrimination and Bias — The Proxy Trap

The biggest regulatory friction with AI underwriting is discrimination. NAIC, EU AI Act, Koreas FSS, and Japans FSA all share the concern.

Core ideas:

- **Disparate treatment**: explicit use of protected attributes (race, sex, religion). Almost always illegal.

- **Disparate impact**: no explicit use, but the outcome disfavors a protected group. ECOA, Fair Housing Act, and similar US statutes apply.

- **Proxy discrimination**: variables strongly correlated with protected attributes (ZIP code, surname patterns, vehicle types) acting as proxies. The NAIC Bulletins headline worry.

Validation approach:

Bias check — demographic parity and equal opportunity

def fairness_report(df, score_col, label_col, group_col, threshold=0.5):

df = df.copy()

df["pred"] = (df[score_col] >= threshold).astype(int)

out = []

for g, sub in df.groupby(group_col):

pos_rate = sub["pred"].mean()

tpr = ((sub["pred"] == 1) & (sub[label_col] == 1)).sum() / max(

(sub[label_col] == 1).sum(), 1

)

fpr = ((sub["pred"] == 1) & (sub[label_col] == 0)).sum() / max(

(sub[label_col] == 0).sum(), 1

)

out.append({"group": g, "pos_rate": pos_rate, "tpr": tpr, "fpr": fpr})

return pd.DataFrame(out)

In practice, stripping protected attributes from training data is not the goal — **acknowledging and validating them is**. Then strong proxies such as ZIP code get reweighted or dropped.

Chapter 22 · Price Discrimination vs Personalization — Where Is the Line

Insurance pricing legally discriminates on risk, provided the discrimination is unrelated to protected attributes. The accepted pattern, then:

| --- | --- | --- | --- |

| Driving record | Legal | — | — |

| Occupation | Legal | — | — |

| Race/religion | — | — | All |

The EU Test-Achats ruling in 2012 outlawed sex-based pricing in EU insurance. Consequently, EU auto, life, and health insurance use sex-neutral pricing by default. US rules vary by state.

Chapter 23 · Comparative Map — US / EU / KR / JP Regulatory Matrix

| --- | --- | --- | --- | --- |

Chapter 24 · Operational Checklist — Before Launching a New AI UW Model

Finally, the practical checklist.

1. **Document data provenance**: source, license, consent for every training and validation dataset.

2. **Bias review**: demographic parity, equal opportunity, group calibration.

3. **Proxy check**: SHAP impact of strong proxies like ZIP code, vehicle type, occupation.

4. **Explainability**: at least five SHAP-based reason codes.

5. **Human review path**: escalation workflow for declines and high-risk cases.

6. **Champion/challenger**: start the new model on 5–10% of traffic.

7. **PSI/AUC monitoring**: drift and performance alerts.

8. **Retraining cadence**: quarterly or semi-annual.

9. **Vendor governance**: SLAs and test results recorded for any third-party model.

10. **Incident response**: SOPs for clawback, redress, and reprocessing when the model decides wrongly.

11. **Log retention**: input, output, and model version traceable for every decision.

12. **Consumer notice**: AI usage and appeal path disclosed.

13. **Regulatory reporting**: NAIC/EIOPA/FSS/FSA forms and cadence.

14. **Capital impact**: K-ICS, Japan ICS, Solvency II capital simulation.

These fourteen are the 2026 baseline. What matters most is that **AI underwriting carries equal weight for efficiency and accountability**. As fast quotes and automated payouts become routine, the social cost of a wrong decision rises in parallel. The `<3초` claim settlement is a marketing line; behind it sit fourteen guardrails.

References

- Lemonade — [https://www.lemonade.com/](https://www.lemonade.com/)

- Root Insurance — [https://www.root.com/](https://www.root.com/)

- Hippo — [https://www.hippo.com/](https://www.hippo.com/)

- Ladder — [https://www.ladderlife.com/](https://www.ladderlife.com/)

- Bestow — [https://www.bestow.com/](https://www.bestow.com/)

- Ethos — [https://www.ethoslife.com/](https://www.ethoslife.com/)

- Policygenius — [https://www.policygenius.com/](https://www.policygenius.com/)

- The Zebra — [https://www.thezebra.com/](https://www.thezebra.com/)

- Insurify — [https://insurify.com/](https://insurify.com/)

- NAIC AI Model Bulletin — [https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf](https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf)

- NAIC Home — [https://www.naic.org/](https://www.naic.org/)

- EIOPA — [https://www.eiopa.europa.eu/](https://www.eiopa.europa.eu/)

- EU AI Act — [https://artificialintelligenceact.eu/](https://artificialintelligenceact.eu/)

- Munich Re aiSure — [https://www.munichre.com/](https://www.munichre.com/)

- Swiss Re Magnum — [https://www.swissre.com/](https://www.swissre.com/)

- SCOR Velogica — [https://www.scor.com/](https://www.scor.com/)

- Korean General Insurance Association (KGIA) — [https://www.kgia.or.kr/](https://www.kgia.or.kr/)

- Korea Life Insurance Association (KLIA) — [https://www.klia.or.kr/](https://www.klia.or.kr/)

- Japan Life Insurance Association — [https://www.seiho.or.jp/](https://www.seiho.or.jp/)

- Dai-ichi Life — [https://www.dai-ichi-life.co.jp/](https://www.dai-ichi-life.co.jp/)

- LIFENET — [https://www.lifenet-seimei.co.jp/](https://www.lifenet-seimei.co.jp/)

- SHAP (SHapley Additive exPlanations) — [https://shap.readthedocs.io/](https://shap.readthedocs.io/)

- LIME — [https://github.com/marcotcr/lime](https://github.com/marcotcr/lime)

- Cape Analytics — [https://capeanalytics.com/](https://capeanalytics.com/)

- ZestyAI — [https://zesty.ai/](https://zesty.ai/)