Before an AI-related exposure is underwritten or a claim is assessed, the parties need to distinguish who operated each system, what it contributed, who suffered harm and which records support that account.
Executive takeaways
- Separate the actor from the insurance interest. Who used AI, what it did and whether the insured suffered a first-party loss or faces a third-party claim are different questions.
- Examine control operation, not merely control design. A documented safeguard may never have operated; a safeguard that operated may nevertheless have been ineffective against the relevant failure mode.
- Follow the evidence without deciding the contract. A technical account can narrow factual uncertainty; legal responsibility and insurance response require separate analysis of facts, law and issued wording.
The decision is what the evidence can actually support
“AI was involved” is not a satisfactory causal account of an incident. It leaves open whether an insured’s AI produced a consequential output, an external actor used AI against the insured, an employee took a decisive action or several causes interacted. It says nothing, by itself, about which insurance contract may respond.
ODA3 Institute proposes a practical discipline for two different moments. Before loss, an enterprise preparing an underwriting submission should identify material AI-dependent workflows, external dependencies, plausible harm pathways and what operational control records are available. After a reported loss, the insured, broker and insurer should separate competing explanations and test them against preserved records. Neither use establishes that the insurer must accept the risk or pay a claim.
For clarity, first-party loss is loss suffered by the insured; third-party liability concerns a claim against the insured by another party. A submission may be an underwriting presentation or information accompanying a notified claim. A brokeradvises on and arranges insurance, subject to its appointment and applicable regulation. An endorsement modifies or supplements the terms of an insurance policy, whether incorporated at inception or added subsequently; it may extend, restrict or clarify coverage.
Harm direction and insurance interest are separate dimensions
Three principal actor-to-affected-party pathways can be examined, with multi-actor sequences potentially combining or overlaying them:
| Illustrative pathway | Possible event | What must not be inferred |
|---|---|---|
| Insured’s AI → insured | Automated pricing or payment error harms the insured | That any first-party policy responds |
| Insured’s AI → another party | AI-assisted service causes alleged customer harm | That the insured is legally liable |
| External actor’s AI → insured or another party | AI-enabled impersonation, fraud or misuse | That the insured’s own AI necessarily failed |
| Multi-actor overlay | External AI, internal AI and human approvals interact | That any single actor caused the entire loss |
These are illustrative inquiry patterns, not a new ODA3 standard or insurance-market taxonomy. They are not mutually exclusive. For instance, AI-enabled impersonation may lead to the insured’s own financial loss; the same operational event may also lead someone else to claim against the insured. Direction of harm identifies potential contributors and affected parties. First-party versus third-party status identifies the insured’s loss or liability interest. They are not interchangeable. A first-party loss does not establish that a first-party policy exists or responds.
The distinction changes which questions may be material to a risk or claims decision, but does not automatically shift a loss from cyber or crime cover to professional or general liability. Each potentially relevant policy has its own insured entities, triggers, exclusions, conditions and limits.
What authoritative sources establish—and do not
The US National Institute of Standards and Technology (NIST) released the voluntary, non-sector-specific AI Risk Management Framework (AI RMF 1.0) on 26 January 2023. It helps frame AI risk-management questions, but does not prescribe an insurance evidence file or establish whether particular controls worked. As checked on 4 November 2026, NIST’s official page states that AI RMF 1.0 is being revised; this article discusses the 2023 edition, not an assumed successor. [1][2]
The Lloyd’s Market Association (LMA) reported an opinion survey conducted in Q2–Q3 2025 with 144 responses overall, 94% from underwriters. The summary analyses four scenarios, each drawing only six to nineteen responses; 144 is not the sample size for any single scenario. The study was conducted in the absence of market underwriting and claims data specifically identifying AI-related exposures and contributions to losses. Its findings indicate respondents’ views, not observed claim frequencies, representative market probabilities or agreed claims-handling practices. [3]
A US wording illustration comes from Alliant Insurance Services’ Insurance Requirements in Contracts manual, version 2026.1, Appendix B. It reproduces three distinct optional Insurance Services Office (ISO) generative-AI exclusion endorsements: CG 35 08 01 26 (products/completed operations liability, printed p. 189), CG 40 47 01 26 (commercial general liability Coverage A and B, p. 190), and CG 40 48 01 26 (Coverage B only, p. 191). Their descriptions and page locations have been checked against publicly indexed manual text, but final visual page inspection remains open. These model forms illustrate why attribution alone is insufficient: contractual treatment may turn on the described harm and its relationship to generative AI, not simply who owned a model. The reproduced forms do not show that any endorsement was attached to a particular issued policy or adopted throughout a market. [4]
The LMA announcement displays 22 January 2026, despite an inconsistent 2025 dateline in its opening text. The survey’s verified collection period is Q2–Q3 2025; this article uses the displayed publication date without treating the conflicting dateline as a separate event. [3]
A payment event does not tell its own causal story
Illustrative scenario — synthetic; no real system, customer, transaction or claim was examined.
A retailer receives an instruction purporting to come from a known supplier and directing payment to a different bank account. A voice message purporting to confirm the change is retained, but whether it was AI-generated remains an unresolved hypothesis. The retailer’s AI-assisted payment-priority tool places the invoice in a review queue, and an employee authorises the transfer.
Three explanations require examination: an external actor deceived the retailer; the internal tool or associated alert workflow materially influenced approval; or ordinary authentication, process or human failures account for the payment irrespective of AI. These explanations can coexist. Original messages, identity checks, model/version and alert records, screen states, approvals, bank instructions, blocked-event records and later interviews have different evidentiary value.
The analysis must answer five distinct questions:
- Did the system produce the recorded output or take the recorded action?
- Did that output or action influence a subsequent system or human decision?
- Did the decision or action contribute to the actual loss?
- What legal responsibility, if any, follows under the applicable law?
- Does the issued insurance contract respond to that loss or liability?
A payment record establishes execution, not its legally material cause. A log may show an AI output without showing that anyone relied on it. Operational contribution does not itself establish legal causation, liability or coverage. Equally, the absence of a recorded AI output does not establish that AI played no material role, especially where records are incomplete or controlled by another provider.
The evidence exchange works in both directions
Insurer → organization. At underwriting, an insurer might ask where AI affects material decisions, what external dependencies exist, who can override a model and what controls have actually been observed. At notification or claim review, the insurer or claims handler may identify the factual uncertainty relevant to an issued policy. These are proposed inquiries, not verified universal insurer requirements.
Organisation → insurer. The insured can provide original records, dates, provenance, observation periods, failed controls and contrary evidence, while identifying vendor-controlled records it cannot access. Evidence that a control was designed does not show that it operated; evidence that it operated does not show it was effective. Control effectiveness must be assessed against the control’s intended purpose, the relevant failure mode, the conditions under which it operated and the outcomes observed.
Evidence retention is subject to storage cost, minimization, applicable law and contractual access. Cloud or model vendors may hold important logs that the insured cannot compel them to preserve or supply. Legal counsel should oversee preservation notices, potential litigation holds and any privilege or confidentiality issues; privilege is not automatic merely because a lawyer is involved. Regulatory notification duties and insurance notice duties must be considered separately. Disclosure duties also depend on jurisdiction and the actual transaction—for example, the UK’s fair-presentation regime for non-consumer insurance—but this article makes no cross-jurisdictional legal assertion. [5]
The following questions illustrate provisional inquiry steps, not a mandated underwriting checklist:
| Decision question | Records to examine | What records could support | What remains uncertain |
|---|---|---|---|
| Who originated the instruction? | Original messages, metadata and authentication records | Preserved communication sequence | Actual actor and AI tool attribution |
| What did internal AI do? | Deployed version, inputs, output and alert records | Recorded system behaviour in the observation window | Influence on human judgement |
| Did safeguards operate and work? | Control tests, approvals, blocked events and exceptions | Design, observed operation and relevant outcomes where tested | Unobserved failures and representativeness |
| What loss followed? | Transfers, recoveries, accounting and claimant records | Recorded transactions and claimed impacts | Legal causation, quantum and allocation |
| What policy questions follow? | Full issued policy, endorsements, notices and facts | Material contractual questions for specialists | Whether coverage ultimately responds |
The table does not replace a policy review. A technically supportable observation is still a different proposition from a legally or contractually established one.
Independent examination requires an explicit boundary
ODA3 Institute’s evidence approach follows Requirement → Control → Evidence → Provenance → Verification → Finding → Bounded Assurance: define what must be established; identify the relevant safeguard; locate records; establish their origin and integrity; test them against criteria; report what they support; and specify the limits of reliance. This is an application of ODA3’s existing evidence philosophy, not a validated insurance-specific standard.
An examiner should declare the system and time window, source completeness, test authorisation, reproducibility, potential contradictory evidence and reasons for any inaccessible records. Later reconstruction must be dated as reconstruction, not presented as contemporaneous observation. An independent examiner can report technical facts; management owns controls and disclosures, while insurers and legal decision-makers retain their respective authority.
The statuses Supported, Partial, Gap and Unknown describe findings from a bounded examination against predefined criteria; they are not insurance ratings.
Supported: The available, verified evidence satisfies the defined criterion within the examined scope.
Partial: Evidence supports identifiable elements of the criterion but does not establish the criterion in full.
Gap: A specified requirement is demonstrably unmet, or a required evidentiary element is established as absent, under the applicable assessment rule. An evidence gap does not, by itself, establish failure of an underlying operational control.
Unknown: The available and accessible evidence does not permit a defensible determination. Missing access, incomplete observation or unresolved provenance may contribute to this status.
Every finding must distinguish the status of the underlying control from the sufficiency of evidence supporting the claim. No formal finding is assigned to the synthetic scenario.
Why this may matter for underwriting and claims
The insurability link is conditional. A bounded causal account may help clarify the risk being presented, identify an uncertainty worth further examination or expose a shared dependency. It does not demonstrate acceptance, pricing benefits, a faster settlement or coverage. Crime/fidelity, cyber, professional indemnity, technology errors and omissions, product liability and general liability may each be relevant for examination, depending on facts and wording; none is presumed responsive.
Common foundation models, cloud services, application components, data suppliers, integration services and updates may connect losses across insured organisations. A compromised shared payment-verification component, for example, could affect multiple users. That is a plausible accumulation pathway, not a portfolio-loss estimate: assessing accumulation would require insured exposure, dependency and credible loss data.
Local financial-sector model governance, information-security and operational-resilience duties can impose recordkeeping or reporting expectations independently of insurance policy wording. Their applicability requires the specific jurisdiction, regulated entity and current controlling rule; no uniform global logging mandate is asserted here. Nor is a single policy or AI assurance exercise a substitute for those obligations.
Executive decisions that follow
| Decision | Accountable owner | Evidence or trigger |
|---|---|---|
| Identify actor, insured-interest and plausible causal accounts before material underwriting or claims submissions | Chief Risk Officer and head of insurance | Bound the claim and preserve competing explanations; revisit if new facts alter the chronology |
| Establish lawful preservation and vendor-evidence access before an incident | CISO, system owner and General Counsel | Versions, logs, exceptions, retention terms, contractual access and legal-hold process |
| Compare the facts with actual issued wording and local duties | General Counsel, insurance lead and broker | Policy, endorsements, notification conditions and applicable rules; revisit on renewal or material system change |
| Commission additional testing only if it can resolve a material uncertainty | Executive sponsor with Risk Committee oversight | Define decision value, likely cost and limitations before scoping examination |
A practical causal-account worksheet
The following is a working discussion aid, not a standard, prescribed insurer submission, certification or evidence rating scheme. Complete it separately for each material scenario or claim:
| Field | Record to complete |
|---|---|
| Decision context | Underwriting submission / notification / claim review; recipient and purpose |
| Systems and actors | System names and versions; operators, vendors, human roles and dependencies |
| Affected parties and interests | Who alleges harm; insured’s own loss versus claim by another party |
| Event chronology | Contemporaneous instruction, output, approvals, execution and recovery records |
| Competing explanations | At least two plausible accounts; evidence that could falsify each |
| Controls | Documented design separately from observed operation and effectiveness |
| Provenance and access | Source, custodian, timestamps, integrity, gaps and vendor restrictions |
| Legal and insurance questions | Refer to counsel and policy specialists; no technical examiner coverage finding |
| Uncertainties and next decision | What is unresolved, who owns the decision and whether more evidence is proportionate |
Illustration only: In the synthetic retailer scenario, the authenticity of the voice message, the extent of the AI tool’s influence and the causal role of the employee’s decision are all unresolved hypotheses. This does not amount to a formal Unknown examination finding because no records were examined against criteria.
The next decision
At the next AI-related insurance renewal, underwriting submission or claim notification, establish four things: which systems and actors were involved; who suffered or alleges harm and what insured interest is affected; which original, corroborating and contradictory records are available; and which material uncertainty remains unresolved. Document what the evidence supports, what it cannot establish, and who is responsible for the next decision. This turns a general assertion of AI involvement into an examinable causal account without treating technical evidence as a determination of legal liability or insurance coverage.
Interest disclosure: ODA3 Institute develops evidence-oriented AI assurance methods and therefore has an institutional interest in the value placed on such evidence. The analytical framework applied in this article — Requirement → Control → Evidence → Provenance → Verification → Finding → Bounded Assurance — is ODA3 Institute’s own. No external validation of the method is claimed.
Notably Absent
No real incident, insured AI deployment, issued insurance contract, notified claim or insurer determination was examined in preparing this article. No verified evidence establishes market-wide insurer adoption of the proposed causal-account approach, improved underwriting acceptance, premium reductions, faster claims settlement, or actual loss accumulation attributable to a shared AI dependency.
The LMA survey provides reported respondent perspectives, not an actuarial loss series or verified claims-handling practice. The reproduced ISO forms demonstrate possible wording structures, not their attachment to any particular issued policy.
The ODA3 evidence method was applied analytically to a synthetic scenario; no operational validation, independent insurer evaluation or formal assurance finding was undertaken.
These limitations describe the evidence examined for this publication. They do not establish that relevant evidence is unavailable elsewhere.
References
- NIST, AI Risk Management Framework. AI RMF 1.0 released 26 January 2023; official page states revision under way. Checked 4 November 2026.
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023.
- Lloyd’s Market Association, Understanding AI Exposures: AI Loss Scenarios Survey Results, based on Q2–Q3 2025 survey (144 total responses; 94% underwriters; four reported scenarios with 6–19 responses each); LMA news announcement displayed 22 January 2026, with an inconsistent 2025 opening dateline. Checked 4 November 2026.
- Alliant Insurance Services, Insurance Requirements in Contracts, version 2026.1, Appendix B, printed pp. 189–191; reproduced ISO generative-AI exclusion forms CG 35 08 01 26, CG 40 47 01 26 and CG 40 48 01 26. Form IDs and printed-page locations cross-checked against publicly indexed manual text on 4 November 2026.
- UK legislation, Insurance Act 2015, Part 2: Duty of fair presentation. UK example only; applicability depends on the insured and transaction.
Methodology note
Source-based analytical review, synthetic scenario, actor/harm-direction comparison and evidence-chain reasoning. No operational test, legal opinion, policy interpretation or insurer validation.
Notably Absent
No real incident, issued policy, insured deployment, actual claim, insurer decision, premium effect, verified insurer adoption, actuarial loss series or portfolio accumulation examined. No direct visual inspection of reproduced Alliant pages established.
