Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Artificial Intelligence in Insurance Decision-Making

Abstract. Artificial intelligence has become part of the institutional architecture through which insurers estimate risk, decide access to coverage, calculate premiums, detect fraud, administer claims, and communicate contractual outcomes. Its economic significance arises from combining predictive inference with the authority of consequential decisions. This monograph studies that combination at the level of insurance contracts, portfolios, operational processes, affected individuals, and supervisory institutions. It compares actuarial prediction with administrative decision rules, analyzes source-data provenance and nonstationarity, examines fairness and explanation under insurance-specific constraints, and evaluates governance across the European Union, United States, and international standard-setting practice. The inquiry is a documentary and analytical synthesis of primary supervisory texts, enacted legal materials, and identified research literature; it reports no original insurer experiment, survey, causal identification exercise, or newly established theory. The central finding is that a model's statistical performance has meaning only within the decision process that gives its output practical force. A well-calibrated loss estimate can still be attached to an unlawful underwriting rule, an opaque claims denial, or a distribution strategy that deepens a protection gap. Conversely, interpretability and human review improve accountability only when reviewers possess authority, relevant evidence, decision records, and operational capacity. The research therefore evaluates a decision as a controlled sequence from data acquisition and target construction through validation, authorization, communication, appeal, monitoring, and correction. This is an analytical ordering of established obligations and research findings, not a new doctrine. The conclusion identifies reviewable controls and supervisory questions for AI-supported insurance decisions and states the limits of cross-jurisdictional inference.

Keywords: Artificial intelligence; insurance; underwriting; ratemaking; pricing; claims; fraud detection; insurance governance; actuarial modeling; model risk; algorithmic fairness; explainability; consumer redress; third-party risk; data protection; AI Act; IAIS; EIOPA; NAIC.

Introduction

Insurance has always made consequential classifications. It joins uncertain future loss to present contractual prices, underwriting rules, mutualized capital and legally enforceable promises. Artificial intelligence changes the scale, speed, granularity and architecture of that classification. A predictor trained on claims, interactions or external data can now inform a customer's quote, a fraud referral, an adjuster's work queue, a renewal decision or a chatbot's account of coverage. At each point, an estimate enters a system of legal rights, discretion, operational incentives and public expectations. The question is therefore what makes an AI-influenced insurance decision reliable, explainable and accountable over its complete lifecycle.

The growth is documented, though adoption figures require disciplined interpretation. EIOPA's 2024 digitalization survey results report AI use among 50% of surveyed nonlife and 24% of surveyed life insurers. Its 2026 generative-AI survey covers 347 undertakings in 25 countries, with nearly two-thirds reporting some active generative-AI use while many projects remain proofs of concept. Different respondents, time periods and definitions preclude treating those percentages as a single time series. The second survey concerns generative systems, whereas earlier insurance modeling also includes generalized linear models, trees, boosting and other predictive techniques.

Regulatory attention has followed these shifts. The IAIS 2025 application paper organizes supervision under proportionality, governance, robustness, transparency, fairness and redress. EIOPA's AI governance opinion interprets existing insurance-sector requirements. The NAIC's AI model bulletin articulates expectations for insurer programs and consumer-affecting decisions in participating US states. These sources converge on organizational accountability even while their legal status, scope and enforcement pathways differ. Their common concern is not an abstract machine acting alone; it is an insurer making or supporting insurance decisions with computational systems.

Research Questions, Scope, and Method

Four research questions guide the inquiry. First, which properties of AI forecasts matter for different insurance decisions, and how can those properties be tested? Second, where do bias, distribution shift, security failures and outsourced model dependencies arise? Third, how should explanations, human supervision and contestation be structured for consumers and regulators? Fourth, how do international standards and jurisdiction-specific rules allocate responsibility?

The unit of analysis is an insurer-linked decision instance: a proposed output, its data and model lineage, its business rule, an authorized disposition, and a possible later correction. The study covers personal and commercial insurance where general principles carry over, but distinguishes products and jurisdictions when statutory treatment depends on them. The dossier addresses the use of AI in decisions. General engineering of artificial intelligence, the history of insurance technology, and technical development unrelated to an insurance disposition belong to other Science inquiries.

Methodologically, this is structured documentary research and comparative institutional analysis. Primary evidence is drawn from IAIS standards and application guidance, EIOPA publications, the NAIC's bulletin and policy materials, Colorado's regulatory notice, the EU GDPR and AI Act, relevant court reasoning, and NIST/OECD materials. Peer-reviewed actuarial and machine-learning work supplies bounded propositions about predictive performance, explanation, calibration and statistical fairness. For every cross-domain inference, the text identifies that the source was not an insurance-specific experiment. No unpublished insurer data, confidential supervisory examinations, randomized trial or independently replicated software benchmark was available. Figures presented below are symbolic demonstrations, not empirical estimates.

Comparison is conducted at the level of legal instrument and decision function. An international principle, an EU regulation, a US model bulletin, a state rule and a nonbinding technical framework must retain their distinct status. Temporal statements are fixed to November 14, 2026. Prospective dates, including the amended AI Act implementation schedule, describe future legal application rather than completed compliance events. Interpretive results in this monograph arise from connecting published evidence with insurance decision consequences; they do not assert discovery of previously unknown empirical effects.

AI Applications across Insurance Processes

An insurance AI system can serve several fundamentally different roles. It may describe a photograph, estimate expected claim frequency, rank suspected fraud, suggest policy language, optimize assignment of adjusters, or recommend an underwriting decision. Those roles require separate assessment because an incorrect textual summary, a biased frequency score and an erroneous coverage denial create different types of harm. A single commercial service may perform more than one role, and its risk classification must follow actual functions and deployment conditions.

In underwriting, the pathway commonly combines exposure and eligibility data, a risk model, contractual appetite constraints and an approval rule. Pricing combines a forecast distribution with expenses, capital, taxes, margin, competition and permissible differentiation. Claims handling combines evidence collection with coverage interpretation, validation of damage and payment authority. Fraud detection is usually a ranking or investigative triage mechanism rather than proof of misconduct. Customer services add information retrieval and generation, with a special risk that fluent text will be mistaken for binding policy interpretation.

Consequently, assessment must begin with the consequential action. One model might have acceptable accuracy for prioritizing a harmless internal queue and unacceptable reliability for automatically declining a claim. The NIST AI RMF is use-case-agnostic and voluntary, but its emphasis on context, risk mapping and lifecycle measurement supports specifying decision context before selecting metrics. The IAIS Insurance Core Principles likewise anchor insurance outcomes in governance, risk management and conduct obligations. The two sources operate at different normative levels; neither transforms a generic benchmark score into evidence of insurance compliance.

The value chain also contains feedback. A declination changes the insured population; a premium increase changes lapse and retention; a claims referral changes investigation intensity; a chatbot answer can change whether a policyholder submits evidence. Later training data therefore partially reflect prior algorithmic decisions. Because only observed claims among retained customers are available, an insurer can mistake a selectively observed target for the underlying insured population. This selective-label problem is important even when the first fitted model was unbiased on its original data.

Decision Economics and Actuarial Foundations

Insurance prediction ordinarily targets future contingent loss, but the target must be specified. In a simplified one-period model, expected pure premium for exposure x is E[L|X=x] = E[N|X=x] × E[Y|N>0,X=x] only under an appropriately defined frequency-severity decomposition and compatible severity conditioning. Here N denotes claim count, Y the average severity conditional on claims, and L aggregate loss. Real-world models may require dependencies, deductible effects, policy limits, reporting lags, inflation, catastrophe tails and exposure periods. A predictor of propensity to purchase or estimated lifetime value is conceptually different from a predictor of insurable loss, even if both enter the same pricing system.

The insurer's final premium also depends on loaded expenses, risk and capital costs, product strategy and applicable regulation. Let an illustrative pricing decision be P(x)=m(x)+e(x)+k(x)+q(x), with m expected covered loss, e expense, k capital/risk load and q any permissible commercial adjustment. This identity is a decomposition for analysis, not a formula required by law. A claim that an AI model is accurate at estimating m does not validate all components of P or the legal admissibility of the features used.

Insurance-specific empirical research shows why methods cannot be assessed in isolation. Henckaerts and colleagues (2021) study tree-based tariff construction with the distinctive properties of claim counts, exposure and long-tailed severity. Their results demonstrate economically relevant modeling possibilities and practical ways to examine tariffs; they do not establish that boosting universally outperforms actuarial GLMs. Richman and Wüthrich (2023) develop LocalGLMnet to improve the interpretive structure of neural prediction on tabular data. These contributions establish a research basis for comparing nonlinear capability with intelligible tariff factors.

A defensible model comparison tests the same decision objective with compatible training exposure, temporal validation periods and documented materiality thresholds. Measures of frequency fit, likelihood, deviance, rank discrimination and portfolio profit can diverge. A high discrimination score can coexist with poorly calibrated premiums, unstable subgroups or unjustifiable contract treatment. The scientific consequence is straightforward: model selection must specify which errors matter economically, for whom, and over what horizon.

Underwriting and Coverage Access

Underwriting is an exercise of contractual selection constrained by actuarial justification, product rules and law. The system may extract facts from submissions, detect omissions, predict loss or suggest coverage terms. In every case the resulting action should be identifiable: accept, request evidence, vary limits, apply an exclusion, refer for human assessment or decline. A model score alone lacks the legal and commercial content of these choices.

Data provenance is central because applicants and consumers may have little knowledge of external variables. A third-party lifestyle proxy or household characteristic could correlate with risk while also duplicating protected or socially sensitive attributes. Data quality review must examine collection context, lawful use, historical completeness and how third-party variables transform into underwriting features. A vendor's assertion that a variable is 'non-sensitive' does not determine the consequences of correlated features. A documented feature rationale must connect insurance relevance, predictive contribution, permissible use and consumer effects.

Underwriting validation should examine both applicants who received coverage and those who did not. Historical loss experience among accepted risks cannot directly identify the counterfactual losses of declinations. Excluding previously rejected applications from performance testing creates a selective sample and may reward the very exclusion rule under examination. An ethical and scientifically defensible audit therefore reports what is estimable from observed data and which claims require additional information or explicit assumptions.

US state frameworks illustrate the operational significance of these questions. The NAIC model bulletin calls for written insurer AI governance and documentation, but adoption is state dependent. Colorado's 2025 amended Regulation 10-1-1 notice describes governance controls for insurers offering specified life, private passenger auto and health coverages when external consumer data and predictive methods are used. A supervisory examination must apply actual local legal provisions and insurer lines, not treat model governance as identical across all American jurisdictions.

Pricing, Risk Pooling, and Affordability

AI pricing can increase granularity by identifying differences in expected loss that broad rating categories conceal. Greater risk sensitivity may mitigate cross-subsidies within an existing tariff, improve selection and help insure risks previously priced too crudely. It can also concentrate price increases on individuals with limited ability to mitigate circumstances or afford coverage. Those are separate effects. The insurer's technical ability to forecast individual risk does not itself decide which social distribution of insurance burdens is legitimate.

Risk pooling is an institutional commitment to finance uncertain losses collectively under policy and market rules. Full individualization is conceptually bounded by stochastic uncertainty, the function of pooling and legal constraints on discrimination and access. A life or health insurer may face different constraints from a motor or commercial insurer. Product differences must appear before any claim of a single fairness criterion for all coverage.

Consider two portfolios with similar aggregate expected cost. A newly granular model reassigns premiums from lower-scored to higher-scored risks, leaving total expected collected premium unchanged. A conventional prediction metric may improve. Yet if higher rates cause the most exposed customers to withdraw, the residual insured group and its loss distribution change. The original validation no longer describes the market equilibrium. This is an analytical example, not a measured causal effect. The feedback must be assessed with behavioral evidence or an explicitly qualified scenario model.

The IAIS supervisory AI guidance recognizes societal concerns about granular pricing and fairness as part of a proportionate supervisory approach. Keller's insurance AI analysis discusses the complexity of transparency and fairness tradeoffs within an insurance-specific context. The strongest synthesis is to assess premium adequacy, lawful differentiation, financial inclusion and coverage continuity together, keeping any distributive judgments explicit rather than hiding them inside a technical loss function.

Claims, Fraud Detection, and Customer Processes

Claims administration is the point where an insurance promise is tested against an event. AI may transcribe notice-of-loss statements, classify documents, identify missing evidence, estimate repair costs or flag inconsistencies. The model's function should be constrained by the contract and established claims process. A predicted fraud score is a hypothesis for appropriately authorized investigation, not a finding of fraud. A predicted low settlement probability is not evidence that a claimant lacks entitlement.

False positives and false negatives have asymmetric consequences. A false referral can delay indemnification, impose additional documentation costs and damage trust; a missed fraudulent claim can increase insurer costs and mutualized premiums. Management should evaluate thresholds against both impacts, including resolution time and error reversal. An accuracy rate pooled across small routine claims and severe disputed claims is an incomplete summary. The audit requires stratification by severity, product, vulnerability and whether the model actually controlled the disposition.

Generative models introduce a separate form of risk. They can produce plausible summaries of exclusions, duties or deadlines that are wrong. The NIST generative-AI profile identifies risk classes specific to generative systems, while EIOPA's 2026 survey describes insurer concerns about inaccurate outputs and security. An insurance chatbot handling contractual questions needs source retrieval tied to the authoritative issued wording, record of the version shown, an escalation path and controls against invented coverage statements. A generic disclaimer is operationally weaker than preventing an unverified answer from being acted upon.

The relevant performance metrics include validated extraction accuracy, agreement with the operative contract, false-positive investigative referrals, average and tail processing delay, complaint reversals, claimant comprehension and accessibility of escalation. An effective claims system should preserve underlying documents and a traceable reason for its eventual disposition even when technical components are replaced.

Data and Model Governance

Data governance starts with identifying which populations and transactions a dataset represents. Underwriting and claims datasets contain timing differences, policy limits, earned exposure, subrogation, recoveries, inflation, settlement practices, catastrophe regimes and missingness. Combining heterogeneous claims without adjustment can train a model on administrative artifacts. The training record should specify provenance, permissions, retention schedule, feature transformations, label definitions and exclusions. Data minimization and lawful bases for personal information are separate requirements from predictive usefulness.

The General Data Protection Regulation governs relevant personal-data processing in the EU, including special-category information, transparency and restrictions on some automated individual decisions. A model trained outside the EU can still enter an EU insurer's legal environment through deployment. Governance therefore covers collection and downstream use, not simply where a model's weights were created. It also covers testing datasets: fairness evaluation may require demographic information subject to particular legal conditions and safeguards. These issues call for a jurisdiction-specific compliance analysis rather than uncontrolled collection of sensitive attributes.

A model register should connect each system version to owner, purpose, intended users, material inputs, training period, provider, validation report, approvals, runtime endpoints, control thresholds, incidents and retirement rules. Records become meaningful when a particular policy or claim decision can be reconstructed from the exact applicable version. If an insurer cannot identify the score or rule used for a challenged decision, aggregate dashboard documentation is insufficient to explain that event.

The EIOPA governance opinion describes relevant expectations under existing insurance-sector legislation, while the NIST risk framework helps organize mapping and measurement activities across use cases. Together they support a practical separation of model-development accountability, insurance product ownership, independent validation and consumer operations. That allocation is an institutional analysis consistent with the cited instruments; it is not asserted as a uniformly mandated organizational chart.

Validation, Calibration, and Robustness

Model validation asks whether a prediction is fit for the precise context in which an insurer acts on it. A complete review includes construct validity of the outcome; sampling and missingness; actuarial and statistical adequacy; leakage and temporal integrity; predictive calibration; business-rule effects; subgroup performance; stability over time; susceptibility to manipulation; and recovery from foreseeable failure. The acceptance threshold should be established before a favorable test result is observed.

Calibration is especially important when numerical predictions become insurance prices or reserve indications. For a binary event with forecast p, calibration asks whether the event frequency among comparable forecasts approximates p. Classification accuracy alone is unsuitable when claim events are rare. In count and severity settings, the appropriate diagnostic is different, and aggregate calibration can conceal materially miscalibrated subpopulations. Guo and colleagues (2017) show calibration problems and temperature-scaling results in noninsurance neural-network benchmarks; the insurance implication is to validate the actual deployed probabilities and losses, rather than to transfer a preprocessing remedy without testing.

Temporal validation must respect chronological dependence. Randomly splitting policy records may leak future information through claims settled later, repeated claimants, related policies or post-decision variables. Proper backtesting preserves feature availability at the decision timestamp, maturity of claims and lagged reporting. Catastrophe years, inflation regimes and fraud responses may change outcome distributions. A model can be accurately estimated for a historical regime and obsolete for the next renewal period.

Ovadia and colleagues (2019) document challenges in uncertainty estimation under distribution shift across general machine-learning benchmarks. The direct insurance conclusion is limited but important: drift thresholds should be checked against changing input and outcome distributions, not only a stable global accuracy dashboard. When labels arrive late, leading indicators such as feature drift and complaint patterns warrant attention while realized-loss validation remains pending.

Robustness also includes adverse conditions. An image estimator may fail on unusual damage; a fraud detector can be gamed; a third-party endpoint can silently alter preprocessing; a large language model can misread a policy table. Testing should document failures and control actions. Abstention, manual review, conservative defaults and temporary suspension are valid risk treatments when model outputs fall outside validated conditions. Such controls require capacity planning to remain effective during catastrophe claim surges.

Operational Security and Third-Party Reliance

Insurance AI depends on data pipelines, storage, identity systems, cloud providers, model services and service-level agreements. The security question spans confidentiality of personal or medical records, integrity of models and input data, endpoint availability, privileged access, prompt or document manipulation, and failure propagation through shared vendors. A seemingly small change to model configuration can change a high volume of prices or claim queues without altering product wording.

The EU's Digital Operational Resilience Act applies ICT risk-management and third-party resilience obligations to covered financial entities from January 17, 2025, with entity-specific scope considerations. This is distinct from AI-specific model requirements. An insurer should maintain service inventories, substitution or exit arrangements where proportionate, incident processes and business continuity tested against credible AI dependency failures. Evidence of a vendor certification does not replace institution-specific impact analysis.

Third-party concentration can occur even when several nominal vendors rely on the same foundation model, cloud platform, data broker or outsourced claims service. A procurement register limited to direct contractual counterparties can miss that hidden correlation. The NAIC AI policy update records attention to third-party models and examination tools, and the IAIS 2025 supervisory approach addresses outsourced AI risks. Practical oversight must therefore request sufficient model documentation, change notifications, audit rights, incident cooperation and data rights to validate the insurer's own use.

Cybersecurity testing requires explicit permission and safe scope. The research here recommends governance controls and failure-mode analysis; it provides no exploit instructions or untested assertion that a particular technology is secure. Material incidents should be traced from model output to consumer impact, correction, reporting obligations where applicable, and the verification that normal operations were restored.

Explainability and Consumer Understanding

Explanation has several audiences with different informational needs. A model developer needs reproducible behavior and diagnostics. A board needs material exposures and control ownership. A supervisor needs test evidence, legal classification and audit records. A customer needs a truthful account of the material reasons for a price, coverage condition, investigation or claim decision, expressed in comprehensible terms with an available route to correction. These are related but distinct explanation products.

Lundberg and Lee (2017) develop SHAP feature attributions for predictions. Attribution helps describe a fitted model's allocation of output among represented features under a chosen background convention. It cannot independently establish that a feature caused a loss, that model inputs were lawful, or that a particular insurance contract was correctly applied. Correlation, proxy variables and interactions can make superficially intuitive feature stories misleading. Explanatory graphics therefore need a statement of their method and limitations.

Rudin (2019) argues for inherently interpretable models in high-stakes settings. This is an important methodological challenge to reflexive use of black-box explanation, but model choice remains a case-specific empirical and institutional decision. Where a transparent model offers adequate performance, it can simplify verification and challenge. Where greater complexity yields a demonstrably material benefit, the insurer must still provide reliable operational reasons and retain the evidence needed to review a challenged outcome.

A usable consumer explanation should distinguish observed information, a predictive inference, a contractual or regulatory rule, and the human or automated action taken. A consumer may correct an inaccurate accident history without disputing the statistical model. A claimant may challenge whether a coverage exclusion applies without needing access to the insurer's training code. Explanation must support the actual right or review available in the jurisdiction, respect proprietary restrictions legitimately applicable, and disclose enough to make procedural redress meaningful. As the revised OECD AI recommendation recognizes, transparency and accountability are linked to context and adverse-outcome challenge.

Bias, Fairness, and Redress

An insurance risk model can exhibit bias through data acquisition, target choice, sampling, features, training, deployment or differential consequences. Bias is not a single scalar. Sampling bias results when insured records misrepresent applicants; measurement bias when losses or health status are recorded differently; label bias when expense or historical settlement is treated as a proxy for underlying need; feedback bias when prior refusals change future observations. A model audit should name the mechanism it investigates.

A striking cross-domain warning is Obermeyer and colleagues (2019), who documented racial bias in a health-management algorithm that used medical spending as a proxy for health need. It was not an insurance pricing experiment. Its relevant lesson for insurers is that an apparently predictive and administratively convenient target can encode unequal access or historical treatment. Before training an insurance model, the organization must identify what its chosen label measures and what alternative social or contractual concept it is meant to represent.

Formal fairness measures can conflict. Chouldechova (2017) analyzes incompatibilities of some calibration and error-rate criteria with differing group base rates in recidivism scoring. Importing a single mathematical fairness metric into insurance therefore requires explicit normative choice, outcome definition and legal review. Equal group error rates, equal risk classification, equal access and premiums proportional to permissible expected loss express different objectives. The appropriate assessment may depend on product, protected characteristics, jurisdiction and remedy.

The legitimacy of risk differentiation is a distinct insurance question. Statistical disparity does not automatically prove unlawful discrimination, and formally accurate loss prediction does not automatically authorize the associated pricing or coverage rule. Analysis must test both substantive law and consumer impact. For some interventions, fairness is pursued through restrictions on variables; for others through underwriting review, pricing caps, access arrangements, explanation, or appeal. Redress can include correction of erroneous data, reassessment with a validated process, refund, reconsideration of coverage, or escalation to an applicable dispute mechanism.

A redress system is effective only when decisions can be revisited, the appeal route is communicated, evidence can be entered, and successful challenges change both the individual outcome and, where necessary, the underlying process. Complaint frequencies alone are not an adequate harm proxy when less-resourced consumers abandon disputes or cannot identify the basis of a decision.

Human Oversight and Decision Accountability

A nominal human-in-the-loop arrangement can fail when the reviewer sees only a score, lacks original documentation, is measured by the rate of agreement with the model, or has insufficient time to exercise independent judgment. Human oversight is meaningful when the reviewer can understand the affected decision, obtain necessary records, disagree with an output, override the recommendation where authorized, provide a reason and trigger correction of recurring errors. Auditability requires proof that these powers function in practice.

Accountability extends beyond the last person who clicked approve. Product management determines permissible risk appetite; actuaries examine tariff construction and technical adequacy; data teams define features and training; model risk teams test performance; claims or underwriting leaders authorize dispositions; legal and compliance professionals assess obligations; executives and boards determine control resourcing. An insurer cannot contract away the responsibility for customer outcomes simply because a vendor supplied the model. The precise legal allocation is jurisdiction-dependent, but the governance necessity follows from the IAIS accountability framework and the NAIC's insurer-program expectations.

An effective approval record identifies the decision use case, material model changes, relevant validation limitations, risk acceptance, authorized user population, explanation mechanism, incident threshold and challenge process. It should identify which decisions can run automatically and which require review. Approvals must be re-opened when a model is repurposed, a third party changes a material component, or complaints reveal a failure mechanism absent from original testing.

Monitoring also has to be connected to remediation. A dashboard showing deteriorating calibration without an assigned response owner provides observation but little protection. Escalation should be tied to actions: stop using an unvalidated variable, require a second assessor, recalculate decisions from a safe version, notify affected customers where required, and document what was corrected. Genuine accountability is observable in how an institution responds to error.

Worked Analytical Cases and Audit Questions

Case: A granular motor insurance tariff

An insurer replaces a broad tariff with a tree-based model trained on several years of reported claims. The new model shows lower aggregate deviance and separates frequency segments more sharply. The scientific audit first checks exposure denominators, policy maturity, inflation adjustment, subgroup calibration and whether internal loss experience includes customers previously declined. It then examines the premium formula, permitted variables, change in price distribution, projected lapse, complaints and effects on access. Higher predictive fit is a finding about one tested objective; the decision to deploy requires separate legal, economic and conduct evaluation. The illustration is hypothetical; it does not estimate any insurer's gain or discrimination.

Case: An automated claim fraud referral

A model assigns an anomaly score to submitted invoices, causing high-scored claims to enter enhanced investigation. The validation set contains investigated cases disproportionately drawn from historic suspicion rules. The observed 'confirmed fraud' labels therefore reflect selection into investigation. A sound audit reports this verification bias, measures claimant delay, estimates false referrals only with defensible labels and creates an appeal and independent review process. An algorithmic high score is not evidence of guilt. If the insurer cannot distinguish investigatory priority from actual coverage adjudication, the decision design is defective.

Case: A generative customer assistant

A customer asks whether storm water damage is insured under a particular issued policy. A language model retrieves an obsolete wording and answers confidently. The relevant controls are version-locked policy retrieval, quotation of controlling clauses, uncertainty detection, an abstention/escalation route and preservation of the interaction record. Measuring linguistic fluency or user satisfaction without measuring contractual correctness would validate the wrong target. The generative risk profile and insurer GenAI survey explain why this error class warrants distinct treatment.

Case: A vendor-issued life underwriting score

An insurer receives an external numerical score without complete training data. It maps the score into automatic declination bands. The audit asks whether a lawful basis exists for underlying data, whether proxies and target labels are understood, whether the supplier can provide meaningful documentation, how appeal and human review operate, and whether the insurer can independently validate performance. The SCHUFA reasoning illustrates why strong reliance on a third-party score can be legally significant in an EU automated-decision analysis, subject to the facts. The insurer cannot equate vendor opacity with adequate explanation.

Common audit questions

Across the cases, the researcher or supervisor should ask: What exactly was decided? Which model version and training labels influenced it? Which feature values existed at that moment? What population and horizon were validated? Which error metric is material to that disposition? Were lawful eligibility and product rules separated from predictive outputs? Can an affected customer obtain a meaningful reason and correction? What happens under drift, outage, adversarial input and a vendor change? Who can suspend the system, who records the incident, and what evidence demonstrates restoration?

These questions synthesize the evidence into a practical audit agenda. They are analytical consequences of existing research and controls, not a newly named or proprietary model of insurance AI.

Quantitative Validation and Decision-Loss Analysis

Statistical validation should serve a decision, and the decision's costs must be stated before model comparisons acquire practical meaning. A frequency model estimated from rare claims might achieve excellent accuracy by predicting no claims at all. It could nevertheless be unusable for premium adequacy because its conditional expected counts would be severely distorted. For actuarial rating, logarithmic scores, frequency deviance, severity error and calibrated aggregate predictions have different purposes. For fraud triage, precision at a feasible investigation capacity may matter more than global area under a classification curve. For claims denial or policy declination, the cost of an unsupported adverse outcome warrants direct control in addition to any predictive metric.

Let D be a proposed insurance decision, Y a later observed outcome and C(D,Y) a stated decision loss that includes financially and legally relevant consequences. A mathematically optimized policy would minimize expected C subject to binding contractual, regulatory and institutional constraints. The expression describes an existing decision-theoretic approach rather than the invention of a new loss doctrine. It reveals a key modeling choice: an insurer that assigns zero cost to unjustified delay will systematically prefer some harmful claims actions even if the mathematical optimizer works perfectly. A revised cost function cannot itself determine legal rights. The legal and contractual admissibility of D is an external constraint that must be enforced and tested.

For a binary loss event, an illustrative calibration assessment partitions predictions into predeclared bands B. Within each band the observed event rate is compared with the average predicted event probability, while sample size and uncertainty are documented. An aggregate portfolio check compares total predicted and realized losses after aligning exposures, maturity and valuation bases. In a heavy-tailed insurance book, a small number of claims can dominate the aggregate; robust conclusions require uncertainty intervals, severity diagnostics and sensitivity to unusually large observations. The calibration literature supports the principle that probability estimates need explicit evaluation, but its neural classification experiments are not an insurance tariff validation study.

A validation design should pre-specify the following rather than improvise thresholds after performance is known: the intended policy or claim action; the training and test observation windows; population inclusion rules; contractual and regulatory restrictions; chosen response variable and its availability date; benchmark estimator; primary predictive and business metrics; subgroup and tail diagnostics; acceptable confidence intervals; missingness sensitivity; override triggers; and procedures for stale or delayed outcomes. The record should capture adverse results alongside successful ones. A model with superior average performance but concentrated error in severe claims requires a materially different decision from one whose performance gain is evenly distributed.

Calibration is also conditional on decisions taken. Suppose a model predicts high loss and the insurer declines coverage; that policy then produces no insured claims for the insurer. Comparing only realized insured losses with rejected applicants' predicted losses would confuse absence of observation with absence of underlying risk. This identification problem cannot generally be solved from insured claims alone. Possible investigations include carefully governed external information, quasi-experimental variation, valid instrumental assumptions or explicit uncertainty bounds; each carries its own limitations. The appropriate finding may simply be that the risk of excluded applicants is unidentified by the retained portfolio.

Risk estimates should be audited at deployment horizons appropriate to their purpose. Price models may experience inflation and selection drift across renewal cohorts; fraud systems may provoke changed adversary behavior within weeks; chatbot errors may emerge immediately from policy document changes. The general evidence on dataset-shift uncertainty shows why confidence statements can degrade after distribution changes. The responsible control is continuing comparison with a baseline and explicitly defined intervention criteria, recognizing that ground truth may be delayed.

Portfolio-Level Effects, Selection, and Protection Gaps

Insurance decisions must be evaluated both individually and collectively. A price or underwriting action changes which risks remain within the pool; retention then modifies the data on which subsequent models are trained. An operationally efficient fraud filter may alter claim reporting and investigation distributions. A customer recommendation system may concentrate distribution on profitable cohorts and leave some households with reduced access. These effects can emerge even if each individual prediction remains reasonably calibrated on a stable benchmark sample.

The portfolio consequences of a granular tariff can be investigated with scenario analysis. Define a baseline portfolio, an alternative decision rule and plausible behavioral responses to price or coverage changes. Estimate expected claims and expenses for each retained population under transparent assumptions; then calculate changes in premium revenue, loss ratio, lapses, coverage take-up and concentration. Stress the conclusions under uncertainty in price elasticity, catastrophe exposure and competitive responses. The resulting estimates should be labeled simulated rather than measured; without causal demand evidence, reported welfare gains or harms remain conditional scenarios.

Insurability has both a risk and an institutional dimension. A person may face a low predicted loss and still lack coverage because the product is poorly distributed, the premium includes fixed expenses, or the claim process is inaccessible. A high-risk person may require socially arranged coverage, cross-subsidy or public support rather than increasingly granular private pricing. The IAIS's fairness and redress guidance places consumer outcomes and the social impact of risk differentiation within the supervisory discussion. Whether a given market should preserve a particular level of solidarity is a policy choice subject to law and democratic institutions; machine learning does not settle that choice.

Model concentration adds a separate systemic channel. If multiple insurers use similar external data vendors, model architectures or catastrophe proxies, they may react similarly to an event even when their direct contract relationships differ. Shared technical assumptions can correlate pricing responses, underwriting withdrawals or service disruptions. The research does not quantify this correlation. It identifies a prudential question for further empirical and supervisory investigation: whether apparently diversified insurer portfolios conceal concentration in their decision infrastructure. The NAIC third-party oversight work and IAIS supervisory standards provide institutional reasons to monitor provider dependence and resilience.

Distributive analysis is most credible when technical and normative claims are reported separately. A change in indicated risk cost can be estimated under documented assumptions. An adverse impact on vulnerable insureds can be measured if appropriate, lawful data and a defensible study design are available. A conclusion that a particular premium difference violates a specific rule requires a separate legal determination. Keeping these statements distinct makes both actuarial research and public debate clearer.

Documentation, Audit Trails, and Remediation Protocols

The practical test of governance is whether a consequential decision can be reconstructed and corrected. An event-level audit record should retain the decision identifier, applicable policy form and endorsements, jurisdiction, decision timestamp, input provenance and effective values, model and rule versions, output and uncertainty, human interventions, communicated reasons, later changes, and evidence considered on review. Sensitive data require access control, retention limits and legal handling. A record that captures every input without purpose limitation can itself create privacy risks; minimal necessary reconstruction is preferable to indiscriminate replication.

Model documentation should distinguish the prediction report from the insurance decision report. The former describes target, estimator, features, validation, drift and model limitations. The latter specifies which actuarial or operational threshold transformed the estimate into a quote, decline, payment, referral or customer instruction, and why that action was permissible. A single model may support many business rules; an accurate model card cannot explain all of them. Conversely, a fully documented underwriting rule can be undermined by an unvalidated upstream data feed.

Independent review has a concrete task sequence. The reviewer identifies the exact decision and controlling legal or policy context, obtains the relevant original records, checks whether the input was accurate and permissible, reproduces the model output where feasible, verifies the decision rule, and evaluates the communicated reason. If the outcome is wrong, the institution rectifies the customer record and considers whether similar decisions require retrospective examination. A review system that cannot change a disposition is a logging procedure, not effective redress.

Escalation thresholds should be defined by material impact, with examples including unanticipated rises in claim delays, systematic error in a subgroup, sudden changes in external scores, repeated policy misquotation, anomalous complaints and a loss of third-party service. Response plans should identify a system owner, an authorized suspension decision, a safe manual process, a method of enumerating affected transactions, and evidence of corrected outputs. After remediation, the organization should test the changed system on the failure mode that triggered intervention. Governance becomes credible through this closed trail from detection to restoration.

For supervisory inspection, anonymized examples are often informative only when accompanied by sampling method and disposition counts. A collection of successful overrides cannot establish that ordinary reviewers regularly exercise judgment. A useful file contains sampling frames, unfavorable findings, time-to-correction, model versions and control effectiveness, with lawful confidential access to the underlying records. The NAIC model bulletin explicitly emphasizes insurer documentation for examinations, while the EIOPA AI opinion locates record keeping and human oversight within sectoral risk governance.

Research Design for Comparative and Empirical Follow-Up

The existing documentary record is strong on supervisory objectives and methodological warnings; it is thinner on observed causal effects of deployed insurance AI. A reproducible study should therefore formulate testable questions about specific decision functions rather than treating aggregate AI adoption as a result variable. One possible project would compare a defined cohort of claim decisions before and after the introduction of an AI triage rule, controlling for claim type, severity, seasonality, staffing and reporting delay. The study would predefine outcomes such as time to resolution, payment reversals and consumer appeals. Simple before-after comparisons would remain vulnerable to confounding, and insurer selection into deployment would require an explicit identification strategy.

A fairness study could evaluate both statistical risk adequacy and access effects for legally relevant groups with appropriate data governance. It should identify whether protected characteristics are available lawfully, whether proxies are involved, and whether differing base rates affect the selected metric. The fair prediction analysis demonstrates why metric selection requires normative justification; it cannot directly establish discrimination in a named insurer. A label-validity study could compare the predictive target to the contractually relevant insured loss and assess whether historical settlement choices contaminate labels. The health algorithm study offers a clear empirical illustration of target mismatch in another domain.

Consumer comprehension could be tested through controlled experiments comparing an insurer's existing reasons with alternative explanation formats, asking whether users can identify the decisive facts, correct an error, understand remaining coverage and initiate a valid challenge. Such an experiment should not measure confidence alone, since a persuasive but wrong explanation can raise confidence without increasing understanding. A deployment experiment must also guard against denying legally required information to control groups.

Finally, cross-jurisdictional comparison should preserve the legal timeline of each system. In the EU, the new 2026 Digital Omnibus Regulation amended the AI Act in July 2026 and changed implementation details relevant to high-risk systems. In the United States, state adoption of the NAIC governance bulletin proceeds through distinct legal mechanisms. A research table that groups these sources as equivalent 'AI rules' without legal status and effective-date columns would be analytically defective.

A robust research program can combine observational insurer data, public regulatory decisions, structured supervisory interviews and appropriately controlled experiments. Each method answers different questions and carries different access and inferential limits. This dossier supplies the conceptually precise questions and source-grounded institutional context, while openly reserving empirical answers for studies that actually acquire and analyze the necessary evidence.

Research Synthesis and Limitations

The literature supports several bounded findings. Actuarial studies such as Henckaerts et al. (2021) and Richman and Wüthrich (2023) show that increasingly flexible prediction can be combined with efforts at insurance-specific interpretability. General machine-learning research on calibration and distribution shift warns that in-sample predictive success is an insufficient deployment test. Research on fairness metric incompatibility and proxy-label bias cautions against allowing a numerical risk score to stand in for a defended social or contractual objective. The supervisory material from IAIS, EIOPA and NAIC places these technical concerns inside insurer governance and customer conduct.

The synthesis yields three practical inferences. First, every consequential AI-assisted decision should be validated at the junction of model output, business rule and customer effect; evaluating the estimator alone omits the location where harm occurs. Second, auditability requires event-level lineage and actual correction authority, because model-level documentation cannot reconstruct a specific dispute if the operative rule or version was lost. Third, governance should monitor changes induced by deployment itself: selective retention, altered investigation rates, drift in observed labels and changes in customer behavior. These are reasoned interdisciplinary conclusions grounded in existing knowledge; no claim of a new theory or empirical discovery is made.

The evidence has significant limits. Cross-country supervisory texts describe institutional expectations but do not demonstrate compliance or causal effectiveness. EIOPA surveys represent their specific samples and terminology, not the entire global insurance market. The cited actuarial studies do not cover all insurance products, and general fairness, calibration and explainability papers are not substitute insurer-field experiments. Some publisher articles were reviewed through accessible abstracts, metadata and available content rather than independently reproducing methods or accessing all underlying data. The research did not test vendor models, conduct jurisdiction-specific legal advice, estimate protected-class effects in an insurer portfolio, or establish a universal preference for interpretable versus complex predictors.

Future empirical work could test appeal outcomes, selective labels, contract-specific explanation comprehension, catastrophe-drift robustness and cross-vendor concentration using responsibly governed insurer datasets. Those are proposed research questions, not claims that the findings already exist. Legal updates after the November 14, 2026 documentary cutoff require fresh verification at operational adoption or publication.

Conclusion

Artificial intelligence in insurance decision-making is best understood as a governed institutional use of predictive and generative systems within contractual and risk-pooling processes. Model performance is indispensable, but its significance depends on data provenance, the target being estimated, the authority of the business rule, the rights of affected persons and the capacity of the insurer to revise mistaken outcomes. Underwriting, pricing, claims, fraud triage and customer communication each demand decision-specific evidence. A single accuracy statistic or general-purpose ethics declaration cannot establish soundness across these functions.

The reviewed sources establish a common direction across actuarial research, international supervisory guidance and national regulation: identify use and consequences; document data and model lineage; test calibration and shift in realistic conditions; examine legally and ethically relevant group outcomes; ensure meaningful reasons and contestation; and keep the insurer responsible for its own deployed decisions. These requirements acquire substance only when they are observable in actual decision records and incident response.

The major analytical conclusion is that AI assessment should follow the full causal and institutional path from information to customer outcome, including the feedback created by previous automated actions. Better prediction can improve insurance efficiency and coverage, while granularity and errors can redistribute burdens or restrict access. Proper governance makes those consequences visible, reviewable and correctable. The monograph presents this conclusion as a synthesis of verified sources and explicit reasoning. It claims neither a new authorial theory nor an independently measured effect.

References