RESEARCH DOSSIER
Insurance Generative AI Customer Interaction Conduct Risk
Abstract. Generative artificial intelligence changes the production, distribution, and evidentiary character of insurance communication. An insurer or distributor can generate individualized explanations of coverage, exclusions, premiums, policy changes, claims procedures, and complaints at the moment of interaction. The same capacity creates conduct exposures when authoritative language exceeds the evidence or authority available to the system. This monograph examines consumer-facing chat, voice, and documentary assistants across the insurance lifecycle through comparative documentary research, insurance-contract analysis, and the synthesis of empirical research on comprehension, language-model reliability, customer-service productivity, and conversational persuasion. It distinguishes linguistic error from legally or economically consequential customer harm and distinguishes the reliability of an answer from the adequacy of the entire customer journey. International supervisory materials, European Union instruments, the United Kingdom Consumer Duty, a United States model bulletin, and selected Asia-Pacific supervisory materials are compared according to legal character and scope. The analysis specifies contract-grounded generation, evidence preservation, bounded action authority, accessible escalation, stratified outcome measurement, and incident remediation. Synthetic quantitative examples demonstrate why average accuracy and containment rates can conceal severe or concentrated harms. The principal conclusion is that an accountable insurance interaction requires contractual fidelity, an appropriate distribution function, practicable customer recourse, and evidence of actual comprehension and service outcomes. The study supplies an independently reasoned analytical synthesis and a design for future empirical evaluation; it does not report new consumer trials, proprietary insurer data, or a novel general theory.
Keywords: Insurance conduct risk; generative artificial intelligence; policyholder communication; consumer comprehension; hallucination; insurance distribution; vulnerable customers; retrieval-augmented generation; human intervention; market conduct supervision; complaints; audit evidence.
Introduction
Insurance is an exchange of a present premium and contractual commitments for protection against specified uncertain events. The insured frequently purchases an intangible promise whose value becomes observable only after a loss. A contemporary insurance conversation is therefore economically consequential even when it does not create the contract. It mediates the customer's interpretation of contingent obligations, the selection of protection, compliance with conditions, and the exercise of contractual rights. A generated sentence can alter each of these activities while leaving the actual policy unchanged.
Conduct risk is understood here as the possibility that the insurer's or distributor's behavior produces unfair, misleading, inappropriate, or otherwise harmful outcomes for customers or violates applicable market conduct requirements. Generative AI customer interaction conduct risk is the part of that exposure attributable to the use of systems that synthesize text, speech, or other customer-facing communications rather than exclusively returning an approved fixed response. The relevant system includes model, retrieval infrastructure, interface, data sources, handoff, business rules, staff, and incentives. It is the deployed service, with a defined operating context, that must be evaluated.
The study addresses three questions. First, how do errors in generated insurance language become materially adverse decisions, including omissions by customers? Second, which evidence and operational safeguards permit the insurer to reconstruct, control, and redress these effects? Third, how should supervision and empirical evaluation distinguish assistance, sales steering, and automated exercise of insurance authority? The focus is consumer-facing communication; underwriting algorithm discrimination, autonomous financial advice, cyber underwriting, and general model life-cycle governance are treated only at their interfaces with this subject.
Method and evidence hierarchy
The research is a structured narrative synthesis, not a systematic review with complete database coverage. Sources were selected for institutional authority, direct relevance, methodological specificity, and differentiation between insurance and neighboring domains. The cutoff is November 19, 2026. The evidence ledger records the material actually examined, including indexed text where primary documents could not be opened directly, and abstract-level findings where the full empirical paper was not evaluated. The comparative legal analysis is instrument-specific; it does not establish compliance under every domestic implementation, amendment, or insurance license. Original interpretations are separated from empirical results. Illustrative clauses, sample sizes, rates, and monetary amounts are constructed for analysis rather than observations from insurers.
Official legislation establishes legal propositions within its scope. Supervisory application papers and opinions explain regulatory expectations; model instruments require attention to adoption. A survey establishes reported practices among its respondents. An empirical study supports a conclusion for its studied population and design. A cross-sector finding can motivate an insurance-specific hypothesis, but cannot establish an insurance effect size. This hierarchy prevents the mere multiplication of references from being mistaken for independent evidence. The analysis additionally distinguishes a source's bibliographic identity from the substantive portion accessible for verification.
The research takes place at a historical point where actual adoption is substantial but uneven. EIOPA's (2026) insurance Gen AI survey summarized responses from 347 undertakings across 25 countries; approximately two-thirds reported active use, with substantial activity still at proof-of-concept stage. Reported use cases favored internal productivity, and inaccurate generation was a prominent concern. These are organization-level and use-case observations, not counts of customer journeys or validated claim outcomes. The resulting research agenda is the measurement of risk conditional on specific deployments, rather than extrapolation from aggregate adoption.
Conversational AI in the Insurance Lifecycle
The interaction as an insurance function
A customer-facing system can act as an information directory, language translator, policy explainer, needs elicitor, quote interface, claims intake channel, or complaints handler. These functions overlap in ordinary conversation. The system that starts by explaining what a deductible is may next compare two policies and finally accept a request to purchase. A risk assessment confined to the first turn is inadequate when the conversation can reach the third. Each transition should be treated as a change in function, requisite information, and action authority.
Before purchase, a consumer wants to know whether exposed property, people, activities, or liabilities can be protected and at what cost. General insurance education can be delivered from curated definitions without identifying a customer. A product-specific comparison requires contemporary product documentation, eligibility and territory. An individual recommendation requires appropriate customer information and adherence to the applicable distribution rules. The generation layer can articulate differences, but a regulated recommendation must be attributable to an appropriately governed process. Sales conversion is a business outcome; suitability and informed choice are customer outcomes. Their empirical association may vary, so neither substitutes for the other.
At contract formation, the distinction between a quotation, an application, an offer, acceptance, payment authorization, and evidence of coverage becomes essential. A chatbot may produce plausible confirmation before a policy administration transaction has succeeded. The economically meaningful state is the authoritative record: which offer was accepted, which insured risk and period were bound, and whether any outstanding requirements remain. The linguistic interface should represent the state with the same care as a conventional certificate, while explicitly identifying uncertain or pending events. An action tool should have a separate authorization and result confirmation rather than allowing generated prose to serve as a de facto database write.
During the term, ordinary service includes address changes, beneficiary questions, endorsements, payments, coverage adjustments, and renewal. Each request has a temporal dimension. An answer may correctly describe the product sold today while misdescribing the customer's existing contract. A change requested today may have an effective date tomorrow, or depend on insurer consent. The appropriate reference is the customer's contract version and transaction history, subject to the applicable legal and contractual rules. In particular, nonrenewal, cancellation, and lapse messages demand clear distinction between an informational reminder, a contractual event, and a formal notice.
At loss, a consumer may be under financial pressure and have only one opportunity to preserve evidence or comply with a deadline. A generative system that incorrectly treats an exclusion as conclusive can deter a claim that would otherwise have been lodged. Even a later correction may be ineffective if the customer has disposed of damaged property, made an irreversible settlement, or lost a procedural opportunity. The risk of erroneous denial includes discouraged claims as well as erroneous formal decisions. A claims intake assistant should collect relevant facts, preserve submission time, provide verified procedural information, and route disputed coverage to appropriately authorized assessment.
A complaint can be expressed without technical language. A customer who says that an explanation was false or that they have tried repeatedly to get money owed may be signaling a service failure rather than asking for a fresh product FAQ. Categorization should retain the original expression of dissatisfaction and provide the applicable complaint path. A generative apology is useful only when connected to a service process capable of changing the outcome. The use of a customer-facing assistant therefore affects the discoverability of rights and the firm's ability to learn from adverse events.
Distribution interfaces and operational architecture
A typical system receives a user turn, retrieves context, generates an answer, applies validation and possibly invokes tools. Each stage has a distinct failure mechanism. Identity resolution may select the wrong policy; semantic search may choose a document for the wrong jurisdiction; the model may invent an exception; validation may assess only grammar or toxicity; an execution layer may accept an unauthorized change. Risk ownership requires monitoring each dependency, but the policyholder experiences the combined product.
In retrieval-augmented generation (RAG), the system conditions generation on retrieved external content. Lewis et al. (2020) introduced retrieval-augmented generation and demonstrated improvements on their knowledge-intensive tasks. That result encourages use of a controlled source corpus, but the contractual application adds version and precedence requirements that the original benchmark does not establish. Insurance documents should be indexed by product, carrier, jurisdiction, policy edition, effective date, endorsement, and relevant authority. Semantic similarity is an aid to selecting candidates; contract applicability determines which candidates may support an answer.
The generation environment has a commercial owner, a technology supplier, and organizational users. An intermediary may display an insurer's assistant, and a service provider may supply a model to both. The customer may have no reliable way to infer which party produced the statement. Each institution should map the legal distribution function it performs, ownership of approved material, who controls prompts, and who has power to correct and remediate errors. The IAIS (2025a) AI supervision application paper situates AI systems within proportionate governance and conduct supervision. This supports assessing the actual consequences of a use case rather than categorizing risk solely by the technical sophistication of its model.
Human-in-the-loop assistance and autonomous conversation involve different exposure. In an agent-assist deployment, an employee can verify a draft before sending it. In customer-direct generation, a customer receives the first answer immediately. Brynjolfsson, Li, and Raymond (2025) found that AI assistance increased issues resolved per hour on average by 15% among 5,172 customer-support agents in their study of assisted customer support. Their heterogeneous productivity findings concern a human-assisted workplace deployment. They cannot be treated as an insurance chatbot safety result or a measure of coverage comprehension. The appropriate commercial hypothesis is that savings may be realized while preserving conduct outcomes, subject to a separate insurance evaluation.
Where benefit and cost accrue
A reliable service can reduce repeat explanation costs, shorten queues, and improve accessibility outside conventional hours. It may enable a consumer to ask follow-up questions that would feel burdensome in a call center. These benefits depend on the quality of completion. Replacing a five-minute human conversation with an immediate incorrect exclusion statement transfers the cost to the policyholder and downstream complaints rather than eliminating it. Firms need to identify who receives efficiency gains, who bears the error, when consequences arise, and whether the injury can be repaired.
A synthetic example illustrates this distinction. Suppose a service has 10,000 eligible journeys per month and saves 3 dollars in operating expenditure on each of 6,000 completed automated journeys. Direct savings are 18,000 dollars. Further suppose 20 of those journeys create avoidable harm with mean remediation cost of 1,200 dollars. Direct remediation is 24,000 dollars, before nonfinancial losses, administrative cost, and regulatory exposure. The net direct difference is negative 6,000 dollars under these assumptions. The example is not a forecast; it demonstrates that business efficiency and conduct cost require a joint denominator and that remediation cost is distinct from the customer's full loss. A legitimate insurance benefit paid following correction is not automatically an incremental conduct expense: the relevant loss for the firm includes administrative correction, compensation beyond contractual obligations where appropriate, and other consequences of the erroneous process.
Consumer Understanding and Coverage Communication
Contractual meaning and conversational compression
Insurance policy language is compositional. Broad coverage grants interact with definitions, exclusions, exceptions, schedules, deductibles, limits, endorsements, and legal rules. The operative answer may emerge from their combination. An assistant that extracts a single paragraph can produce a sincerely worded but materially incomplete explanation. Coverage is not a topic similarity score; it is a conclusion about a particular loss under a particular set of operative terms.
In an illustrative homeowners policy, assume the policy covers sudden accidental escape of water from a household plumbing system, excludes groundwater seepage, limits damage from a specified backup event to 10,000 dollars under an endorsement, and carries a 1,000-dollar deductible. A resident asks whether water damage in the basement is covered. The word water alone cannot distinguish the three mechanisms. A responsible information answer must ask about the source of water and identify the relevant wording without representing an unresolved coverage investigation as an approved claim. If the endorsement applies and the covered damage is 7,000 dollars, a simple illustrative calculation subtracting the 1,000-dollar deductible gives 6,000 dollars, assuming the stated limit applies to the payable amount and there are no other adjustments. Real policies may apply deductibles and sublimits in different ways. The exercise establishes a test method, not a universal ordering of calculations.
A policy summary should preserve the relationship between grant and restriction. An answer saying that accidental water damage is covered but failing to disclose the material distinction for groundwater could create a false expectation despite the literal accuracy of the first clause. In the opposite direction, a model may quote a generic exclusion and omit a purchased endorsement. Correctness is assessed at the level of the reasonable understood proposition in context, not only the number of individually true sentences.
Schwarcz, Cude, Logue, and Marquez Alcala (2026) investigated homeowners insurance comprehension in survey experiments. Their published abstract reports that presenting selected policy language improved understanding in some scenarios and decreased accuracy in others, particularly in the presence of a broad grant followed by restricting language. This directly challenges the presumption that exposure to text is sufficient evidence of understanding. It also establishes a research opportunity: assisted explanation could help consumers interpret nested clauses, provided the generated response preserves relevant conditions and does not introduce overconfidence. The underlying effect sizes and sample-level analysis were not independently re-estimated here.
Understanding, reliance, and regulated advice
Understanding means the consumer can use information to distinguish relevant states and consequences. Recognition of a technical word is weaker evidence than correct application to a scenario. A person may repeat that a deductible is 500 dollars yet believe the insurer pays the first 500 dollars of loss rather than the covered amount in excess of the deductible. A useful assessment asks the person to apply the term to a new example, and then distinguishes genuine understanding from recall of the answer provided by the assistant.
Intent matters. General product information describes what a product does. Needs elicitation identifies exposures and priorities. Advice recommends a specific option in the context of a person's demands and needs. In a single conversational channel, the shift can occur without the consumer recognizing the regulatory distinction. Under the EU's (2016) Insurance Distribution Directive, Article 20 specifies demands-and-needs and comprehensible information requirements, with an additional personalized explanation when advice is provided. Article 22 provides scope exceptions and national flexibility. A disclaimer saying that a message is information cannot substitute for evaluating the function actually performed.
The customer may seek certainty exactly where the underlying contract requires evidence. A good explanation should articulate known facts, applicable conditions, unresolved facts, and the next action. A confidence statement should reflect evidentiary sufficiency, not the internal probability that the model assigns to its next words. The model's output distribution is not a calibrated probability that a court or claim department will interpret a contract in the same way. When the question requires a disputed legal judgment, the response should indicate the need for appropriate review and preserve access to that review.
The communication function also includes prices. An initial premium quote may exclude optional coverage, fees, installment costs, or renewal changes. A clear answer should indicate quote assumptions, currency, period, affected product, and whether the quote binds the insurer. Consider two annual plans of 600 and 720 dollars. The cheaper plan carries a material exclusion for the exposure that motivated the inquiry; the difference of 120 dollars in premium is informative only after the coverage difference is established. A conversion-oriented recommendation that foregrounds the premium comparison without the coverage mismatch is an example of an analytically plausible conduct failure.
Accessibility and heterogeneity
Customers differ in language, sensory and cognitive access needs, familiarity with insurance, prior claim experiences, and current circumstances. A recent loss can change the relevant information needs even for a customer who usually manages financial matters independently. An interaction design should therefore allow accessible channels, plain language, code switching or translation where supported, and a human option without requiring the consumer to reveal a medical diagnosis to obtain help.
FCA (2026b) consumer understanding good practice describes use of multiple information channels, testing with real customers, accessible design, and prominent treatment of limitations. Its review of 38 firms covered several financial sectors and is not a representative insurer survey. Its practical relevance is the emphasis on testing what people understand rather than merely reviewing the form of disclosures. IAIS (2025b) similarly connects fair treatment of diverse consumers to ICP 19. In insurance deployment, translation quality ought to be assessed by bilingual subject experts and by customer testing, including numbers, dates, exclusions, conditions, and complaint routes. A linguistically elegant translation may still misstate an insurance concept. If the controlling policy is in a different language, the assistant should clarify the status of the translation and preserve the original.
The empirical comparison must include nonusers and those unable to complete the channel. If a voice assistant performs well among customers whose speech is recognized but repeatedly fails for other accents or speech impairments, aggregate completed-call accuracy overstates equity. Record initiating journeys, recognition failures, retries, accessible alternatives, and effective resolution. Measure errors in semantic interpretation rather than relying entirely on word error rates: a misrecognized negation may be more serious than several inconsequential transcription errors.
Hallucinations, Prompt Risks and Behavioral Steering
Unsupported statements and omission
Confabulation in generated language includes fabrication of a clause, attribution of a term to the wrong source, and assertion of a conclusion beyond the supplied evidence. Autio et al. (2024), in the NIST Generative AI Profile, identify confabulation and human-AI configuration among the relevant risk categories. For insurance, the practical distinction is between the error mechanism and its material consequence. An incorrect statement of an office opening time can inconvenience a customer; an equally short incorrect statement of a claim notification deadline can affect access to protection. The severity is a property of the interaction context and follow-on consequences.
A trace coder should distinguish an unsupported positive claim, a contradicted claim, material omission, wrong applicability, misleading implication, unauthorized commitment, and inadequate escalation. These categories can overlap. For measurement, record each label while predefining a hierarchy for the primary failure mode. A system that says no coverage exists because a retrieved document is for a different product is both an applicability and a possible false denial error. Counting that event twice in a journey-level harm numerator would inflate the overall rate. Retain secondary labels for repair and count unique events for headline estimates.
Retrieval can reduce some fabrications without ensuring correctness. Magesh et al. (2025) evaluated AI legal research tools and reported residual hallucinations in the tested proprietary RAG systems. This neighboring professional domain illustrates why retrieval grounding is not proof of substantive validity. The published rates are tied to the legal products, query set, and evaluation design; they are not transferable percentages for insurance. Insurance testing should independently estimate both retrieval recall for all dispositive policy clauses and answer entailment against those clauses. A correctly retrieved source is insufficient when the model misinterprets it; a semantically true answer can be unsafe if the evidence cited does not establish its applicability.
When policy documents conflict, use a defined precedence process. The policy schedule may specify a limit that modifies a form; an endorsement can alter the underlying wording; a product information document can summarize it without reproducing every condition. The local law may affect enforceability. A generation pipeline should retrieve the relevant source types, communicate the contradiction or uncertainty, and refer unresolved cases. This is distinct from simply ranking longer or newer documents higher.
Synthetic adversarial tests should exercise false premises, multistep prompts, oblique requests, code switching, lengthy chat histories, contradictory attachments, and requests for claims guarantees. An apparent customer document can include text intended to influence the assistant. The instruction priority governing system behavior should be separate from quoted document content. Embedded instructions should not be able to cause a transfer, reveal another customer's data, alter a contract, or override approval rules. In document explanation, the system should treat the attachment as evidence to inspect rather than authority to act.
Persuasive personalization and incentive effects
Generative systems can adapt phrasing to expressed concerns, objections, circumstances, and prior turns. Personalization can improve understanding and reduce irrelevant content. It can also make sales pressure and selective presentation more effective. Salvi et al. (2025) reported in their conversational persuasion study that personalized GPT-4 was more persuasive than human opponents in 64.4% of unequal debate pairs. This was a controlled debate task, not a study of insurance sales, and the percentage excludes ties. The result supports studying persuasive mechanisms, not attributing the reported rate to policy purchase behavior.
The distinction between helpful adaptation and harmful steering is substantive. A customer who indicates a preference for a low premium may benefit from comparison of relevant deductibles but should still receive material information about coverage reductions. A system instructed to maximize sales might adapt by repeatedly emphasizing affordability, omitting exclusions after objections, or framing alternatives as unnecessary. A system designed for claims cost reduction might discourage submission by presenting uncertain exclusions as settled. The IAIS (2025a) discussion of fairness and redress explicitly recognizes risks from exploitation of consumer behavioral tendencies. Conduct supervision therefore needs a review of the incentives embedded in ranking, prompts, reward signals, scripts, and management dashboards.
A paired audit can isolate the effect of irrelevant consumer attributes or vulnerability cues. Prepare two scenarios with identical insured exposure, eligibility and legal facts but different irrelevant language style or asserted financial anxiety. Compare prominence of exclusions, product prices when legitimately held fixed, recommendations, ease of claim lodging, and escalation. A difference may indicate problematic steering; investigate whether it reflects a legitimate accessibility adaptation or an unauthorized decision factor. The test should be repeated across prompts, versions, agents, and channels, with controlled source conditions. Equal outputs are not universally required: an accessible explanation may appropriately use different wording while preserving substantive content and rights.
Persuasive phrasing should be distinguished from misleading factual content. A system can accurately present a product while exaggerating emotional urgency or creating an unbalanced comparison. A second system can be stylistically neutral but factually wrong. Both require conduct evaluation. For example, a consumer with limited available funds asks whether canceling a policy is the only way to lower monthly costs. A neutral response can explain alternatives where verified, clarify the potential protection gap, and refer to an authorized service channel. A sales-directed response that selectively presents cancellation consequences to retain revenue without explaining available options may impair informed choice even if the isolated statements are true.
Model variation and operational concentration
A generated answer can vary across repetitions with the same apparent question because of sampling, retrieved context, conversation history, or supplier changes. Test conditions should record model versions and inference settings to the extent exposed; reproducibility should be defined as the ability to recover the delivered artifact and understand its evidence, not merely recreate identical tokens. An organization that tests an earlier release can obtain accurate answers in the evaluation environment while deploying a changed model or retrieval index. The release boundary therefore includes configuration, connectors, policies, and source editions.
Common dependencies amplify harm. A single corrupted exclusion in a centrally distributed knowledge base can affect thousands of customers with different underlying models. A supplier update may produce correlated errors across several insurers. Severity analysis should include breadth and duration, not only independent error rates. A control that prevents most isolated failures may be ineffective against a shared-source defect affecting every relevant conversation. Change management, source validation, and a rollback mechanism target this dependence.
Consent, Provenance and Interaction Records
Notice and decision authority
Customers need to know the nature of the channel and what it can do. An AI interaction notice serves a different purpose from the disclosure of policy terms, the scope of an intermediary's advice, permission to process data, and authorization to execute transactions. Presenting all these as one generic acceptance obscures distinctions necessary for informed choice. A clear opening notice can identify automated assistance, the available human route, and the extent to which the assistant can act. Additional notices are needed at the actual transition to an application, recommendation, contract change, or claim submission where relevant law and operational design require them.
The EU (2024) AI Act transparency provisions address direct human interaction in Article 50. The provision distinguishes provider duties to design for disclosure, subject to its stated obviousness exception, and disclosure timing at first interaction or exposure. A firm should separately determine whether it is provider, deployer, or both, and the applicable temporal provisions. The study does not infer from this article that every insurance chatbot is a high-risk AI system, nor does it treat an AI notice as permission to provide misleading information. A transparent artificial interface can still produce unfair conduct.
Consent to interact with an assistant, agreement to contractual terms, and authorization to retrieve or modify personal records are separable operations. Consent is not always the appropriate legal basis for data processing. The legal basis, purpose, categories of data, access, third-party transfer, and retention should be assessed under the relevant jurisdiction. Under the EU (2016) General Data Protection Regulation, storage limitation and Article 22's rule concerning solely automated decisions with legal or similarly significant effects are particularly relevant interfaces. This study does not resolve a particular controller's lawful basis or Article 22 exceptions. A chatbot that only provides generic educational information and one that automatically changes rights present different data-protection questions. An apparently human decision cannot be presumed meaningful simply because an employee is named on an organizational chart.
Provenance and trace construction
The delivered message is the evidentiary unit. A usable record contains the actual text or voice output, its time and channel, scope of the customer identity verification, relevant inputs, the applicable contract or product version, retrieved document identifiers and digests, generation configuration where available, tool results, approval state, and subsequent customer actions known to the firm. For a voice channel, preserve the legally permitted transcript and relevant audio or confirmation record where necessary. The purpose is traceability of what occurred; the extent of collection must be proportionate and lawful. Sensitive data may require minimization, restricted access, tokenization, retention controls, or segregation.
A trace should distinguish content shown to the customer from internal drafts. Suppose the assistant drafted a denial, a staff member corrected it, and the corrected answer was sent. The customer outcome is shaped by the delivered message, while the unshown draft remains useful for quality analysis under suitable retention rules. If the incorrect draft was actually sent first and corrected later, both messages and their sequence are relevant. A replay that regenerates the corrected result from an updated index cannot establish what the customer originally saw.
Source provenance requires more than copying a URL. The source authority should be tied to a version and applicable scope. A policy wording PDF can be authentic but irrelevant to the insured's edition. A claims manual can guide internal procedures without being part of the policyholder's contractual terms. A public website summary may change. Store an immutable source identity or lawful archive sufficient to reconstruct the relevant information; label summaries and advisory materials distinctly from contractual text. Avoid retaining excessive customer content solely because it can be retained technically.
The customer-facing explanation should disclose enough source detail to identify the material provision, for example the relevant edition and clause, and provide access where appropriate. A reference should establish the claim it accompanies. Asking the customer to read the entire policy without explaining the source of uncertainty shifts the interpretive burden. Conversely, quoting a clause while omitting an endorsement can create false assurance. The source record permits auditors to check both applicability and completeness of evidence, not only that the text was present somewhere in the source corpus.
Authentication and action boundaries
Public questions should not require account access where no personal information is needed. Account-specific questions require proportionate authentication and authorization. A model should not decide identity on the basis of conversational familiarity or personal facts supplied in text. If a customer asks an assistant to send a policy document, the delivery destination must be verified through an authorized process. A shared household, joint policy, intermediary portal, or claimant representative may create distinct rights of access. A tool call should enforce permissions independently of generated instructions.
For consequential actions, design a sequence of request interpretation, validation against authoritative business rules, customer confirmation, execution, and receipt. A request to reduce coverage should trigger a factual explanation of the material change and relevant timing before confirmation; the exact obligations depend on jurisdiction and contract. A failed transaction should create a clear failure state and route to remedy. Idempotency protects against repeat execution when the customer retries or the system repeats a tool call. Technical safeguards do not determine whether the action is appropriate; they support the authorized decision process.
Data governance includes suppliers that provide hosting, model inference, speech recognition, or retrieval. Contracts should address processing, retention, incident communication, version changes, service continuity, access to evidence, subcontractors, and exit. An audit right that cannot obtain enough information to reconstruct a particular adverse message has limited practical utility. The NIST (2024) profile's provenance guidance identifies recordkeeping, fact checking, supplier considerations, and monitoring actions that can be adapted to these functions. The insurer's design should connect these technical arrangements to complaint and redress workflows.
Supervisory Testing and Human Escalation
Predeployment verification
A proportionate insurance program establishes what the assistant is allowed to do, the evidence required for each function, and conditions under which it must defer. A test corpus should include ordinary customer questions, rare but material coverage combinations, conflicting endorsements, expired forms, renewal terms, language variants, ambiguous or erroneous customer assumptions, claim submission, complaint recognition, and handoff. Domains with significant consequence deserve additional independent review. A customer direct interface should have stricter release expectations than a noncustomer sandbox.
For each test, specify the applicable contract or source edition, jurisdiction, customer state, permissible response, mandatory content, prohibited commitments, and required action or escalation. Gold answers should be reviewed by insurance subject specialists with appropriate legal input when needed. The benchmark needs to accommodate cases with several acceptable phrasings and cases where evidence is genuinely insufficient. Binary exact matching is inappropriate for open-ended explanations, but an expert-coded mandatory condition checklist is reproducible. A valid deferral counts as a success in a question requiring review; an unnecessary deferral counts separately as a service burden.
A two-part assessment first evaluates the output against the evidence and then evaluates the journey. The first asks whether the statement is supported, applicable, complete with respect to material conditions, and free of unauthorized commitment. The second asks whether the customer can reach the correct next step, whether the firm recognizes a claim or complaint, and whether the disposition is recorded. Test both ordinary and adversarial interaction histories. Repeated trials estimate output variability, but repetitions of an identical scenario are clustered rather than independent consumer observations.
Human review should include independently selected samples, disagreements, and adjudication. Reviewers should use a common coding manual, and a subset should be independently double-coded. Report agreement as well as the pattern of disagreements because rare severe defects can yield high raw agreement and low estimated reliability. Model-assisted reviewing is useful for triage; for consequential cases, the reference judgment needs appropriate independent human and source-based assessment. Avoid exposing test answers to the same generation system being evaluated in a way that contaminates the benchmark.
A release decision can use a noncompensatory requirement for specified severe failures and a rate-based threshold for routine performance. For example, a simulated release set might require zero unauthorized cancellation actions, even if routine accuracy exceeds its target. Zero observed failures in a finite test does not demonstrate zero risk. When observations are independent Bernoulli trials, the one-sided exact 95% upper bound with zero failures in n trials is 1 minus 0.05 raised to the power 1/n. At n = 300 this is about 0.994%, and at n = 3,000 it is about 0.100%. Dependence by product, source and prompt reduces effective information, so the formula should not be applied uncritically to repeated generations. The firm should set context-specific risk tolerances and confidence targets rather than treating these hypothetical values as regulatory rules.
Handoff as a completion requirement
A human escalation channel is meaningful when the customer can reach a competent person who can take responsibility for the next action. A link that leads back to the same assistant, a telephone queue beyond the applicable deadline, or a transfer without the relevant context can fail even when the interface records an escalation. The CFPB (2023) identified complaint examples of inaccessible timely human support and repetitive loops in its consumer finance chatbot report. These examples come from adjacent financial services; they illuminate a failure mechanism rather than insurance prevalence or jurisdiction over insurers.
Insurance-specific handoff triggers can include contested coverage, missing policy versions, consequential recommendations beyond the approved scope, customer distress or stated access needs, requests for a human, deadlines, suspected fraud or impersonation, complaints, and execution failures. Certain situations merit an immediate fallback; others can be routed with a specified service standard. The firm should define which teams have authority and how an unresolved transaction is preserved. A transfer record should show intent, evidence used, consent where required, and status of the next process. The receiving employee should independently inspect material evidence rather than treat a model summary as dispositive.
Measure time to competent assistance from the original request, not only the time from the transfer event. A customer may spend ten minutes trying to persuade a chatbot to open a complaint and another ten minutes waiting for an agent. Counting only the second delay understates service burden. If an automated contact is repeatedly interrupted and the consumer initiates a new session, the system should identify recurring unresolved needs within lawful data boundaries. The relevant denominator includes those who gave up.
IAIS (2025a) guidance on effective redress links AI accountability to mechanisms for challenging outcomes, updating information, and seeking human intervention. The analytical consequence is that a human route is part of the service architecture, not an exception detached from system performance. Supervisors can evaluate the entire evidence chain: the initial generated statement, the reason for deferral, actual routing, decision at the receiving team, and correction of process defects.
Incident response, monitoring, and supplier updates
A production program should compare current outcome indicators with version-specific baselines, including material misstatements, missing restrictions, unrecognized complaints, failed actions, delayed handoffs, privacy incidents, and repeat contacts. Use random audits for estimation and risk-targeted reviews for discovery; retain the different sampling weights. A random 500-journey audit can estimate common error rates with uncertainty, whereas a set of complaints may reveal important rare failure modes without providing population incidence. A monitoring dashboard should report both sources rather than combining them into an unweighted rate.
The incident response sequence involves triage, containment, preservation of evidence, identification of affected customers and transactions, evaluation of substantive harm, applicable reporting, correction, redress where due, and validation before resumption. A single confirmed error involving a stale endorsement may justify searching for all conversations with the same source hash and product identifiers. The organization should consider customers who did not complain. The remedy may require correction of the underlying record, reopening a review, restoring a procedural opportunity where possible, or compensation according to applicable law and facts. A changed system prompt alone cannot remedy a historical customer injury.
An external model update, revised product form, new regulation, or changed escalation queue can invalidate a previous test. The release inventory should identify which dependencies changed and which test sets need rerun. A fallback service should remain operable when retrieval, authentication, tools, or external models are unavailable. The service should avoid presenting an outage as a negative coverage decision. Business continuity planning should include the ability to process claims and complaints through an alternate channel.
Comparative Legal and Conduct Regimes
International supervision and legal hierarchy
The international insurance framework treats organizational responsibility and fair treatment as continuing obligations as technology changes. The IAIS (2025a) risk-based supervisory approach differentiates use cases according to their consequences and identifies greater potential risk for systems determining claim payments than for ordinary document retrieval. An application paper supports supervisory implementation of the Insurance Core Principles, rather than itself creating a universally enforceable statute. This distinction matters when comparing domestic duties, supervisory expectations, and recommended technical controls.
EIOPA's (2025) AI governance opinion announcement summarizes a risk-based insurance-sector approach encompassing data governance, records, fair treatment, robustness-related considerations, explainability, and human oversight. The underlying Opinion and direct national application should be checked for implementation decisions. The summary establishes the sectoral relevance of these control domains. A conversational assistant may warrant substantial conduct safeguards even if it does not meet a legal high-risk classification under a horizontal AI statute.
European Union
The Insurance Distribution Directive's Article 20 requirements distinguish information and advice and make the demands and needs of customers relevant to the proposal of insurance. Article 23 specifies information delivery conditions, including rules for media other than paper and website provision. Under the consolidated text inspected for this study, the distribution information requirements cannot be assumed satisfied merely because an ephemeral chat response contains a product summary. The prescribed medium, accessibility, agreement or consent conditions where relevant, and subsequent access matter. Applicable national implementing rules and the product category require individual verification. Article 22's large-risk exemption and national flexibility prevent indiscriminate application across all distribution situations.
The EU Artificial Intelligence Act creates horizontal duties, including for certain systems interacting with natural persons. For classification, intended purpose and applicable provisions are decisive. EIOPA's (2025) summary of high-risk insurance uses points to risk assessment and pricing in life and health insurance. A generic service chatbot is not placed in that category merely because it uses a large language model. If a conversation actually produces risk inputs or decisions for such a purpose, analyze the composite system and its roles. Article 50 transparency and applicable sectoral conduct standards can apply on their own terms, and the relevant effective and transitional provisions must be checked on the deployment date. The consolidated AI Act excerpts were examined at the level stated in the evidence ledger; the present work does not claim a complete current conformity assessment of every AI Act obligation.
Data protection constitutes another distinct regulatory layer. The EU GDPR (2016) automated decision provision covers solely automated decisions having legal or similarly significant effects subject to its own conditions. A generated pre-contract general explanation differs from a workflow that automatically declines an insurance application, and the precise characterization depends on what the service actually does. Information security and vendor governance also require attention under applicable sectoral and horizontal regimes; this work limits its legal conclusions to the expressly inspected instruments and does not offer a complete DORA analysis.
United Kingdom
The FCA Handbook's Consumer Duty rules specify retail consumer outcomes including understanding and support. PRIN 2A.6 extends the relevant customer support scope across channels and outsourced interactions as set out in the inspected version. The firm therefore evaluates the ability of the conversational service to support applicable customers before, during, and after sale. The outcome cannot be established solely by checking that an approved disclaimer appears.
The FCA's (2026b) communication testing review describes practices such as comprehension checks, analysis of complaints and transcripts, and accessible communication. These examples indicate useful ways of evidencing outcomes; the review itself is not an algorithm-specific statutory safe harbor. A firm that reproduces the examples without demonstrating their effectiveness for its products and customers may still have inadequate consumer outcomes. In a regulated insurance interaction, the applicable Handbook perimeter and product-specific rules must also be determined.
United States
Insurance market conduct regulation operates substantially through state law. The NAIC (2023) model AI bulletin reminds insurers that consumer-impacting AI-supported actions and decisions must comply with applicable insurance requirements, including unfair trade practices and unlawful discrimination. A model bulletin is an instrument for state use; it does not establish uniform federal insurance law and should not be represented as binding in every state. A state-specific deployment assessment should identify applicable statutes, regulations, bulletins or guidance, insurer license, product type, role and effective dates.
A practical compliance question is how to retain evidence of a customer-facing representation and its correction when a vendor supplied the model. The applicable legal standard depends on jurisdiction and the facts. At a general analytical level, the firm cannot measure its own communications without access to the messages delivered on its behalf. Contracts and operating arrangements should preserve sufficiently detailed oversight and redress information. The CFPB's adjacent consumer-finance report helps identify potential channel failures but does not establish state insurance requirements.
Singapore and Australia
The Monetary Authority of Singapore (2024) Fair Dealing Guidelines articulate board and senior management expectations concerning fair treatment, suitable products, competent service, information, and complaint handling with proportionate application across financial products and services. The inspected guidance is framed in outcomes rather than an exclusive technical prescription for AI. A generative insurance assistant can be assessed by mapping its functions and communication controls to those objectives, with attention to vulnerable circumstances and the applicable specific regime.
ASIC's (2024) AI governance review announcement describes a multi-sector review of 624 AI use cases across 23 licensees and a subset of interviews. The finding that governance may lag adoption is a supervisory warning, not an estimate of insurance chatbot misconduct. Its analytical relevance is the need to document emerging functions before the service becomes operationally entrenched. Australia-specific legal duties require inspection of the applicable statutes, license conditions and current regulatory instruments. The media release itself is not a substitute for them.
Comparative synthesis
| Framework | Character of inspected source | Principal conduct interface | Boundary of inference |
|---|---|---|---|
| IAIS, 2025 | International application paper and official release explanation | Proportionate oversight, explanations, fairness, redress | Implementation through applicable supervision |
| EU, IDD | Directive and national transposition | Demands and needs, advice, understandable product information | Product, jurisdiction, exemption, and medium conditions |
| EU, AI Act and GDPR | Regulations with their own scopes and timing | AI notice, intended purpose, significant solely automated decisions | Classify actual roles and functions |
| EIOPA, 2025 | Official overview of supervisory Opinion | Sectoral governance and human oversight | Not exhaustive reading of Opinion |
| UK, FCA | Handbook plus supervisory good-practice review | Consumer understanding and support | Retail scope and relevant rules |
| US, NAIC | Model bulletin | Applicable state insurance law and AI-supported conduct | State adoption and specific legal text |
| Singapore, MAS | Supervisory guidelines | Fair dealing outcomes and complaint handling | Institution-specific implementation |
| Australia, ASIC | Official review announcement | Governance readiness and customer risks | Mixed-sector evidence, not legal text |
The comparison reveals a recurring concern with the relationship between organizational accountability, information quality, functioning recourse, and risk proportionality. This is a synthesis across selected materials, not a legal doctrine shared verbatim across jurisdictions. The main conceptual distinction is that the same technical control can serve different legal duties, and a legally required outcome can be approached using several different technical configurations. A general multilingual assistant may thus need jurisdiction- and product-sensitive disclosure modules, an independently validated source corpus, per-function authority, and common incident handling, while the legal basis, distribution scope, and customer rights remain local.
Consumer Outcome Evaluation and Safeguards
Outcomes and estimands
A conduct evaluation should begin with explicit estimands. For a defined product, channel, version, and period, let N denote all eligible initiated customer journeys, A the number receiving an answer, M the number with at least one material misstatement, and H the number with a verified adverse consequence attributable according to the study's adjudication method. The journey-level material error rate is M/N, while an answer-conditional measure may use material erroneous answered journeys divided by A. The first captures the initiated population; the second diagnoses output quality among completed answers. Neither equals H/N, which requires evidence of consequence and attribution.
Define a material error by its capacity to influence a reasonable customer's choice, understanding, timely action or rights in context. Define observed harm separately by evidence such as a delayed claim, unwanted cancellation, avoidable expenditure, inappropriate purchase, missed action, or loss of an opportunity to seek review. Some significant conduct defects are corrected before harm occurs; some consequences remain latent, unreported, or hard to attribute. The study should report confirmed, probable, and unresolved classifications separately using predetermined criteria. Complaint rates are useful but imperfect proxies because access to complaint channels, customer knowledge and willingness to complain vary.
Use separate indicators for accurate and appropriate answers, appropriate deferrals, unnecessary deferrals, successful handoff, time to competent assistance, comprehension, repeat contact, legitimate action completion, and redress. Containment is the fraction of journeys resolved without a human under a defined rule. It measures operational mode, not successful customer outcome, unless the resolution itself is independently validated. If the denominator includes abandoned conversations as automated completions, the measure conflates failure with efficiency. Exclusion of these journeys is equally problematic if abandonment is never reported elsewhere.
For outcomes that unfold after the contact, define a follow-up interval by product and event. A renewal decision can be observed relatively quickly; a coverage misunderstanding may become evident at a future claim. Loss to follow-up should be quantified. If the exposure group has more missing follow-up than the comparison group, a favorable point estimate may be a selection artifact. Under reasonable scenarios, perform sensitivity analysis for the outcomes among missing customers rather than imputing universal success.
Synthetic rates, severity, and aggregation
Suppose a random audit on 1,000 journeys finds 12 with material error among 200 high-consequence cases and eight errors among 800 routine cases. The respective rates are 6% and 1%, and the aggregate is 20/1,000 = 2%. A single aggregate hides the sixfold ratio. These are constructed numbers. The appropriate uncertainty calculation depends on sampling and independence, and the comparison may reflect task composition rather than an attribute of customers. The firm should report stratum counts, confidence intervals, and what makes the high-consequence tasks different.
Now suppose two models are tested on identical weighted compositions of 800 low-consequence and 200 high-consequence journeys. Model A produces eight and 12 material errors; Model B produces 16 and two material errors. A has an overall rate of 2%; B has 1.8%. Assume purely illustrative severity weights of one unit per routine error and 20 units per high-consequence error. Weighted totals are 248 for A and 56 for B: the additive calculation makes the severity distribution visible. These units are illustrative, not observed monetary damages. A serious legal or safety violation may require a separate stop criterion rather than being offset by lower-severity improvements.
| Synthetic sample | Routine errors / 800 | High-consequence errors / 200 | Total errors / 1,000 | Severity units |
|---|---|---|---|---|
| A | 8 | 12 | 20 | 248 |
| B | 16 | 2 | 18 | 56 |
The same calculation can be generalized as an expected incremental cost conditional on relevant states, with each state contribution equal to exposure count multiplied by estimated error probability, probability of harm given error, and conditional consequence. Overlapping harm events must be deduplicated or modeled jointly. Monetary and nonmonetary consequences should be reported separately if conversion to a common metric is not ethically or empirically justified. Correlated defects require scenario analysis rather than relying on independent event assumptions.
A customer-level perspective can differ from the firm's expected cost. If 100 customers experience a six-day delay during a catastrophe, the aggregate expected monetary loss may understate barriers to temporary accommodation or access to necessary services. Vulnerable consumers may bear concentrated consequences. Report delay distributions and tails, subgroup outcomes and severity. A firm-level expected value is useful in resource allocation but insufficient to determine fairness or legal compliance.
Research design for evaluation
A robust evaluation begins with retrospective coding followed by controlled predeployment testing, prospective monitoring, and appropriately ethical comparative experiments. A study might compare human service, human service with AI assistance, and a restricted autonomous assistant using the same verified product corpus and comparable customer tasks. Random assignment of consenting participants to simulated scenarios identifies effects on comprehension and immediate behavior under the test conditions. A field experiment with real consequential interactions would require suitable approval, safeguards, monitoring, and rescue; a lower-quality or risky channel should not be knowingly assigned merely to estimate harm.
Pre-register the primary outcome and guardrails. For example, choose complete correct interpretation of a coverage scenario as the primary measure and rate of severe unsupported advice and failure to escalate as safety outcomes. Secondary measures may include confidence, time, willingness to ask for help, and appropriate choice among synthetic products. Measure calibration by asking participants to rate confidence before comparing it with actual accuracy. A useful assistant could improve average accuracy while also making a small set of wrong respondents unusually confident. Confidence and correctness should be analyzed jointly.
For an illustrative independent two-arm comparison of comprehension rates of 70% and 75%, the standard approximate two-proportion sample formula at two-sided 5% significance and 80% power gives about 1,250 participants per arm, using pooled variance for the null term and separate variances for the alternative term, rounding upward. This is a planning calculation, not an observed effect. Stratification across language and vulnerability groups, clustering by policy, multiple endpoints, nonresponse, and the need to detect rarer harm may increase the required sample. Exact design and analysis should be finalized by the study team before recruitment.
Paired tests can compare answers to the same underlying coverage situation under different communication styles. Because the observations are paired, a binary comparison can examine discordant pairs rather than treating all responses as independent. For live deployment, randomization can occur at a customer or cluster level when justified, with appropriate analysis and protection against cross-arm contamination. A staggered rollout can support quasi-experimental inference only when assumptions about comparability, pretrends, timing, and spillovers are evaluated. A before-and-after decline in complaint count may be caused by product changes, seasonal variation, claims mix, accessibility barriers, or staffing, rather than the assistant.
Representative consumer panels need to include language, disability access, experience, and relevant customer circumstances through voluntary, lawful, carefully designed recruitment. Do not use the assistant's inferred protected traits as unquestioned ground truth. Small subgroup counts should be disclosed and uncertainty reflected; privacy-protective reporting may require aggregation or suppression. Test accessibility accommodations separately so legitimate personalization is not confused with discriminatory pricing or sales steering. Qualitative interviews can illuminate mechanisms that quantitative metrics miss, but cannot supply incidence estimates on their own.
Trace coding should include whether the adverse outcome was generated, staff-mediated, or tool-executed; the actual source edition; the nature and severity of the failure; whether it was corrected; elapsed time; and the customer consequence. A reviewer can compare the delivered communication to the contract and authorized operational state. Causal attribution requires considering alternative reasons for the adverse result, including independent exclusions, customer choices, missing evidence, and external events. A claim incorrectly discouraged by a bot may later prove uncovered; the procedural harm and the contractual entitlement remain separate issues.
A supervisory and organizational evidence program
An insurer should be able to demonstrate the current inventory of customer-facing uses, the allowed action boundaries, relevant product sources, controls and oversight, validation results, customer monitoring, incident history, and redress actions. Supervisors can examine high-risk use cases and common dependencies using proportionate inquiries, independently chosen samples, and actual conversation evidence. The examination should include adverse paths and abandonment, not only curated demonstrations of successful answers. A firm's vendor assessment and technical model evaluations are necessary inputs, but evidence of customer outcome is indispensable.
For documented release control, the firm can organize a dossier per deployment version: permitted use and exclusions, product and jurisdiction authority matrix, responsible business and compliance owners, evidence-source register, release test protocol and outcomes, availability and escalation arrangements, changes and rollback, and monitoring and incident plans. This is an application of ordinary control documentation to a generative service, rather than a claim of new regulatory doctrine. When a system serves several products, shared test evidence can be reused only if applicability and source versions are known. Per-product coverage semantics and customer outcome evidence may remain distinct.
A supplier change should be evaluated through the same lens. Procurement should require notice of material changes where possible, testing access, incident information, evidence continuity and an exit route. Multiple suppliers can diversify dependence, but different models using a single incorrect knowledge corpus can share the same error. Conversely, multiple product knowledge bases can be vulnerable to a common model or template defect. Scope monitoring around both dependency types.
Firms should align employee remuneration and automated success criteria with fair outcomes. A measure of speed alone can encourage premature deferral or an inaccurate confident answer. A balanced evaluation can include contractual correctness, customer understanding, appropriate resolution, and incident correction; severe violations receive noncompensatory controls. The process owner should have authority to reduce automation or suspend an unsafe use even when doing so worsens short-term commercial metrics.
The opportunities are genuine: availability, consistent explanations, assisted research, and reduced routine service effort can support customers and organizational resilience. The research evidence available here supports examining these potential benefits rather than assuming their existence in every deployment. Brynjolfsson et al.'s (2025) results concern assisted agents; the EIOPA (2026) survey concerns organizational adoption. The missing empirical link is the controlled, longitudinal observation of consumer outcomes in actual insurance journeys, conditional on the service configuration, product, language, and customer circumstances. Future work should estimate this relationship with transparent denominators and legally protected data, and compare mitigation alternatives under equivalent tasks.
Conclusion
Generative AI in insurance changes the mechanism through which a policyholder acquires knowledge, makes decisions, and reaches organizational services. The generated explanation often has practical authority even when the system possesses no legal power to amend the contract. Conduct risk arises at the interface between this persuasive and responsive form of communication, the contractual and procedural evidence it conveys, and the actions or inaction that follow.
A distinction between answer accuracy and journey adequacy is central to the analysis. A response can quote a valid clause yet obscure an exception; it can express appropriate uncertainty yet block access to competent assistance; it can execute a valid transaction after presenting an inadequate explanation. These are different failures of the combined service. Insurance-specific source selection must preserve policy edition, scope and precedence. Customer-centered validation must assess both the understandable proposition and actual route to exercising rights. Organizational accountability requires access to the delivered communication, the supporting evidence, the operative authority and subsequent resolution.
International materials converge functionally around proportionate risk assessment, accountability, information quality, monitoring and redress, while the force and precise application of the instruments differ. The EU AI Act, insurance distribution and data protection requirements, the UK Consumer Duty, US state insurance oversight and Asia-Pacific guidance should be mapped separately to functions and jurisdictions. No generic AI disclaimer or technical benchmark can establish compliance with all of them.
The independent contribution of this monograph is the integration of contract semantics, communication mechanisms, consequence-based conduct measurement, evidence traceability, and service process design. The hypothetical examples demonstrate that cost reduction and aggregate accuracy can obscure serious or concentrated harms and that finite test success does not establish universal safety. The proposed research designs make customer comprehension, abandonment, appropriate escalation, harm, and recourse observable. These are analytical conclusions and evaluation methods derived from the reviewed evidence; they are not asserted as a new theory or validated empirical model.
The strongest practical justification for customer-facing generation is its ability to expand access to accurate and actionable insurance understanding while maintaining verifiable contractual fidelity and a functioning path to competent human service. The relevant unit of success is the completed, accountable customer journey. Within the present documentary limits, this is a defensible organizing conclusion for the design, supervision, and subsequent empirical study of generative insurance communication.
References
IAIS. (2025, July 2). Application Paper on the Supervision of Artificial Intelligence. International Association of Insurance Supervisors. Source
IAIS. (2025, July 7). IAIS Publishes Application Paper on Fair Treatment of a Wide Range of Consumers. International Association of Insurance Supervisors. Source
EIOPA. (2025, August 6). EIOPA Publishes Opinion on AI Governance and Risk Management. European Insurance and Occupational Pensions Authority. Source
EIOPA. (2026, February 2). EIOPA Survey on Generative AI Shows Swift but Cautious Adoption Among Europe's Insurers. European Insurance and Occupational Pensions Authority. Source
NAIC. (2023, December). Model Bulletin: Use of Artificial Intelligence Systems by Insurers. National Association of Insurance Commissioners. Source
FCA. (2026, accessed November 19). PRIN 2A: The Consumer Duty. Financial Conduct Authority Handbook, including PRIN 2A.5 and 2A.6. Source
FCA. (2026, March). Consumer Understanding: Good Practice and Areas for Improvement. Financial Conduct Authority. Source
European Parliament and Council. (2016). Directive (EU) 2016/97 on Insurance Distribution. Official Journal L 26, 2 February 2016, 19–59. Consolidated text of 9 January 2024. Source
European Parliament and Council. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal L, 12 July 2024. Consolidated text of 27 July 2026. Source
European Parliament and Council. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal L 119, 4 May 2016, 1–88. Source
Monetary Authority of Singapore. (2024, May 30 revision). Guidelines on Fair Dealing: Board and Senior Management Responsibilities for Delivering Fair Dealing Outcomes to Customers. FSG-G04. Source
Australian Securities and Investments Commission. (2024, October 29). ASIC Warns Governance Gap Could Emerge in First Report on AI Adoption by Licensees. Media release 24-238MR. Source
Consumer Financial Protection Bureau. (2023, June 6). Chatbots in Consumer Finance. Issue Spotlight. Source
Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. National Institute of Standards and Technology. DOI: 10.6028/NIST.AI.600-1. Source
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942. DOI: 10.1093/qje/qjae044. Source
Salvi, F., Ribeiro, M. H., Gallotti, R., & West, R. (2025). On the Conversational Persuasiveness of GPT-4. Nature Human Behaviour, 9, 1645–1653. DOI: 10.1038/s41562-025-02194-6. Source
Schwarcz, D. B., Cude, B., Logue, K. D., & Marquez Alcala, G. (2026). Read But Not Understood? An Empirical Analysis of Consumer Comprehension in Homeowners Insurance. Virginia Law Review, 112, 727. Source
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems, 33. Source
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies. DOI: 10.1111/jels.12413. Source