RESEARCH DOSSIER
New Commercial Risk Classification and Underwriting Uncertainty
Abstract. Novel commercial activities pose an underwriting problem that begins before a credible loss distribution can be estimated. Underwriters must identify the relevant enterprise activity, describe the units exposed, specify the covered event and financial interest, determine the applicable contract wording, evaluate preventive controls, estimate dependence across insureds, and decide whether risk-bearing capacity can support the resulting obligation. Uncertainty accumulates across these tasks. Missing claims are particularly misleading when an activity is new, past events would have been coded elsewhere, insurance coverage has changed, or incident reporting is incomplete. This monograph investigates the formation and revision of commercial risk classes under such conditions through a comparative reading of actuarial scholarship, insurance-market classification practices, supervisory standards, insurance law, and recent evidence concerning cyber, climate, and generative-artificial-intelligence exposures. It distinguishes observed loss, economic exposure, contractual indemnity, statistical estimation, expert judgment, portfolio accumulation, and insurer capacity. The analysis demonstrates that assigning a conventional class code can support administration without establishing empirical homogeneity; that credibility methods can stabilize estimates without creating missing information; and that underwriting decisions depend jointly on probability assumptions, contract interpretation, capital constraints, and the willingness of policyholders to accept exclusions or retentions. An original illustrative calculation makes those dependencies explicit without claiming empirical calibration. The principal contribution is a source-grounded synthesis of the evidentiary decisions and institutional constraints that govern emerging commercial risk underwriting. The research identifies testable questions, disciplined documentation requirements, and the conditions under which broader access to insurance may be supported by improved evidence and appropriately bounded contracts.
Keywords: Commercial insurance; emerging risk; risk classification; underwriting uncertainty; actuarial credibility; ambiguity; expert judgment; model risk; insurability; insurance capacity; contract wording; accumulation risk; artificial intelligence; cyber risk; enterprise exposure; supervisory governance.
Introduction
Commercial insurance translates specified exposures into conditional promises of financial recovery. The translation is difficult when the insured enterprise activity has emerged more quickly than the underwriting categories, loss records, claims practices, or contractual language available to describe it. A manufacturer using a self-learning control system, a cloud-dependent logistics platform, a company licensing generative-AI outputs, and an industrial facility operating a novel energy-storage process may each have familiar physical assets, employees, counterparties, and liabilities. Their particular configurations of activity, dependency, causal loss pathways, and contractual claims can nevertheless be unfamiliar to an insurer. The underwriting problem arises at the intersection of the enterprise's changing operations and the insurer's need to define an obligation before complete loss experience exists.
This monograph approaches classification as a practical epistemic act: an underwriter connects a specified exposure to an administrative reporting category, a portfolio of comparable risks, and a proposed insurance contract. Those connections have different purposes. A code organizes information; a statistical class is meant to support estimation; a contract creates the terms of a possible payment; a solvency assessment determines whether the insurer can sustain that payment in adverse states of the world. Conflating these purposes produces false confidence, particularly when the class has few observations or a shared underlying cause could affect many insured firms.
Internationally, the problem is significant but uneven. The OECD's Global Insurance Market Trends 2026, based on authority-supplied information across more than fifty jurisdictions, records climate, cyber, and geopolitical concerns during 2025; it does not provide a validated micro-level loss distribution for any specific novel business activity. Lloyd's risk-code administration illustrates the coexistence of mature classification infrastructure and ongoing judgment about placement. The IAIS Insurance Core Principles place risk identification, underwriting and capital within the insurer's governance architecture. These sources jointly establish the institutional setting, while their different data populations prevent a simple aggregation of statistics.
The research question is how a defensible commercial underwriting decision can be made when the exposure class is new, claims history sparse, structural change material, and coverage language open to competing interpretations. Four subsidiary questions organize the analysis: which features make risks meaningfully comparable; what can be inferred from sparse or selected observations; how do policy terms change the modeled risk; and how are capital, accumulation and legal constraints translated into offered limits, prices, exclusions and retentions? The method combines document analysis with conceptual comparison, algebraic examples, and source-critical evaluation. The calculation below is explicitly constructed, not estimated from an insurer dataset. No independently collected insurer loss run, confidential policy schedule, claim file, judicial record or full premium quote is represented as having been inspected.
Primary Foundations and Scope
The classification problem and its history
Risk classification has historically served several insurance functions: recording types of business, establishing rating distinctions, monitoring concentrations, and reporting underwriting experience. Its practical necessity does not make a particular classification objectively complete. Lloyd's maintains official risk-code listings and guidance for its market; that guidance is the source for current Lloyd's code selection, not a universal ontology of enterprise risks. The need to use judgment where no narrowly matching category exists is especially revealing. A code may be assigned so that a transaction can be processed while the substantive exposure still requires bespoke analysis.
Risk management standards describe an ongoing process rather than a once-for-all taxonomy. The ISO 31000:2018 guidelines address identification, analysis, evaluation, treatment, communication, monitoring, and integration into governance. The companion IEC 31010:2019 standard concerns selection and application of assessment techniques under uncertainty. Neither standard assigns a premium, fixes a classification for all insurance markets, or resolves a contested insurance contract. Their relevance lies in procedural discipline: selection of an assessment technique must follow the decision and its available evidence.
A newly named class is not necessarily a new cause of loss. It may be a new combination of old hazards, a change in technology, a novel contractual duty, a changed exposure scale, a previously unrecorded aggregation, or a change in how an insurer reports existing business. Distinguishing these cases matters. If an apparently new class merely relabels an established operation, historical loss information may remain relevant after adjustment. If the business changes the dependence among insureds, an old marginal frequency can be seriously misleading even when individual perils look familiar. The correct historical question concerns continuity of the mechanism, not continuity of its label.
Objects, units, and the policy boundary
For research purposes an enterprise exposure is a defined combination of activity, asset or duty, location or network, time at risk, causal event pathway, and potential economic consequence. The exposure unit can be a location-year, machine-hour, revenue unit, active customer, software transaction, insured shipment, or contractual obligation; its selection must correspond to the mechanism being modeled. One annual policy covering many sites cannot automatically be treated as one statistically independent observation. Equally, one reported incident may generate several coverage demands across property, business interruption, professional liability, and cyber forms.
The insurance claim is a separate object. A physical event can produce economic loss even where a policy excludes that cause, requires a different trigger, imposes a waiting period, or attaches above a deductible. Let X denote the enterprise's economic loss and let g(X,W,J,T) denote the indemnity determined by the wording W, applicable legal setting J and effective time T. The insured loss Y=g(X,W,J,T) need not equal X. The function g is a conceptual representation of policy application, not a claim that every dispute can be reduced to arithmetic. Actual recoveries may additionally depend on adjustment, documentation, litigation, insurer performance and payment timing.
A classification designed to estimate X may therefore differ from one used to estimate Y. Enterprise risk managers typically care about total cash-flow disruption; insurers need contract-responsive losses; reinsurers must determine covered aggregation across ceded policies; supervisors assess residual liabilities and capital adequacy. Scientific comparability requires the relevant unit and perspective to be named in every inference.
Risk, ambiguity, and the limits of observed frequency
The classical distinction between measurable risk and deeper uncertainty remains relevant. Ellsberg's analysis of risk and ambiguity illustrates why decisions under incompletely specified probabilities cannot always be represented by the same behavior as choices with well-defined chances. In commercial underwriting, this distinction appears when an insurer can state a point estimate but has little justified confidence in the generating process, or when experts disagree about which events belong in the loss set.
The operational question is not whether a probability may be written down: any number can be written down. It is whether the probability refers to a specified exposure population, observation period, event definition and contract, and whether the model remains credible under plausible changes in those definitions. A low count can reflect a genuinely infrequent event, weak detection, excluded losses, an immature market, or reporting lag. An absence of historical paid claims cannot establish an absence of potential loss. Conversely, very large hypothetical losses should not automatically be interpreted as high expected loss. Frequency, severity, dependence and tail scenarios are different quantities requiring different evidence.
Sources and comparative method
The documentary population reviewed here includes official insurance-supervisory materials, market classification and scenario material, statutory insurance provisions, international risk standards, the OECD's cross-jurisdiction insurance monitoring, and scholarly work on decision ambiguity and actuarial credibility. The IAIS supervisory framework is applied as a reference for insurer governance; its principles are not a directly applicable domestic statute in every jurisdiction. EIOPA's Solvency II Article 45 text provides a specific European regulatory example. The UK Insurance Act 2015 supplies a specific common-law statutory illustration of commercial risk disclosure. The U.S. NAIC artificial-intelligence materials show a different, state-mediated approach to insurer governance.
Evidence is assessed within its stated scope. An institutional guidance webpage can verify the institution's current published position, but it cannot validate a proposed premium for a particular company. A scholarly publisher abstract verifies publication metadata and a stated research argument, not findings hidden in inaccessible pages. A survey of corporate insurance decision-makers indicates reported awareness or purchasing interest among respondents, not the frequency of insured claims. The distinction between source content and analytical inference is maintained in the accompanying machine-readable evidence ledger.
Emerging Enterprise Exposure Classes
The exposure record as a scientific object
A useful exposure description begins with a dated operational map. It records the enterprise's business model, revenues and operational volumes, people and physical locations, counterparties, technologies, security and safety controls, contractual duties, geographic reach, insured interests and dependencies. It then maps potential initiating events to loss mechanisms and potential insurance responses. The object of study is the actual configured operation during the proposed policy period. An industry label alone typically lacks that granularity.
Exposure changes can be discontinuous. A software provider may extend an internal forecasting tool to customer-facing safety recommendations; a manufacturer may shift from selling a product to operating it continuously; a logistics group may centralize activity in one platform. Each change can alter duty of care, outage concentration, accumulation, claim frequency, and aggregation language. Underwriters require versioned exposure information, including the date a new operational function was introduced and the date each control was deployed. A claims series spanning both configurations has to be segmented or explicitly adjusted.
This approach follows the management logic in the ISO risk-management guidelines, while the choice of measurement and risk analysis technique should be proportionate to the data and decision, as emphasized in the IEC assessment-technique guidance. An audit trail should retain the original submissions, source dates, transformations and exclusions. Without that record, the same numerical estimate cannot reliably be reproduced after a renewal, claim dispute or supervisory challenge.
Platform dependence, generative AI, and technology risk
The emergence of commercial generative AI offers a useful contemporary example because the underlying consequences can cross established policy lines. A model output may influence product functionality, professional advice, advertising content, intellectual-property exposure, privacy, security operations, or employee decisions. The Geneva Association's 2025 business survey on generative AI covered 600 corporate insurance decision-makers in six markets and reported substantial interest in insurance for related risks. Its sample supports findings about surveyed attitudes and adoption, not observed insured loss rates, global prevalence or the contractual availability of identical protection in every market.
For an AI-enabled business, the underwriting unit should identify the deployed model or service version, use case, human control, decision consequence, training and monitoring constraints, third-party contracts, and conditions of withdrawal or rollback. A model's inaccuracy might cause a contractual service failure, a bodily injury claim, an intellectual-property allegation or only a reputational loss. These consequences are not interchangeable insured events. NIST's AI Risk Management Framework provides a voluntary framework for managing AI-related risks and trustworthiness. EIOPA's 2025 AI-governance opinion clarifies relevant governance expectations for insurance AI systems under existing sectoral requirements. Neither document warrants a factual claim that all AI-related business losses are covered or that all enterprise AI uses fall within one regulatory category.
Cyber and cloud dependencies illustrate another difficult feature: shared infrastructure. Numerous firms may experience interruption from a single service provider or vulnerability. An insurer with many seemingly independent commercial policies can be exposed to one initiating cause through a portfolio-level network. The NIST Cybersecurity Framework 2.0 resources offer a governance and control reference, while Lloyd's Realistic Disaster Scenarios explicitly include cyber as a catastrophe stress category. The framework and scenarios support examination of controls and accumulation; they do not establish a frequency distribution for a specific cloud outage or a universal policy trigger.
Physical, legal, and organizational novelty
Emerging risk classes extend beyond software. New storage chemistry, automated factories, offshore infrastructure, novel therapies, complex supply chains and unfamiliar service contracts may change failure mechanisms or create legal duties that have little tested claims experience. The novelty may attach to the insured's activity, the claimant's theory, the coverage wording, or a rapidly increasing population of exposed units. A class can appear stable at the level of property peril while liability exposure changes through shifts in product design and reliance by third parties.
Physical-hazard classification is commonly supported by engineering examination of equipment, energy sources, combustible material, maintenance, contingency arrangements, protections and propagation paths. Legal exposure classification requires different evidence: governing law, business representations, professional duties, indemnities and insurance clauses, regulatory permissions, claimant populations, and temporal application of statutes or policy terms. Organizational novelty requires analysis of ownership, vendors, outsourced operations, business-continuity responsibilities, and the practical authority to change controls. Integrating these descriptions improves understanding, yet they should remain separate where evidence and legal consequences differ.
A policyholder may describe a change as commercially innovative; the insurer's question is whether it changes the loss mechanism and the set of contractual outcomes. Some innovations reduce hazard through better monitoring, while introducing service concentration or unfamiliar liability channels. The actuarial comparison must therefore examine the net change in exposure and controls rather than infer an adverse effect from novelty alone.
Administrative codes and underwriting homogeneity
The Lloyd's risk-code documentation reveals that a code belongs to a reporting and processing system. Its guidance is specifically intended for the Lloyd's market. Codes can contain criteria relating to territory, policy form, or class of business. A firm performing several functions may legitimately require multiple risk descriptions even though the outward transaction has one policy number. The historical existence of a code cannot prove that risks assigned to it follow the same distribution.
Statistical homogeneity is a testable proposition about the conditional distribution of losses after relevant covariates have been considered. A class that includes both a standalone software vendor and a cloud service used by thousands of essential facilities may exhibit severe within-class variation. Pooling may improve precision only when the sources of variation are explicitly controlled and the pooled experience is sufficiently transferable. The distinction is crucial: the classification label may be operationally necessary, and the underwriting evidence may simultaneously justify customized terms or a separate statistical segment.
The scientific test of a new commercial class should track changes in frequency, severity, tail behavior, dependence, exposure basis, and potential contract response relative to an explicitly named comparison population. A comparison population selected because its outcomes happen to look favorable creates selection bias. Comparisons should be pre-specified where possible, include unsuccessful and uninsured risks, and record why excluded exposures were treated as noncomparable.
Evidence thresholds and measurement defects
Sparse observations raise at least four separate problems. First, low event counts widen sampling uncertainty. Second, historical events are only informative if their reporting processes and definitions match the target. Third, large claims can arrive or develop long after the underlying event, so immature cohorts understate eventual cost. Fourth, a common-cause loss can create correlation across insureds and across insurance lines. Standard errors based on independent, identically distributed observations can therefore understate material uncertainty.
The absence of paid claims is especially weak evidence where the enterprise has operated briefly, the policy was absent, the deductible absorbed losses, or claims have not been made or resolved. Near misses, operational incidents, regulatory notices, internal loss records, vendor outages, litigated allegations and claims information can all be relevant but have different reporting incentives and data quality. A useful source ledger distinguishes these sources rather than merging them into one observed loss series. The OECD's 2026 insurance-market report offers macro context for the importance of emerging risks; micro underwriting calibration still depends on appropriate granular observations.
Expert Judgment and Model Uncertainty
Frequency, severity, dependency, and coverage
For a defined exposure class, a conventional aggregate-loss representation writes annual gross loss as L = sum of S_i over N events, where N denotes event count and S_i the economic severity of event i. This representation is useful because it separates the event process from the amount lost when an event occurs. It must be expanded when claims from one event are dependent, when annual exposure changes, or when a single event produces losses in several insured interests. The fundamental underwriting question is which L is being modeled: all economic consequences, reported incidents, insured claims at the primary layer, or recoveries under an excess attachment. Each can have a different distribution.
Define the contract-responsive loss as C = h(L,W,T,J,D,M), where W denotes effective wording, T the period, J the applicable legal setting, D a retention or deductible, and M the applicable limit and sublimits. The function h is a convenient analytic description of contract operation, with legal uncertainty represented through alternative admissible interpretations rather than hidden inside one numerical parameter. It is generally false that the expectation of C is obtained by multiplying the expectation of L by an arbitrary average coverage percentage. For example, exclusions and attachment points disproportionately affect different portions of the severity distribution; correlation between the peril and contract response can change results further.
For stable and well-observed classes, an insurer can estimate model parameters from claims and exposure history. For a new class, both the sampling distribution and its underlying data-generating mechanism may be uncertain. It is useful to distinguish parameter estimation uncertainty from structural uncertainty about the model family, and both from uncertainty about which events the contract will respond to. Ambiguity in the severity distribution cannot be resolved merely by adding a contract exclusion; a clearer wording can, however, reduce uncertainty about the insurer's financial promise if its coverage scope is narrowed or precisely defined.
Credibility, pooling, and transferability
Classical actuarial credibility provides tools for combining a limited class-specific experience with broader information. Bühlmann and Gisler's course in credibility theory develops the theoretical basis and practical application of such weighting. A familiar representation is m_hat = Z m_own + (1-Z)m_reference, with a weight Z between zero and one, although any substantive use requires justified definitions of both estimates, relevant variability, the exposure measure, and the model assumptions. The equation is an existing actuarial idea used here for explanation, not a newly proposed underwriting model.
If m_own is derived from a few incidents and m_reference from a larger but structurally different industry, the weighted result can be numerically stable while substantively biased. Credibility does not replace the obligation to show that the reference book is pertinent. Transferability requires comparison of operational functions, technology, geography, policy wording, protection standards, event recording and legal environments. A high exposure volume with almost no covered events is not automatically equivalent to a highly credible zero loss cost; censoring, delayed reporting and undisclosed events can dominate apparent experience.
For emerging classes, partial pooling across appropriately chosen subgroups may be preferable to an undifferentiated market average. Hierarchical or Bayesian methods can explicitly represent class-to-class variation, provided their prior information and assumptions are disclosed and sensitivity-tested. The strength of shrinkage should be an observed or justified modeling choice, not an intuitive guarantee against large losses. A posterior distribution incorporates assumptions about the prior and likelihood. Its narrowness does not establish truth if the relevant event type was absent from the modeled possibilities.
Elicitation, disagreement, and decision incentives
Expert judgment is indispensable when physical tests, exposure data and loss histories are incomplete. It can identify plausible failure sequences, contractual contingencies, correlations and data sources that a purely statistical exercise would miss. Yet experts may have common training, access to the same industry scenarios, market incentives, or shared omissions. Counting several such opinions as independent evidence exaggerates precision. The underwriting file should document each expert's remit, evidence reviewed, assumptions, elicitation method, range estimates and changes following challenge.
Elicitation should ask for well-defined quantities rather than broad intuitions. Asking for the annual chance of at least one specified outage affecting the named insured units is more interpretable than asking how risky the industry is. The former remains uncertain, but it can be checked against incident logs, near misses, vendor dependencies, cyber threat intelligence and historical analogues. Asking experts for a median, lower plausible bound and upper plausible bound can expose disagreement, while the probability assigned to a range needs careful definition. A range without a stated degree of support is a scenario interval, not a statistical confidence interval.
Ellsberg's 1961 ambiguity analysis provides a theoretical warning against treating all uncertainty as if it were known probabilistic risk. In an underwriting committee this matters because the premium can be sensitive to an expert's selection of a particular scenario set. An effective committee records alternative views and the evidence that would favor one explanation over another. A veto based on an undisclosed worst-case intuition cannot be independently evaluated; an unconditional commitment to the cheapest estimate cannot be scientifically justified either.
Commercial incentives complicate the process. The prospective insured benefits from demonstrating mitigation, the broker seeks a marketable submission, the insurer must manage profitability and portfolio concentration, and a reinsurer may use its own aggregates and contract definitions. These incentives do not invalidate evidence, but they determine which data are produced, verified and emphasized. Independent review of critical assumptions is particularly valuable where decision-makers face immediate premium or capacity pressures.
Model risk, validation, and monitoring
Model validation asks whether the methodology fits its decision use, whether the input data and transformations are reliable, whether implementation reproduces intended calculations, and whether outcomes remain credible after deployment. The Federal Reserve's 2026 revised model-risk guidance describes a risk-based approach in banking. It is used here as a methodological comparison, not as an insurance-sector regulation. Its revision of previous banking guidance is important: a current reference should be dated and not mechanically presented as a still-current universal rule from 2011. The IAIS risk-management requirements provide the more directly relevant insurance-supervisory context.
Model testing for a new business line cannot be reduced to one conventional backtest. Rare, high-severity losses provide little near-term performance information. Appropriate checks may include independent calculation, sensitivity to plausible ranges, benchmark comparison with analogues, stress scenarios, exposure reconciliation, and tests of whether model behavior changes reasonably when controls or policy terms change. Monitoring should also identify when the underlying activity changes, when incident reporting improves, or when large losses reveal a previously omitted mechanism.
The emergence of algorithmic underwriting introduces a second distinction: models used to estimate the insured enterprise's risk and models used by an insurer to make underwriting decisions. The insurer's predictive system may be technically sound on historical classes yet extrapolate poorly to a new one. NIST's AI risk-management framework and EIOPA's insurance AI governance opinion support attention to validity, monitoring, human oversight and data governance. The frameworks do not establish that a generative-AI tool can replace legally required underwriting judgment or that regulatory AI categories apply identically in all jurisdictions.
Contract language as a source of uncertainty
Coverage language identifies the obligation an insurer is prepared to undertake. Novel technology can make familiar terms such as direct physical loss, professional service, occurrence, wrongful act, system failure, pollutant, computer system, dependent property, or accidental event contested in particular fact patterns. Their meaning depends on the complete policy, endorsements, temporal context, jurisdiction and applicable legal interpretation. No abstract risk class determines coverage without that analysis. The insurer may understand the engineering event yet remain uncertain whether it would meet the terms of a particular contract.
Contract versions must therefore be attached to underwriting data. If the insurer's historical portfolio insured a narrower trigger, an observed claims ratio cannot be transferred to a broader draft simply by applying ordinary inflation or exposure trend. Likewise, a coverage expansion can increase the value of insurance to a policyholder while increasing ambiguity until its terms and claims procedures become operationally tested. Underwriting should compare the same exposure under alternative wordings, identify their triggering events and limits, and show how each changes retained economic loss and insurer liability.
Commercial disclosure is another legal dimension. In the United Kingdom, Part 2 of the Insurance Act 2015 addresses the duty of fair presentation for non-consumer contracts. The section defines what the insured must present and addresses circumstances known or that ought to be known. This example establishes the importance of information flow, but the statutory duty is jurisdiction-specific. Its application to a particular policy requires the relevant territorial, temporal and case-law analysis. A generic statement about an international duty of disclosure would be materially inaccurate.
A transparent illustrative calculation
Consider a hypothetical business that provides software-enabled scheduling to an industrial network. Assume, solely for demonstration, that one defined event could cause an economic loss of USD 20 million. Analysts disagree whether the annual probability for the specified business configuration is 1%, 2%, or 4%. Under a simplified binary-event model, annual expected economic loss is respectively USD 0.20 million, 0.40 million, or 0.80 million. These amounts are arithmetic consequences of assumptions, not empirically estimated loss costs or proposed insurance premiums.
Suppose a hypothetical contract would indemnify 60% of that same economic loss if the defined event occurs, with no attachment or limit interaction and no possibility of multiple events. Expected indemnity is then USD 0.12 million, 0.24 million, or 0.48 million. If instead the coverage language is ambiguous and the analyst has two legally plausible, illustrative responses—zero payment or payment of 60%—the result depends on the contract interpretation as well as the occurrence probability. Averaging those contract outcomes at an invented 50% chance would conceal an unverified legal assumption. A more transparent approach is to show each interpretation separately and state which wording change would settle the intended contractual response.
The example also shows why a large line size is not equivalent to a large expected loss. The USD 12 million possible indemnity under the covered response is a tail amount; its annual expectation under the assumed 1% event probability is USD 0.12 million. An insurer must also consider capital, reinsurance, administration, expenses, profit objectives, claims uncertainty, event clustering and the ability to pay after adverse developments. An insurer might decline the full limit while accepting a smaller layer, or insist on defined triggers and risk controls. Such decisions cannot be inferred from the expected value alone.
If ten insured enterprises use the same underlying supplier, independence would make a simple portfolio model potentially inaccurate. A common outage could lead to losses on several contracts in one event. The insurer's aggregate tail loss is then driven by the network structure and overlap of insured obligations; multiplying a single-policy expectation by ten gives an expected value only under suitably defined assumptions and says little about correlated tail capital. Lloyd's 2026 realistic-disaster-scenario approach illustrates the institutional importance of stress assessment. This constructed calculation is not fitted to Lloyd's data and does not represent a documented market quotation.
Capacity and Insurability Decision
From technical possibility to market capacity
Insurability is a multidimensional decision. The relevant loss event must be described; contractual boundaries must be intelligible and enforceable; prospective losses must be economically supportable; risk accumulation must be manageable; and the buyer must be willing and able to pay for the offered protection. A risk may be insurable in a narrowly specified layer while broader coverage is unavailable at the requested price or limit. Different insurers and reinsurers can reach different outcomes because their existing portfolios, capital positions, access to reinsurance, expertise, legal interpretations and risk appetites differ.
The capacity decision is distinct from whether a risk exists. A company may face material exposure and obtain no insurance, or obtain partial coverage through a high retention, a sublimit, a parametric trigger, a named-peril policy, facultative reinsurance or a specialized market. A refusal to offer an unrestricted limit is evidence about the insurer's decision under its constraints, not proof that the underlying peril has no measurable distribution. Conversely, an offered price is evidence of a transaction or quote, not proof that the modeled distribution is correct.
The Geneva Association's generative-AI insurance demand findings illustrate the difference between reported interest and demonstrated capacity. A survey can record what firms want to purchase; an underwriting analysis requires the requested terms, prices, available limits and loss evidence. Broad survey estimates cannot substitute for transaction-level observations of quoted, bound, rejected and withdrawn coverage.
Capital, aggregation, and stress scenarios
An insurer's capacity depends on possible losses across its whole book, including simultaneous or sequential events. The IAIS enterprise-risk-management principles require an insurer to identify, measure, report and manage material risks and their interrelationships. In the European Union, Article 45 on ORSA requires assessment of overall solvency needs in light of the risk profile and business strategy. Article 45 also specifically states that the ORSA itself does not calculate a capital requirement. It would therefore be incorrect to describe an insurer's own-risk assessment as an automatic replacement for the Solvency Capital Requirement.
Stress analysis tests exposures that a limited historical frequency dataset may not capture. Lloyd's Realistic Disaster Scenarios include compulsory market scenarios and syndicate-defined events representing material potential losses beyond specified standard events. The method is relevant to novel commercial classes because underwriters can construct an internally coherent loss narrative even where statistical calibration is weak. A scenario is a conditional account of a possible loss; it becomes a probability estimate only with further assumptions and evidence. Stress testing should not be mistaken for proof of a particular recurrence period.
Common technological infrastructure is a distinctive accumulation mechanism. If an insured book has numerous cloud-dependent enterprises, diversity of industry or geography may not imply diversity of failure causes. The same point extends to widely deployed technology, shared suppliers, legal changes and common model vulnerabilities. The 2026 OECD risk and vulnerability comparison documents supervisory concern with interconnected risk categories but does not supply firm-level dependence parameters. A defensible aggregate analysis must make its own exposure mapping explicit.
Reinsurance can increase gross writing capacity, but protection depends on the terms of the reinsurance treaty, its event and aggregation definitions, attachment, exclusions and counterparty performance. Reinsurance should therefore be modeled as a distinct contract layered over the primary insurer's obligations. A ceded portion of expected loss does not imply that all plausible tail losses have been transferred. Capacity reports should distinguish gross exposure, net retention, retrocession where relevant, collateral or liquidity constraints, and stress losses after recoveries.
United Kingdom, European Union, and United States
Comparison across jurisdictions is most reliable when legal rules and supervisory guidance are separated from commercial practice. The United Kingdom's statutory fair-presentation duty creates a specified disclosure framework for non-consumer insurance contracts. It should inform underwriting documentation when applicable, but it does not dictate a universal classification taxonomy or solve every contractual interpretation. The policy and applicable legal version remain essential.
The European Union's Solvency II system makes the insurer's own risk profile relevant to solvency governance. EIOPA's ORSA guidance adds an institutional explanation of the assessment process, while Article 45's legal requirements provide the controlling EU-level text at the page reviewed. Application depends on the effective legal consolidation, transposition, supervisory practice and the particular undertaking. For AI underwriting tools, EIOPA's 2025 opinion addresses existing sectoral obligations for relevant AI systems; it should not be described as a newly created standalone law.
The United States has state-based insurance supervision rather than a single general federal insurance underwriting code. The NAIC's insurance AI governance materials describe the NAIC model bulletin adopted in December 2023 and subsequent supervisory work. A model bulletin is not automatically binding law in every state: implementation and exact requirements must be checked at the relevant state and date. NAIC's enterprise risk-management overview also distinguishes U.S. regulatory mechanisms from direct adoption of IAIS principles. These variations matter when comparing underwriting decisions, reporting obligations and available remedies.
An international insurer should therefore maintain a jurisdiction and contract matrix: issuing entity, insured location, admitted status if relevant, governing law, policy language, local endorsements, data governance obligations, and claims payment mechanisms. One global risk label may conceal different legal promise boundaries. Cross-border underwriting comparisons without these controls risk attributing regulatory differences to the intrinsic novelty of the business itself.
Business risk financing and contractual negotiation
The insured's problem is to choose an economically feasible allocation of risk among retained losses, contractual indemnity, operational mitigation, contingent liquidity, and other financial instruments. A proposed insurance limit should be evaluated against plausible cash-flow needs, balance-sheet resilience, covenant obligations, and duration until recovery. An insurer may offer a low layer against defined events while leaving consequential loss or systemic interruption in the enterprise's retained risk. This can be commercially valuable if the insured understands the boundary and can finance the remainder.
Deductibles and retentions also affect incentives and claims verification. A retention can align the insured with prevention and reduce the insurer's frequency of small claims; however, it can be unusable if a novel activity produces highly correlated losses that overwhelm enterprise liquidity. Higher limits may require multiple carriers, coordination among primary and excess contracts, or new reinsurance arrangements. The insured should distinguish the nominal program limit from the amount that could be collected under a specified event after sublimits, exclusions and attachment provisions.
Evidence improvements can change the bargaining outcome. Verified operational controls, reliable asset registers, incident logs, auditable technology versions, independent safety testing, and a clear incident definition may permit narrower uncertainty margins. Their effect depends on the risk mechanism: excellent controls over an enterprise's own network will not remove a shared upstream dependency. The insurer's premium may remain high where common exposure and capital requirements dominate expected individual claims. The point is an empirical one to be tested against actual quote and loss experience.
Comparative case analysis
Three documentary cases demonstrate different evidentiary limits. First, the Lloyd's classification system shows how a functioning insurance market formalizes reporting categories. Its evidence establishes coding and processing infrastructure; without granular risk and claim records it cannot establish the economic adequacy of a new class's premium. Second, Lloyd's catastrophe stress scenarios show how insurers assess severe potential accumulations; these scenarios are not observations of realized losses at stated frequencies. Third, EIOPA's natural-catastrophe protection-gap dashboard combines economic loss, insured loss and risk information across countries. It demonstrates the importance of distinguishing insured from total economic losses, while its peril and geographic scope does not estimate the premium of an emerging non-catastrophe business line.
A fourth case concerns commercial generative AI. The Geneva Association's corporate survey documents demand-side perceptions among its sampled respondents. The EIOPA AI opinion documents supervisory governance concerns. NIST's voluntary AI risk framework documents a general approach to system risk management. Their intersection can inform underwriting questions, but none establishes the claims experience of the proposed insured class. A complete empirical test would need actual deployed systems, coverage wordings, incidents, claims, limits and prices tied to the same risk population.
These comparisons reveal a recurring problem in commercial insurance research: public evidence is often richest for governance and broad trends, while precise loss and contract data remain confidential. Original analysis can identify this gap, distinguish the inferential steps and test a constructed numerical example. It cannot honestly claim a statistical discovery or actuarial calibration without access to the necessary observations.
Evidence Robustness and Source-Critical Validation
Evidence hierarchy and traceability
A rigorous assessment of novel commercial exposures requires a hierarchy organized by what each record can actually prove. Executed policies and endorsements establish contractual text for a stated period and parties, subject to applicable legal interpretation. Dated exposure schedules and verified operational logs document specified attributes of activity and controls. Claim files can establish alleged and adjusted losses, reserve changes, payments and denial reasons, although claim databases may contain classification or reporting error. Government laws and supervisory texts establish requirements within their effective jurisdiction and date. Market bulletins and industry surveys establish their own published observations or positions. Academic publications establish theories, methods or reported results within their study populations.
These evidentiary categories should be linked, never silently substituted. For example, a vendor incident report does not establish insured loss; an insurance claim payment does not establish the full economic consequence of an event; a public regulatory warning does not provide a policy's governing terms. A transaction-level dataset ideally connects the exposed activity and location, incident, economic consequence, claimant, asserted policy provision, coverage decision, reserve, settlement, and payment date. Linkage itself requires confidentiality controls, common identifiers, entity matching and a record of corrections.
A classified risk should be traceable backward to observable features. The underwriter should be able to explain why an operation was assigned to one class, which alternative classes were considered, and whether the classification changes the insured promise or simply the statistical reference population. The source-critical method follows the original-document emphasis in the Lloyd's risk-code guidance, the risk-process discipline of ISO 31000:2018, and the governance requirements in the IAIS Insurance Core Principles. These are complementary sources with different normative force.
Data comparability and selection bias
Insurance data are selected by application, underwriting decision, attachment of coverage, policyholder risk mitigation, claim reporting and settlement. If only written policies are analyzed, rejected or uninsurable applicants disappear from the observed sample. If only paid claims are analyzed, excluded losses, uninsured events and unsettled claims are absent. When a novel business line has been insured only after stringent engineering reviews, its apparently favorable claims ratio cannot safely be extrapolated to all firms operating the same technology. The observed insured population is conditional on the acceptance rules.
An exposure denominator is critical. A company reporting one incident after one year of operation cannot be compared directly with an enterprise recording one incident across millions of transactions unless the event opportunity and relevant loss mechanism are aligned. For emerging software or platform exposures, transaction count, users, critical operations and financial dependencies each measure different dimensions. For property exposures, replacement value alone may omit time-dependent interruption and concentration in common infrastructure. A carefully chosen denominator increases comparability; it does not erase structural differences.
Loss development is a second comparability problem. Liability claims can take years to emerge and settle, especially where legal duties are newly tested. Observed incurred losses, paid claims, case reserves and ultimate estimates have different information content. An immature class may exhibit short-term paid loss ratios that are lower than its eventual developed cost. Contract form can change the relevant event date, retroactive protection, reporting obligations and claims-made conditions. Any comparison must disclose which development and temporal definition has been used.
The statistical discipline associated with credibility theory and applications is useful precisely because it makes the information in a small class relative to a broader reference explicit. But the researcher must still test whether the reference group is similar enough. A large data set on the wrong activity can be less informative than a smaller, well-described exposure series. This is a problem of external validity and mechanism, rather than sample size alone.
Counterfactuals, sensitivity, and replication
A scientific claim that a classification rule improved underwriting requires a counterfactual: what would the outcome have been under a defensible alternative rule for comparable risks? Common evaluation designs include out-of-time validation, matched cohorts, controlled policy-version comparisons, incident-level causal analysis, and simulated decisions using frozen historical submissions. Each has limitations. Temporal validation may be confounded by market cycles and exposure changes; matching may omit unrecorded risk attributes; simulation cannot establish realized catastrophe outcomes; policy comparisons may differ in legal enforceability.
Performance measures should reflect the decision. Predictive calibration can be tested across bins of predicted frequency where sufficient events exist, but severe tail predictions cannot be validated by short periods without material stress testing. Discrimination measures can assess relative ranking of risks, while a ranking improvement need not imply accurate price level or contract scope. For an underwriting classifier, evidence of fair and reproducible decisions can matter alongside expected loss, yet fairness and legal requirements vary by jurisdiction and protected category. The NAIC's insurance AI governance overview and EIOPA's AI-governance opinion highlight the governance dimension without yielding a single universal classification rule.
Sensitivity analysis should vary at least: exposure size and basis; frequency and reporting completeness; severity and inflation; dependence and common service providers; contract trigger; retention, sublimits and aggregate limit; reinsurance recovery; legal interpretation; and capital stress. Where reliable probability bounds are unavailable, results should be shown as scenario comparisons with transparent assumptions. Ranking one scenario as the most plausible demands evidence. Substituting a point estimate after viewing a wide interval should be recorded as a committee decision, not retroactively described as a statistical inference.
Reproducibility requires durable data-version identifiers, source links, model code or calculation procedures, policy text identifiers, review logs, and a clear separation between original data and engineered variables. An independent analyst should be able to reproduce the numerical calculation with the same assumptions and identify the causes of any discrepancy. Access restrictions may prevent public sharing of confidential insurance records, in which case the published methods should describe the data population, access limitation and verification actually performed. Documentary transparency is possible without disclosing sensitive insured information.
Legal-version and wording controls
Insurance contracts depend on exact language, endorsements, period, territory, choice-of-law rules and the factual trigger of a claim. Legal analysis must retain the effective statutory version and distinguish enacted text, later amendments, supervisory guidance, court interpretation and market practice. The UK Insurance Act's fair-presentation provisions are a specific legislative source, whereas the EU's Solvency II ORSA provisions govern insurer risk management rather than the content of every insured's coverage. Treating these as interchangeable legal requirements would create a category error.
A complete wording comparison should examine insuring clause, definitions, exclusions, endorsements, conditions precedent where applicable, territorial and temporal scope, causation language, treatment of multiple events, defense costs, sublimits, attachment, and claims-notification requirements. These terms determine which economic events may produce a covered claim. Underwriting files often summarize wording changes for business purposes; research must obtain the underlying documents before declaring that one version legally covers a specified contested event. In this monograph, specific hypothetical policy interpretations are described only as constructed alternatives because no signed form for the example was available.
Insurance-policy data are also time-indexed. A claim in one underwriting year may be handled under a form written before the current classification or regulation. Retrospective recoding can support analysis only if the original coding and policy form remain recoverable. The same discipline applies to emerging-risk terminology: the year a category is formally introduced in a reporting system may be later than the year an operational mechanism first appeared. Both dates should be retained when claiming novelty.
Scientific and editorial limits
The study is an analytical documentary synthesis. Its primary-source corpus consists of accessible public standards, laws, supervisory pages and market documentation. Its scholarly foundation includes published work on ambiguity and credibility at the publisher or bibliographic-record access level. It uses the OECD's 2026 international survey as contemporary context and the Geneva Association's 2025 corporate survey as evidence of the attitudes of its respondents. No full unpublished underwriting transaction dataset, insurer-specific rate filing, confidential claims triangle, engineering inspection dossier, judicial coverage file, reinsurance treaty or audit of actual insurance AI models was obtained.
Consequently, the study establishes distinctions, testable dependencies and institutional decision pathways rather than original empirical coefficients or a new priority claim. The model calculation is illustrative. The presence of in-text links establishes the identifiable source for a particular statement, while the evidence ledger states how that source was inspected and what was actually confirmed. Source availability, scholarly metadata, public institutional authority and proof of a particular economic effect remain separate judgments. This discipline is necessary for an international scientific work intended to outlast changes in market terminology and web presentation.
Integrated Research Synthesis and Unresolved Questions
What classification can establish
A useful commercial risk class establishes a reproducible relationship between a described exposure population, an event mechanism, an estimation reference and an administrative purpose. That relationship can be revised with evidence. It has scientific value because it directs the collection of data, the definition of comparison groups and the identification of accumulation pathways. It cannot by itself establish a contractual promise, the probability of a loss, or the capital available to pay it. These are separate claims requiring separate documentation.
The source comparison supports four concrete findings. First, public market-code systems solve essential reporting problems, as demonstrated by Lloyd's risk-code materials, while finer risk comparability must be assessed empirically. Second, actuarial pooling and credibility improve the use of sparse information only when the reference class and observation processes are defensible, consistent with Bühlmann and Gisler's treatment. Third, insurance capacity depends on risk interaction and financial resilience, consistent with the IAIS governance framework and EIOPA's ORSA requirements. Fourth, changes in wording, law or business configuration can invalidate apparent historical comparability even where the class label remains unchanged.
These findings are original in their comparative synthesis and precise application to the designated research question. They are interpretive conclusions based on existing scholarship and sources; they do not constitute a newly created doctrine, proprietary risk model or independent statistical discovery. Their value is in clarifying what a particular document, metric or decision can establish and where the inference must stop.
Joint conditions of workable insurance
Commercial risk transfer becomes more workable when the activity and exposure units are observable; plausible event pathways and controls can be described; data provenance and limits are understood; the coverage trigger and exclusions can be administered; and insurer capital can withstand related losses within the intended limit. These conditions interact. Better data can support estimates without making a broad ambiguous policy affordable. Clearer wording can reduce claims disputes while leaving accumulation risk high. More reinsurance may increase available capacity while introducing dependency on treaty scope or counterparty recovery. Effective underwriting therefore requires a coordinated judgment across evidence, contract and capital.
The analyst should present to decision-makers an internally consistent account of insured and uninsured outcomes. A firm that believes it has insured a new cyber-enabled operational exposure may in fact hold property cover for physical damage, liability cover for specified claims, a separate cyber policy, and a substantial uninsured cash-flow risk. An insurer may possess good marginal claims statistics but inadequate data on common service providers. The difference between the enterprise's total loss and the insurer's actual promise is the central economic consequence of imprecise classification.
Formal governance can improve the quality of decisions by recording assumptions and alternatives. The ISO assessment guidance supports deliberate selection of risk analysis methods, while the NIST AI framework and 2025 EIOPA opinion provide relevant references when algorithmic underwriting itself is used. No single guidance document removes the need to inspect the actual policy, customer, exposure or model. Practical scientific rigor lies in preserving the correspondence among those objects across renewals and later claims.
Implications for institutions and research
Insurance institutions benefit from keeping administrative classification, actuarial classification and policy coverage mapping interoperable but independently auditable. Reporting taxonomies can retain current market codes while recording additional structured exposure descriptors. Submission systems can version operations, controls, event definitions and wording. Claims systems can record both economic incident cause and policy-responsive claim outcome. Pricing and capacity committees can examine loss ranges, dependence scenarios and contractual alternatives without hiding disagreement behind one point estimate.
Policyholders benefit from a clearer statement of the financial consequences of an insured and uninsured scenario. A balanced discussion of enterprise preparedness should include loss prevention, risk finance, claims documentation and dependency mapping. Public institutions and supervisors have an interest in systemic concentration and protection gaps, especially where novel activities become widely shared infrastructure. The OECD international insurance comparison and EIOPA protection-gap dashboard are relevant contextual evidence, but they should not be treated as substitutes for insured-specific analysis.
Research design should prioritize access to linked exposure, incident, policy, quote, claims and capital records. A longitudinal study could track a technology's underwriting history from first deployment through risk-code assignment, wording revision, claims emergence, reinsurance availability and subsequent renewal decisions. Such a design would permit evaluation of whether improved disclosure or narrower coverage reduced price dispersion, rejection rates or later disputes. It would also permit the study of market segmentation and whether high-quality disclosure improves access differently for large and small enterprises. The causal question remains open until comparable data and a defensible identification strategy exist.
Unresolved empirical questions
An important open question concerns identification of true novelty. How can researchers separate the emergence of an unprecedented loss mechanism from new combinations of older processes, changes in coding practice and newly observed incidents? A second concerns the value of expert evidence. Under what conditions do structured elicitations improve predictive calibration, especially when experts share information and correlated errors? A third concerns coverage. How often does new wording reduce real claims disputes, and when does it mainly shift losses back to businesses through exclusions or restrictive triggers?
Further questions concern insurer capacity and market evolution. How do shared cloud providers, AI model components, or technology suppliers transmit loss across apparently diversified commercial books? How does reinsurance respond when systemic events span multiple established lines? What market evidence can distinguish genuine unwillingness to insure from a temporary shortage of capital, legal uncertainty, inadequate submission quality or an unacceptably broad requested limit? How do these results vary across state-based U.S. rules, EU prudential requirements and other international regimes? Available public materials identify the questions and institutions but do not resolve them through a unified micro-level dataset.
Finally, research must evaluate the informational consequences of underwriting itself. Insurers can induce enterprises to document operations and controls, while strict evidence requirements can make new coverage difficult to obtain for smaller firms. Claims and policies then produce selected data that influence subsequent classification decisions. The resulting selection must be modeled in any empirical analysis of long-term market performance. The appropriate objective is better evidenced and more intelligible risk transfer, assessed through actual outcomes rather than the mere appearance of a new insurance product.
Conclusions
New commercial risk classification is a decision about the relationship between an evolving enterprise activity and a set of insurance, statistical and institutional records. The research confirms that this relationship is broader than assignment of a market code. Classification should identify the exposure unit, business configuration, loss mechanism, data provenance, contract boundary, and possible common-cause aggregation before its associated pricing or capacity conclusion is treated as scientifically supported.
Sparse claims experience is informative only within its observation process. Claims that have not been reported, losses excluded by contracts, immature liability development, and differences between historical and current technology can materially change the meaning of a low observed count. Credibility methods assist the combination of evidence, but their validity depends on the transferability of the comparison population. Expert judgment is necessary for emergent hazards and must be documented, challenged and distinguished from measured frequency.
Coverage uncertainty is economically consequential. The same enterprise event can generate different insurer liabilities under different terms and governing laws. The illustrative calculation demonstrates how an assumed event probability and a contract response jointly determine expected indemnity, while portfolio dependence determines an additional dimension of capital exposure. Its numerical results are illustrative arithmetic, not empirical pricing findings. Commercial capacity depends on the insurer's risk portfolio, available capital, reinsurance, stress losses and the requested allocation of risk.
The international documentary evidence reinforces the separation of functions. Lloyd's classification and stress materials establish concrete market practices; IAIS principles and EIOPA rules describe solvency governance; the UK Insurance Act defines a specific commercial disclosure duty; NAIC materials concern U.S. supervisory coordination; NIST standards address voluntary technology-risk management; and the OECD and Geneva Association reports provide carefully bounded market and survey context. None alone establishes a validated rate or a universal legal conclusion for an innovative business. Their integration permits a stronger scientific account of the conditions under which underwriting judgments can be defended.
The principal result is a disciplined research synthesis: class definitions must be reproducible, loss estimates must preserve their source and uncertainty, contracts must be read as operative documents, and capacity must be evaluated across dependent exposures. These conclusions support precise future empirical research and the construction of clearer insurance contracts. The present manuscript makes no claim of a new theory, a statistically calibrated emergent-risk tariff, completed external peer review, or a released public Research Dossier.
References
Bühlmann, Hans, and Alois Gisler. 2005. A Course in Credibility Theory and Its Applications. Universitext. Berlin: Springer. https://link.springer.com/book/10.1007/3-540-29273-X.
Ellsberg, Daniel. 1961. “Risk, Ambiguity, and the Savage Axioms.” Quarterly Journal of Economics 75 (4): 643–669. https://academic.oup.com/qje/article-abstract/75/4/643/1913802. DOI: 10.2307/1884324.
European Insurance and Occupational Pensions Authority (EIOPA). 2015. “Guidelines on Own Risk Solvency Assessment (ORSA).” European Insurance and Occupational Pensions Authority, publication page and listed later file versions. https://www.eiopa.europa.eu/publications/guidelines-own-risk-solvency-assessment-orsa_en.
European Insurance and Occupational Pensions Authority (EIOPA). 2025. “Opinion on Artificial Intelligence Governance and Risk Management.” August 6. https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en.
European Insurance and Occupational Pensions Authority (EIOPA). 2025. “Dashboard on Insurance Protection Gap for Natural Catastrophes.” Web dashboard and accompanying methodology, updated December 5, 2025. https://www.eiopa.europa.eu/tools-and-data/dashboard-insurance-protection-gap-natural-catastrophes_en.
European Insurance and Occupational Pensions Authority (EIOPA). n.d. “Own Risk and Solvency Assessment: Solvency II Single Rulebook, Article 45.” Consolidated online presentation of Directive 2009/138/EC, consulted November 16, 2026. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2128_en.
Federal Reserve Board. 2026. “SR 26-2: Revised Guidance on Model Risk Management.” April 17. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm.
Geneva Association. 2025. “Nine in Ten Businesses Show Interest in Insurance Cover for Gen AI Risks, New Geneva Association Report Shows.” Press release describing Gen AI Risks for Businesses: Exploring the Role for Insurance, October 2. https://www.genevaassociation.org/press-releases/nine-ten-businesses-show-interest-insurance-cover-gen-ai-risks-new-geneva.
International Association of Insurance Supervisors (IAIS). 2024, updated online. Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups. ICP/ComFrame online tool. https://www.iais.org/icp-online-tool/.
International Electrotechnical Commission (IEC). 2019. IEC 31010:2019, Risk Management—Risk Assessment Techniques. 2nd ed. Geneva: IEC. ISO catalog record. https://www.iso.org/standard/72140.html.
International Organization for Standardization (ISO). 2018. ISO 31000:2018, Risk Management—Guidelines. 2nd ed. Geneva: ISO. https://www.iso.org/standard/31000.
Lloyd's. n.d. “Risk Codes.” Market resources, underwriting; current code lists and guidance linked on the market page. Consulted November 16, 2026. https://www.lloyds.com/market-resources/underwriting/risk-codes.
Lloyd's. 2026. “Realistic Disaster Scenarios.” Market resources, underwriting; webpage including 2026 scenario specification. https://www.lloyds.com/market-resources/underwriting/realistic-disaster-scenarios.
National Association of Insurance Commissioners (NAIC). n.d. “Artificial Intelligence.” Insurance Topics, including description of the December 2023 model bulletin and work through 2026. Consulted November 16, 2026. https://content.naic.org/insurance-topics/artificial-intelligence.
National Association of Insurance Commissioners (NAIC). n.d. “Enterprise Risk Management.” Insurance Topics; U.S. supervisory framework overview. Consulted November 16, 2026. https://content.naic.org/insurance-topics/enterprise-risk-management.
National Institute of Standards and Technology (NIST). 2023, updated online. “AI Risk Management Framework (AI RMF 1.0).” Gaithersburg, MD: NIST. https://www.nist.gov/itl/ai-risk-management-framework.
National Institute of Standards and Technology (NIST). 2024, updated online. “Cybersecurity Framework (CSF) 2.0.” Gaithersburg, MD: NIST. https://www.nist.gov/cyberframework.
Organisation for Economic Co-operation and Development (OECD). 2026. Global Insurance Market Trends 2026. Paris: OECD Publishing, November 2. DOI: 10.1787/c8be6585-en. https://www.oecd.org/en/publications/global-insurance-market-trends-2026_c8be6585-en.html.
Organisation for Economic Co-operation and Development (OECD). 2026. “Risks and Vulnerabilities Facing the Insurance Sector.” In Global Insurance Market Trends 2026, chapter 6. https://www.oecd.org/en/publications/global-insurance-market-trends-2026_c8be6585-en/full-report/component-8.html.
United Kingdom. 2015. Insurance Act 2015, c. 4, Part 2, Duty of Fair Presentation, including relevant later effective versions. Legislation.gov.uk. https://www.legislation.gov.uk/ukpga/2015/4/part/2.