Supervisors Probe Insurer Models for Explainability Failures After Complaints of Unexplained Policy Cancellations

By [Staff Reporter]

Who: Insurance supervisors and regulators in the United States and Europe, including state insurance departments, the National Association of Insurance Commissioners (NAIC) and the European Insurance and Occupational Pensions Authority (EIOPA). What: They are stepping up scrutiny of insurers’ use of artificial intelligence and machine-learning models after a wave of consumer complaints and watchdog reports about unexplained policy cancellations and opaque underwriting and servicing decisions. When: Actions and guidance have been issued across 2024–2025 and intensified into early 2026, with formal bulletins, consultations and market surveys published between July 2024 and mid‑2025. Where: The developments span major insurance markets — the United States and the European Union — and reflect cross‑border supervisory concern. Why: Regulators say opaque automated decisioning creates consumer harm risks — wrongful cancellations, discriminatory outcomes, and missed notices — and that insurers must be able to “meaningfully explain” automated outcomes to consumers and examiners. (dfs.ny.gov)

Lead details
Insurance supervisors have moved from exhortation to action, instructing examiners to demand technical documentation, governance records and audit trails for models that touch policy issuance, renewal, cancellation and claims — and pursuing targeted reviews where consumer complaints indicate unexplained or incorrect cancellations. State bulletins and national consultations make clear supervisors expect insurers to document the life cycle of decisioning systems, show human oversight where required, and explain model outputs in ways consumers and examiners can assess. (oci.wi.gov)

What supervisors are doing now

• U.S. state regulators have issued guidance and operational expectations. The Wisconsin Office of the Commissioner of Insurance warned insurers in a March 18, 2025 bulletin that decisions made using AI “must comply with the legal and regulatory standards” applicable to underwriting and policy administration and said examiners may request information and documentation during investigations. The bulletin urges insurers to maintain written AI programs and governance commensurate with consumer harm risk. (oci.wi.gov)

• New York’s Department of Financial Services finalized circular guidance in July 2024 that requires insurers to analyze external consumer data and AI systems for discrimination, demonstrate actuarial validity, and maintain corporate governance and transparency over third‑party AI services. Superintendent Adrienne A. Harris said regulators must ensure “the implementation of AI in insurance does not perpetuate or amplify systemic biases,” signaling enforcement appetite if documentation or safeguards are lacking. (dfs.ny.gov)

• At the European level, EIOPA launched a consultation and a market survey in 2025 to gather data about insurers’ use of generative AI and to set supervisory expectations, explicitly requiring that insurers be “able to meaningfully explain the outcomes of AI systems.” EIOPA’s work is intended to align sectoral rules with the EU AI Act and to equip supervisors to investigate conduct and operational risk tied to automated decisioning. (eiopa.europa.eu)

Regulators’ grounds for action: complaints, surveys and technical warnings
Regulators cite three converging signals: expanding model use, rising consumer distrust and technical limits in explainability techniques.

• Widespread adoption. National surveys conducted or coordinated by regulators show rapid adoption. A NAIC health‑insurer survey published May 20, 2025 found 84% of participating health plans reported using AI/ML in some capacity; earlier NAIC surveys reported similarly high usage in auto and homeowners lines. Regulators say that ubiquity increases the need for consistent governance and supervisory oversight. (content.naic.org)

• Consumer complaints and anecdotal harms. Academics and consumer advocates have documented cases where algorithmic opacity complicated appeals of denials or cancellations; researchers warn that AI may “lock in” poor past decisions if used to replicate human practices without oversight. Consumer lawyers and policy advocates say unexplained cancellations and abrupt nonrenewals have multiplied as carriers automate lapse, nonrenewal and rescission workflows, leaving customers without clear, audit‑ready reason codes and sometimes without timely notice. “AI systems can make decisions based on incomplete or biased data, leading to unfair treatment of policyholders,” said consumer attorney Chip Merlin. (news.stanford.edu)

• Technical limits of explainability. Computer‑science researchers caution that explainable AI (XAI) methods remain immature and can give a “false sense of security” — attribution methods, post‑hoc explanations and saliency tools can be inconsistent, misleading or insufficient to reconstruct a model’s causal logic in high‑stakes cases such as rescission or cancellation. That technical reality underpins supervisors’ insistence on lifecycle controls, audit trails and human‑in‑the‑loop governance. (arxiv.org)

How unexplained cancellations occur — a technical and operational anatomy
Insurance policy cancellations and nonrenewals are operationally complex: they involve billing systems, premium‑collection logic, policy administration platforms, underwriting flags, fraud detection and communications engines. Insurers increasingly layer machine learning scores into triage and automation: lapse‑risk models that flag nonpayment, fraud models that recommend rescission, and churn models that prioritize retention outreach. When these models are poorly instrumented or insufficiently integrated with document evidence and human checkpoints, the result can be an automated cancellation with a terse, opaque notice rather than an audit‑ready explanation. (insurnest.com)

Industry and supervisory observers point to particular failure modes:

• Data mismatch: incomplete billing feeds or failed joins can make a customer look delinquent.
• Proxy discrimination: variables correlated with protected characteristics can produce disparate outcomes even if direct attributes are not used.
• Model drift: scoring systems trained on historical patterns can become misaligned after business changes or market stress.
• Automation gaps: cancellations executed by batch jobs or robotic integrations that lack coherent “reason‑code” mapping to policy clauses produce notices that make little sense to consumers or examiners. (oci.wi.gov)

Supervisory tools and legal levers
Regulators are not merely publishing guidance; they are equipping supervisors with tools and expectations to probe models in market conduct exams.

• Documentation and the “AIS program”: State bulletins and DFS guidance expect insurers to maintain written AI programs (sometimes called AIS programs) describing model purpose, governance, testing for bias and drift, thresholds for human review, and consumer disclosure policies. Examiners can request model inventories, validation reports, training data provenance and vendor agreements during exams. (oci.wi.gov)

• Targeted data requests and reviews: EIOPA’s consultations and surveys are designed to inform supervisors’ ability to undertake thematic reviews — for example, sampling applications, renewal and cancellation files to check whether automated processes applied consistent, lawful reason codes and whether consumers received adequate notice and appeal rights. (eiopa.europa.eu)

• Consumer remedies and enforcement: Where regulators already found improper cancellations or deficient notice, they have ordered remediation, mandated governance changes, or proposed fines under existing unfair‑trade and anti‑discrimination statutes. Supervisors stress that many existing laws are technology‑agnostic: the same consumer‑protection and actuarial‑validity rules apply whether a human or a model makes the call. (oci.wi.gov)

Voices from the sector — insurers, technologists and advocates
Insurers defend their use of automation as efficiency and risk‑management tools but acknowledge the need for better controls. Market participants have pushed back against blanket bans and urged proportionate, risk‑based oversight.

• Regulators’ rationale. “The results show that more and more companies are using AI and are cognizant of applicable state regulations and guidance,” NAIC representative Commissioner Humphreys said in discussing the association’s survey findings. The NAIC has urged insurers to adopt governance consistent with its AI principles and model bulletin. (content.naic.org)

• Consumer perspective. “AI systems can make decisions based on incomplete or biased data, leading to unfair treatment of policyholders,” said Chip Merlin, a Florida policyholder attorney, describing the stakes for households facing unexplained coverage lapses. Consumer advocates say the opacity sometimes compounds harm: cancelled homeowners or auto policies affect mortgage covenants, licensing and access to replacement coverage. (timesofsandiego.com)

• Technical perspective. Researchers caution that explainability tools can be misleading. An analysis of explainability methods argues the field currently lacks standard metrics and that explanations must be assessed for fidelity and usefulness, not only plausibility. Supervisors increasingly demand not only post‑hoc rationales but demonstrable links between model internals, testing regimes and decision records. (arxiv.org)

Business responses and market ripples
The prospect of supervisory scrutiny has prompted several industry moves: insurers are investing in governance, vendors are marketing “explainability” layers, and specialty underwriters have launched products to insure AI performance risk.

• Governance investments. Many carriers have stood up AI governance committees, model‑risk units and independent validation functions; NAIC survey respondents reported testing for bias and model drift and documenting model outcomes. Regulators expect those programs to mature from policy statements to auditable practice. (content.naic.org)

• Third‑party risk and vendor scrutiny. Insurers frequently rely on external models and data providers. Supervisors are directing firms to map vendor models and contracts, and to retain the ability to audit and validate third‑party algorithms — a point regulators emphasize in guidance and exam requests. (oci.wi.gov)

• Insurance for AI failures. The market has begun to price the risk: Lloyd’s and other underwriters in 2025 developed products to cover losses tied to algorithmic errors and chatbot failures, signaling recognition of systemic liability exposures where model errors cause customer harm or litigation. (ft.com)

Legal and regulatory fault lines
Supervisors’ actions reveal unresolved legal questions that will shape enforcement and market practice.

• Explainability standards vs. technical reality. Regulators demand that firms be “able to meaningfully explain” outcomes, but technical researchers warn that many explanation methods do not meet rigorous standards of fidelity. The tension creates a practical challenge: how to set enforceable, meaningful explanation requirements that reflect technical limits and still protect consumers. (eiopa.europa.eu)

• Cross‑jurisdiction convergence. The EU AI Act treats underwriting in life and health as “high‑risk,” imposing strict controls; EIOPA’s opinion and surveys seek sector‑specific supervisory convergence. In the U.S., the state‑based system — NAIC principles, state circulars and exam practices — produces patchwork standards that may evolve through enforcement actions, model bulletins and multistate coordination. Firms operating internationally face overlapping, sometimes diverging expectations on documentation, data provenance, and consumer notice. (mdpi.com)

• Remedies for affected consumers. Because cancellations can have cascading harms — loss of mortgage compliance, inability to obtain replacement policies, exposure to uninsured losses — regulators are examining remediation frameworks: reinstatement protocols, notice corrections, refunds and compensatory payments for identifiable harms where misclassification or procedural failures are found. (oci.wi.gov)

What supervisors are demanding from insurers now (practical checklist)
Regulators’ public materials and consultations point to a converging supervisory checklist insurers are expected to meet for AI systems:

• An inventory of AI systems that affect underwriting, pricing, renewal and cancellation decisions. (content.naic.org)
• Written AIS/AI programs with roles and responsibilities, governance, escalation thresholds and human‑in‑the‑loop criteria. (oci.wi.gov)
• Validation, testing and bias‑mitigation reports, including model‑drift monitoring and champion‑challenger frameworks. (content.naic.org)
• Documentation of data provenance and third‑party agreements; contractual audit rights and access to training data where possible. (oci.wi.gov)
• Consumer‑facing notices and reason codes that link automated outcomes to policy terms and appeal routes in a way examiners can verify. (dfs.ny.gov)

A path forward: balancing innovation, consumer protection and technical realism
Industry and regulators broadly agree on the goal: harness AI’s efficiency while preventing harm. But striking that balance requires technical investments and supervisory realism.

• Better explainability metrics. Researchers advocate for standardized, use‑case‑aware metrics — consistency, fidelity, plausibility and usefulness — to assess whether an explanation genuinely reflects the model’s decision and helps a consumer or examiner evaluate fairness and correctness. Regulators have signaled interest in measurable standards rather than rhetoric. (arxiv.org)

• Operational traceability. Insurers must redesign cancellation pipelines so each automated action emits an auditable reason code tied to business rules and policy clauses; evidence stores should preserve the inputs used for each decision for specified retention periods. Supervisors have made such traceability a focus of exams. (oci.wi.gov)

• Proportionate human oversight. Not all automated decisions require the same level of human review; supervisors are pushing for risk‑based thresholds that ensure high‑impact cancellations and rescissions receive elevated scrutiny. (eiopa.europa.eu)

• Consumer remediation frameworks. Where automation causes erroneous cancellations, firms should have fast remediation and customer outreach pathways — reinstatements, refunds, and documentation to ease transition to alternative coverage if reinstatement is impracticable. Regulators have flagged remediation as an expected outcome of supervisory findings. (oci.wi.gov)

Conclusion — supervision in a field of growing systemic stakes
Supervisors’ shifting posture — from guidance to targeted probes and data collection — reflects a broader reckoning: AI is not a peripheral efficiency tool but a central architecture in modern underwriting and policy administration. As models grow more embedded across pricing, renewal and cancellation workflows, the costs of opaque or faulty automation are not merely individual consumer complaints; they are systemic risks to market conduct, solvency reporting and public trust. Regulators in the U.S. and Europe have signaled they will use examinations, policy guidance and, if necessary, enforcement to ensure that insurers can explain and justify automated decisions in ways that protect consumers and preserve market integrity. For insurers, the imperative is now operational: map, document, test and be able to answer why a customer’s coverage was ended. (oci.wi.gov)

Sources and reporting notes
This article draws on regulatory bulletins and consultations, industry surveys and academic analyses: Wisconsin Office of the Commissioner of Insurance (March 18, 2025 bulletin); NAIC Health AI/ML Survey (May 20, 2025); European Insurance and Occupational Pensions Authority (EIOPA) consultations and surveys (Feb–May 2025); New York Department of Financial Services circular guidance (July 11, 2024); Stanford Health Policy analysis on AI in health‑insurance decisioning (Jan. 6, 2026); academic preprints and evaluations of explainable AI (2024–2025); and reporting on market responses, including specialty insurance for AI errors. Direct quotes in the article are attributed to the sources cited above. (oci.wi.gov)

— End —