Executive takeaways
- Scope every performance claim before it is relied on: confirm that its task, population, version and period match the deployment.
- Check where AI-related loss sits in your insurance programme: some specialist insurers describe pricing AI underperformance on evidence, while US standard liability forms now allow some generative AI exclusions.
- Assign an owner for evidence after change: a new threshold, data source, permission or population may require reassessment of an earlier finding.
A performance claim travels only as far as its operating conditions
An enterprise deploying artificial intelligence (AI) holds a performance figure from a supplier, pilot or internal test. The Board’s decision is not whether the figure is impressive, but whether it describes the system actually running, on its actual inputs, with its actual authority.
That distinction now has commercial weight. Some specialist insurers describe pricing AI performance risk on evidence. In the United States, standard policy forms now allow insurers to exclude some generative AI exposures. Without records connecting a performance claim to a defined system version, population and operating period, the claim remains insufficiently substantiated for independent examination.
ODA3 Institute’s assessment is that independent examination can strengthen the evidential basis of a performance claim by testing whether it corresponds to deployed behaviour and controls. Whether that examination influences underwriting, pricing or coverage is a separate question.
The subject is insuring losses from an organisation’s use of AI, not insurers using AI in their own underwriting.
What the public record establishes
One reinsurer describes pricing AI performance against agreed metrics and historical evidence. Munich Re, the German reinsurer, describes its aiSure™ offering as insuring the performance of AI models for both AI providers and organisations using their own or purchased models [1]. Its published frequently asked questions (FAQ) page describes two steps: reviewing the model development pipeline for risk scenarios such as unrepresentative training data, data drift (a change in the inputs a model receives after deployment) and weak updating or monitoring; then deriving an estimate of future underperformance from historical performance data [1]. A 2023 whitepaper adds that insurer and client agree the key performance indicators before detailed assessment, with the insurability decision and premium following the insurer’s own review [2]. A 2021 brochure depicts a provider guaranteeing performance to its customer, with Munich Re insuring the provider, and notes that cover is not available in all jurisdictions [3].
The FAQ also states that the offering is model-agnostic, including generative AI, and that tailored covers can pay without requiring a legal liability element [1]. These are the insurer’s descriptions; its estimator, criteria and prices were not examined.
Public examples connect third-party assessment with insurance offerings. Armilla, which Lloyd’s lists as a managing general agent and Lloyd’s coverholder — a firm authorised to bind cover on behalf of insurers in the Lloyd’s market — offers AI liability cover and performance warranties alongside its own validation of model performance [4][5][6]. A published assessment scheme for AI agents describes audits by accredited third-party firms and periodic reassessment [7]. In February 2026, a participating AI developer announced an insurance arrangement associated with certification under that scheme [8]. These are descriptions and announcements by the involved organisations, not independent verification of the underlying assessments, executed policy terms, pricing or claims outcomes. What remains less transparent is how insurers evaluate the underlying audit evidence, which findings influence underwriting decisions, and whether evidence developed outside a particular insurance-linked scheme receives equivalent consideration.
In the United States, standard liability forms now allow AI exclusions. Verisk states that its Insurance Services Office (ISO) 2025 general liability multistate filing addresses generative AI [9]. The form text, reproduced with ISO’s permission by the forms-analysis service FC&S, shows three endorsements with January 2026 edition dates — CG 40 47, CG 40 48 and CG 35 08 — that exclude bodily injury, property damage, or personal and advertising injury arising out of generative AI, which the forms define broadly as machine-learning systems trained on data that create content or responses [10]. The endorsements are optional for insurers to attach [11]. Berkley Insurance Company’s endorsement titled "Artificial Intelligence Exclusion (Absolute)", form PC 51380 (June 2024 edition), amends its directors and officers, employment practices and fiduciary liability coverage parts to exclude claims "based upon, arising out of, or attributable to" the use or development of AI [12]. Insurance Journal reports that Berkley also uses the exclusion in errors and omissions products, and that adoption of the ISO forms remains uncertain [13]. Whether any of these forms appears on a given policy depends on the insurer, the state filing and the renewal. For other markets, the lesson is to read the wording, not to assume the same change.
Technical guidance points the same way. The US National Institute of Standards and Technology (NIST) AI Risk Management Framework, a voluntary 2023 framework, recommends measuring performance in deployment-like conditions, monitoring in production, documenting limits on generalisability, and involving internal experts who did not build the system or independent assessors [14]. It is guidance, not an underwriting requirement.
Insurers need the deployment, not the benchmark
An underwriter’s useful questions concern correspondence, not scores. ODA3 Institute proposes four: what task and population the claim concerns; what counts as a failure and how it is observed; what harm could follow, given the workflow and safeguards; and what evidence connects the examined system to the proposed deployment. Munich Re’s agreement of metrics before assessment [2] reflects the same logic.
Four determinations must stay separate. Technical failure asks whether the system behaved other than intended. Financial loss asks whether that caused a quantifiable loss, and to whom. Legal liability asks whether someone is responsible under the applicable law or contract. Insurance coverage asks whether a specific policy responds under its wording. One does not establish another: an intercepted error causes no loss, while a small error in a workflow with authority to act can be costly. A performance cover may even pay without liability being established [1].
Organisations can make a claim traceable, or leave it as an assertion
Management owns the system, the controls and the evidence. To substantiate rather than assert, an organisation states the version, evaluation population, sampling method, how labels were set and disputes resolved, the failure definition and the operating period. It discloses inputs left out of the evaluation, and whether observations came from a laboratory test, a pilot or production.
An independent examiner can then assess whether that evidence exists, attaches to the deployed configuration and supports the result, distinguishing documents showing a control was designed from records showing it operated. Neither guarantees performance beyond the sampled cases. Insurers make their own underwriting and pricing decisions; Munich Re, for example, describes conducting its own technical review [2]. Coverage depends on the contract, facts and applicable law; contested interpretations may be adjudicated. Receiving an external report, considering it, relying on a specific finding, allowing it to influence terms, or substituting it for internal due diligence are distinct forms of reliance.
| Executive or underwriting question | Evidence to examine | What it can show | What remains uncertain |
|---|---|---|---|
| Does the evaluation reflect this deployment? | Dataset provenance, sampling method, comparison with production inputs | Whether tested inputs resembled live ones when examined | Future inputs; cases never sampled |
| Are reported outcomes dependable? | Label definitions, adjudication records, disputed examples | Whether outcomes came from a stated, repeatable process | Complete ground truth; rare events |
| Is the deployed system the evaluated system? | Version, configuration, dependency and release records | Correspondence between evaluated and deployed builds | Unrecorded modifications |
| Do safeguards contain errors? | Escalation tests, fallback and human-review records, action logs | Whether controls operated in examined cases | Untested conditions; actual loss severity |
| Which changes call for reassessment? | Change log covering thresholds, data sources, permissions, population | Whether changes are detected and reassessed | Whether a future change is material before it is assessed |
These are evidence questions, not a submission format or a rating scheme.
System changes can require re-examination of earlier findings
An AI performance finding is bounded by the system configuration, operating population, evaluation conditions and observation period examined. A subsequent change does not automatically invalidate the finding; it may, however, undermine the proposition that the evidence originally supported.
Changing a decision threshold may alter observed error rates. Replacing a retrieval source may affect the relevance of earlier evaluations. Expanding tool permissions can introduce consequences that were not possible under the tested configuration.
Management should maintain a traceable record of material changes, identify which earlier conclusions could be affected and determine what needs renewed examination. An independent examiner can assess the resulting evidence but cannot guarantee behaviour under unexamined conditions.
Whether a change also creates a disclosure or notification obligation to an insurer depends on the applicable contract and law. Operational reassessment and contractual notification are separate decisions.
Illustrative scenario — synthetic, not based on an examined system
A document-classification model routes applications into queues. The supplier reports high overall accuracy, but uncommon documents cause the most consequential errors. An examination would compare aggregate and subgroup results, inspect escalation of uncertain outputs and test the fallback queue. A bounded conclusion might be that the safeguard operated for tested cases in the examined configuration, with small subgroups disclosed. If the model later sends decisions directly to applicants, the routing evaluation no longer establishes that the new authority is controlled. No real supplier, policy or ODA3 Institute examination is represented.
What leadership should decide, and who owns it
Retained risk may be larger than assumed if exclusions have been added; evidence may need re-examination after change; and coverage turns on wording, not expectation.
| Action | Owner | Decision or evidence needed |
|---|---|---|
| Scope each performance claim used with customers, the Board or insurers | Chief Risk Officer, with the system owner | Task, population, version and period per claim. Would change if supplier inputs cannot be matched to production: treat as supplier-reported |
| Map material AI uses to each policy and identify AI exclusions | Head of insurance or risk transfer, with General Counsel | Current liability, professional, directors and officers, and cyber wordings. Would change if no wording addresses AI loss: escalate as potentially uninsured |
| Define changes that trigger reassessment of earlier findings | Chief Information Security Officer or accountable AI system owner | Change log and escalation route. Would change if version records are missing: present no earlier finding as current |
| Ask the insurer, in writing, what it would examine and what it would conclude | Head of insurance or risk transfer | Recorded response. Would change if the insurer will not consider external evidence: justify the programme on operational grounds alone |
Notably Absent
The following were not located in the sources checked on 15 October 2026:
- A public account of how an insurer uses third-party AI audit findings in underwriting, or whether evidence produced outside an insurance-linked scheme receives equivalent consideration.
- A documented instance, within the sources examined, showing precisely how independent examination changed an underwriting decision, terms, exclusions, limits or price.
- Public data on the frequency or severity of claims paid under AI performance policies.
- A published insurer method for assessing agents that act through tools, despite Munich Re’s model-agnostic statement [1].
- A court ruling interpreting the cited ISO generative AI endorsements or Berkley exclusion within the legal and insurance sources examined; no comprehensive court-database search was performed.
ODA3 Institute examined publicly reproduced exclusion endorsement wording but did not examine any complete issued insurance policy, customer system, underwriting file or claims file. Whether these endorsements were attached to any particular policy was not established.
The decision the evidence now enables
Boards can require every material AI performance claim to state its evidence boundary before it is used for approval, guarantees or insurance discussions. Independent examination can make such claims more traceable and testable against deployed system behaviour. That does not establish whether an insurer will accept the findings or change its decision. The consequential question is what evidence an insurer actually relies on, for which decision, under which conditions, and with what remaining uncertainty.
References
[1] Munich Re. "Frequently Asked Questions — Insure AI (aiSure)." Munich Re, publication date not stated. https://www.munichre.com/en/solutions/for-industry-clients/insure-ai/faq.html (accessed 15 October 2026).
[2] Munich Re. "De-Risking AI Ventures: How Munich Re Assesses AI Performance Risks." Munich Re, 2023. https://www.munichre.com/content/dam/munichre/contentlounge/website-pieces/documents/MunichRe-De-Risking-AI-Ventures-Whitepaper.pdf (accessed 15 October 2026).
[3] Munich Re. "aiSure™ – Implementing Artificial Intelligence with Confidence." Munich Re, 2021. https://www.munichre.com/content/dam/munichre/contentlounge/website-pieces/documents/aiSure-Flyer_2021.pdf (accessed 15 October 2026).
[4] Armilla AI. "Affirmative AI Coverage." Armilla AI, publication date not stated. https://www.armilla.ai/coverage (accessed 15 October 2026).
[5] Mixides, T. "Armilla reveals purpose-built AI liability insurance amid rising legal and regulatory pressures." Reinsurance News, 9 May 2025. https://www.reinsurancene.ws/armilla-reveals-purpose-built-ai-liability-insurance-amid-rising-legal-and-regulatory-pressures/ (accessed 15 October 2026).
[6] Lloyd’s. "Armilla AI." Lloyd’s Lab Accelerator alumni profile, publication date not stated. https://www.lloyds.com/insights/lloyds-lab/programmes-and-initiatives/lloyds-lab-accelerator/alumni/armilla-ai (accessed 15 October 2026).
[7] Artificial Intelligence Underwriting Company. "AIUC-1." AIUC, publication date not stated. https://www.aiuc-1.com/ (accessed 15 October 2026).
[8] Mancini, M. and Kvist, R. "ElevenLabs secures first-of-its-kind AI Agent insurance." ElevenLabs, 12 February 2026. https://elevenlabs.io/blog/aiuc-announcement (accessed 15 October 2026).
[9] Verisk. "From Risk to Endorsement: Four Key Emerging Risks Shaping the Latest ISO General Liability Multistate Filing." Verisk, July 2025 (date inferred from URL; not stated on page). https://core.verisk.com/Insights/Emerging-Issues/Articles/2025/July/Week-4/Emerging-Risks-in-ISO-General-Liability-Multistate-Filing (accessed 15 October 2026).
[10] Sorrell, K. "General Liability Endorsements — Assault or Battery, Generative AI, Human Trafficking." FC&S (PropertyCasualty360), 6 October 2025. Includes copyrighted material of Insurance Services Office, Inc., with its permission. https://www.propertycasualty360.com/fcs/2025/10/06/general-liability-endorsements—assault-or-battery-generative-ai-human-trafficking/ (accessed 15 October 2026).
[11] Germond, N. "Verisk to Roll Out New General Liability Exclusions for Generative AI Exposures." Independent Insurance Agents & Brokers of America (Big I), 21 October 2025 (later updated). https://www.independentagent.com/vu_resource/verisk-to-roll-out-new-general-liability-exclusions-for-generative-ai-exposures/ (accessed 15 October 2026).
[12] Berkley Insurance Company. "Artificial Intelligence Exclusion (Absolute)." Endorsement form PC 51380 00 (06-24). Copy published by Hunton Andrews Kurth LLP. https://hunton.sitepilot11.firmseek.com/assets/htmldocuments/noindex/PC-51380-00-06-24-Artificial-Intelligence-Exclusion-Absolute.pdf (accessed 15 October 2026).
[13] Jergler, D. "Insurer Interest in AI Exclusions Growing as Risk Becomes Omnipresent." Insurance Journal, 22 July 2026. https://amp.insurancejournal.com/news/national/2026/07/22/878480.htm (accessed 15 October 2026).
[14] National Institute of Standards and Technology. "Artificial Intelligence Risk Management Framework (AI RMF 1.0)." NIST AI 100-1, January 2023. https://doi.org/10.6028/NIST.AI.100-1 (accessed 15 October 2026).
© 2026 ODA3 Pvt Ltd.
Methodology note
Source-bounded documentary review of insurer statements, policy-form wording and guidance; four evidence questions and leadership actions are ODA3 analytical proposals; illustrative synthetic document-classification scenario, not an empirical examination. Technical failure, financial loss, legal liability and insurance coverage considered separately.
Notably Absent
The following were not located in the sources checked on 15 October 2026:
- A public account of how an insurer uses third-party AI audit findings in underwriting, or whether evidence produced outside an insurance-linked scheme receives equivalent consideration.
- A documented instance, within the sources examined, showing precisely how independent examination changed an underwriting decision, terms, exclusions, limits or price.
- Public data on the frequency or severity of claims paid under AI performance policies.
- A published insurer method for assessing agents that act through tools, despite Munich Re's model-agnostic statement [1].
- A court ruling interpreting the cited ISO generative AI endorsements or Berkley exclusion within the legal and insurance sources examined; no comprehensive court-database search was performed.
ODA3 Institute examined publicly reproduced exclusion endorsement wording but did not examine any complete issued insurance policy, customer system, underwriting file or claims file. Whether these endorsements were attached to any particular policy was not established.
