Identify the loss, insured party and policy trigger before relying on an AI insurance label.
Research edition: 5 November 2026 · DN Agent Insurance Question Map v0.1
What Matters
AI agent insurance is not one interchangeable promise. Identify whose loss is involved, what event or performance failure must be established, who is insured and which terms apply. Official providers describe different approaches. DN’s question map helps prepare a broker discussion; actual coverage remains undetermined until the applicable wording and facts are assessed.
Consider three hypothetical incidents. An agent places an expensive but authorized order with the wrong supplier. A customer alleges that the agent’s research caused a financial loss. An attacker manipulates the agent into sending money.
All three involve AI. That label does not establish the same cause, claimant, contract or insurance response.
For businesses moving agents from experiments into operations, the useful question is therefore more specific than “Are we insured for AI?” It is: which loss, arising from which activity, is addressed by which wording?
- Last verified: 5 November 2026. Official public materials from Munich Re, Armilla AI, CFC and iTOO reviewed.
- Evidence class: provider descriptions and announcements, not an independent assessment of issued insurance terms.
- Original asset: ten illustrative loss scenarios, forty broker questions, a dated source-review file and a blank preparation brief.
- Not performed: no insurance quote obtained, policy purchased, claim filed or claims outcome tested. Individual purchase and geographic eligibility were not verified.
- Conclusion boundary: no insurer ranking, insurability score or coverage verdict. Use the complete wording and qualified advice for the relevant jurisdiction.
The Signal: Providers Are Taking Different Approaches to AI Risk
The following comparison describes what the providers publish. It does not establish that a particular reader, agent deployment or incident qualifies. The rows are alphabetical, not a ranking.
| Provider | What the official material describes | Question to take forward | What DN has not established |
|---|---|---|---|
| Armilla AI | A standalone affirmative AI liability approach, including descriptions of agent mistakes and output-related claims | How does the issued form address the particular AI-enabled service and third-party allegation? | Reader eligibility, actual limits, exclusions or claim response |
| CFC | A programme announced on 25 June 2026 to add explicit AI language across seven policy lines | Which applicable product, form and regional wording address this exposure? | That every form or territory has identical wording or availability |
| iTOO | An aiSure offering framed around pre-agreed performance targets or KPIs, with a broker route for enquiries | Which metric, threshold, loss and claim process would be agreed for the deployment? | A quoted premium, purchase eligibility or the outcome of a future claim |
| Munich Re | aiSure approaches for providers and deployers, discussing performance warranties, contractual liabilities, AI liabilities and financial losses | Is the proposed arrangement a vendor warranty, direct protection for the deployer, or another tailored structure? | That the different arrangements share the same trigger, beneficiary or conditions |
Armilla’s page identifies it as a Lloyd’s Coverholder and managing general agent and describes third-party loss and agent-action exposures. Those public descriptions do not establish that a deployer’s own mistaken payment is reimbursable.
Munich Re’s aiSure material distinguishes provider-facing performance warranties from arrangements for corporations using AI. That distinction gives a buyer a useful starting question: who holds the protection and who is entitled to receive compensation under the actual arrangement?
iTOO’s product page says a performance KPI is agreed before issuance and must be measurable and verifiable. It directs prospective buyers to a broker and describes the offering as working alongside existing policies. DN has not verified any reader’s eligibility or binding terms.
CFC’s announcement spans products including technology E&O, professional liability, media and cyber. It is an announced programme, not evidence that every reader’s existing policy has already changed.
DN Alpha Thesis: Start With the Loss, Then Read the Promise
A technical failure category is not enough to identify an insurance response. “The model hallucinated” might describe an inaccurate draft that was caught before use, a customer allegation, a breach of a performance promise, or a direct operational loss.
DN’s approach is a loss-to-policy question map. Describe the loss and affected parties first. Then ask which insuring clause, performance trigger or contractual promise is being considered, and which definitions, exclusions and conditions govern it.
This is a research framework for asking better questions. It is not a legal classification system or a recommendation to buy a particular product.
| Situation | First question | Assumption to avoid |
|---|---|---|
| Your business loses money | Does the relevant provision address this kind of own loss? | Third-party liability language automatically reimburses your balance sheet |
| A customer brings a claim | Who is insured and is the service or activity within scope? | Every contractual promise is covered because AI was involved |
| A promised KPI is missed | Which metric, threshold, population and period were agreed? | A performance guarantee is equivalent to a general liability policy |
| Malicious interference is suspected | How do the applicable cyber, crime and AI-related terms address the actual facts? | A valid credential or signed instruction settles the coverage analysis |
An Authorized Error Is Not Automatically Outside Cyber Wording
It would be equally misleading to assume that cyber insurance can only matter when an attacker is involved. CFC’s illustrative examples, published 30 September 2026, include an agent removing a critical business system and describe a possible system-failure response under its Cyber Proactive Response approach.
The same page warns that availability differs across products, territories and jurisdictions and that coverage is subject to the applicable terms. Its examples are illustrative. DN has not tested that response or assessed a reader’s policy.
The practical takeaway is to ask about the actual event definition. Do not infer either inclusion or exclusion from the product label. For payments, keep technical authorization and user intent separate when describing the incident to advisers; the legal and insurance treatment still requires its own review.
Use the DN Agent Insurance Question Map
Open the scenario closest to the issue you need to discuss. Each section identifies the loss to describe, questions for a broker or insurer and evidence to prepare. More than one scenario may apply. None returns a covered or uncovered verdict.
Use the complete policy, schedule, endorsements, specific facts and applicable law with a licensed broker or insurer and qualified advisers. The map does not create insurance, submit an enquiry or replace a coverage opinion. For an existing incident, promptly consult the applicable notification and response requirements rather than treating this article as a claims procedure.
The agent makes an authorized but wrong purchase
Loss to describe: Your business incurs its own financial loss, with no customer claim yet.
Ask about own financial loss, performance triggers and the actual definition of a covered event.
- Does the proposed wording cover our own loss, a third-party claim, or both?
- Which failure, KPI or event must be established before the policy can respond?
- How are trading, investment, procurement or other sector-specific activities treated?
- Which deductibles, limits, exclusions and proof requirements apply to this scenario?
Prepare: Task and authority record; Actual loss calculation; Model, tool and policy versions; Relevant acceptance or KPI evidence.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
A customer claims that the agent’s work caused a loss
Loss to describe: A third party alleges loss arising from a service your business delivered.
Ask about AI liability, professional liability or technology E&O wording for the described service.
- Who is the named insured, and is the relevant AI-enabled service within the insured activities?
- What must be alleged or established for the third-party claim to fall within the insuring clause?
- How are contractual promises, indemnities and liability assumed under an agreement treated?
- How are defense costs, settlements, deductibles and applicable limits handled?
Prepare: Customer agreement and service scope; Claim or complaint chronology; Delivered output and supporting records; Relevant versions and human interventions.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
An attacker manipulates the agent into sending money
Loss to describe: There is suspected malicious interference as well as a financial consequence.
Ask how cyber, crime, social-engineering and AI-related provisions interact; the tool does not classify the event legally.
- How does each relevant wording define unauthorized activity, fraud and social engineering?
- Does technical use of valid credentials or a signed instruction affect the analysis under that wording?
- Which exclusions, sublimits, consent requirements or overlapping provisions need clarification?
- What incident response, notification and evidence-preservation steps should we follow now?
Prepare: Preserved authorization and transaction records; Incident timeline; Access and tool records retained securely; Relevant incident-response communications.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
The agent reveals confidential information in its output
Loss to describe: Information may be disclosed through an output without a conventional network intrusion.
Ask about privacy, confidentiality, output-related liability and response-cost provisions.
- Does the wording address a disclosure through model output without a separate security breach?
- Which affected parties and categories of information are within the scope?
- Which response costs, third-party claims or investigations are addressed, if any?
- What confidentiality, notification, retention and jurisdictional requirements should our advisers assess?
Prepare: Redacted output reference; Data-access scope; Affected-party assessment; Incident chronology without publishing sensitive data.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
The agent misses a promised performance target
Loss to describe: A vendor has made a performance commitment to a customer.
Ask about the performance warranty, contractual obligation and the party entitled to claim.
- What exact metric, population, period and threshold were agreed before inception?
- Who is insured, who holds the warranty, and who can receive a payment?
- How are model changes, data drift, regrading and human assistance handled?
- Which exclusions, dispute process, loss calculation and notification terms govern the promise?
Prepare: Agreed metric and contract version; Dated baseline and subsequent results; Configuration and change history; Relevant warranty and policy documents.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
An AI-enabled workflow stops operating
Loss to describe: Your operations are interrupted, potentially involving an upstream provider.
Ask about interruption, system-failure, dependency and restoration wording for the actual cause.
- Which cause of interruption must fall within the wording?
- Are the affected systems and upstream services within the defined scope?
- Which waiting periods, loss measures, restoration costs and limits apply?
- What evidence distinguishes a covered event from an ordinary service or commercial dispute?
Prepare: Service dependency map; Outage and recovery chronology; Financial-loss calculation; Relevant provider communications.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
A third party challenges an AI-assisted output
Loss to describe: The complaint concerns rights in output, training material or a delivered service.
Ask about the particular intellectual-property, media, output or liability provisions.
- Which alleged rights and activities are addressed by the wording?
- How are training inputs, generated outputs and downstream publication distinguished?
- What defense, settlement-consent, exclusion and territory terms are relevant?
- What should we preserve while qualified legal and insurance advisers review the allegation?
Prepare: Allegation and publication chronology; Input and output provenance references; Relevant permissions or licenses; Contract and policy references.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
A regulator investigates an AI-enabled process
Loss to describe: There may be investigation costs, remediation obligations or a potential penalty.
Ask about investigation and defense costs and any legally insurable penalties; do not assume all fines can be insured.
- Which investigations, insured persons and AI-enabled activities are within the wording?
- Which defense or response costs are addressed, if any?
- Which penalties, if any, are legally insurable and included in the applicable jurisdiction?
- What notification, cooperation, consent and document-preservation obligations need immediate review?
Prepare: Regulatory communication; Process and governance records; Change and incident history; Policy and endorsement references.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
The same model error affects many customers
Loss to describe: A repeated failure can generate multiple losses, complaints or claims.
Ask about aggregation and shared limits rather than assuming a new limit for every event.
- How does the wording group related errors, incidents, models or claims?
- Which per-claim, per-model, aggregate or other limits apply to these facts?
- Do defense costs or earlier claims reduce any applicable limit?
- How are the relevant dates, prior events and notifications determined under the policy?
Prepare: Linked incident and claim register; Affected model and deployment versions; Loss and defense-cost records; Relevant policy periods and notices.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
The deployed agent changes after insurance is arranged
Loss to describe: The insured deployment may have a new model, tool, permission boundary or use case.
Ask about disclosure and change requirements; the map does not decide whether coverage continues or is lost.
- Which changes must be notified, agreed or documented under the actual terms?
- Does the described deployment still match the insured activities and scheduled systems?
- What new test, governance or performance evidence should be supplied?
- What written clarification or endorsement, if any, is needed before relying on the changed deployment?
Prepare: Before-and-after deployment inventory; Known and unknown changes; Relevant retest and control records; Broker and insurer correspondence.
Coverage remains undetermined. Ask for review of the applicable wording and facts.
Ten-scenario map JSONForty broker questions CSVBlank broker briefProvider-description reviewMethodology and limits
The Evidence Pack: Explain the Deployment You Actually Run
A polished demonstration does not tell an underwriter everything about an agent’s authority or operating context. A preparation brief should identify the business entity, use case, model and tool dependencies, access, money-moving permissions, oversight and how exceptions are handled.
Then connect that deployment to evidence. DN’s acceptance-testing lab helps distinguish a format check from substantive task criteria. The benchmark-integrity framework asks what attempts, exclusions and assistance sit behind a score. Neither establishes that a system is insurable.
If the deployment changes, retain the before-and-after record. The evidence-validity planner helps frame which claims may need review. Ask the broker or insurer what the actual policy requires for changes to the model, workload, tools or permissions. Do not assume that every change automatically preserves coverage or voids it.
Use redacted references and secure channels for sensitive records. Do not put customer data, credentials, private policy documents or incident details into a public comment or shared research file.
Read Limits, Defense Costs and Aggregation Together
A large headline limit cannot answer every practical question. Ask which limit applies to the relevant entity, model, claim or event, whether related losses are grouped, and whether defense costs reduce an available limit. Request clarification of the actual deductible, retention, waiting period or other condition where relevant.
For a repeated agent error, the treatment of related claims may matter as much as the amount of any advertised limit. For a performance promise, the covered metric and loss calculation may be decisive. These are questions for the actual contract, not numbers this article can infer from a website.
Similarly, do not infer that a provider’s reference to regulatory exposure means all fines can be insured everywhere. Ask qualified advisers about the jurisdiction, the particular penalty, insurability and the issued terms. Investigation and defense expenses require separate attention.
A Better Procurement Sequence
- Inventory the deployment: identify who owns it, what it can do and the business activities involved.
- Describe concrete loss scenarios: separate your own loss, third-party allegations, performance commitments and suspected malicious interference.
- Collect existing documents: policy schedules, endorsements, contracts, warranties and relevant deployment evidence.
- Prepare focused questions: use the map to identify unclear triggers, insured parties, activities, limits and conditions.
- Request written clarification: ask the licensed broker or insurer to assess the actual facts and documents, including local availability.
- Retain the decision record: record what was clarified, what remains uncertain and which changes require a new discussion.
For a small team, start with one material workflow and the legal entity running it. For an AI vendor, distinguish the promises made to customers from the insurance arrangement behind those promises. For an enterprise, coordinate technical owners, legal advisers, risk managers and the broker around the same deployment record.
Insurance is also not evidence that an agent has the correct permissions or that its controls work. Continue appropriate prevention, evaluation, monitoring and incident preparation. A sales claim about protection should not substitute for a controlled deployment.
The DN Research Desk offers paid, scoped research on deployment claims, dependencies and available evidence. Bring the policy and coverage questions to the appropriate licensed broker, insurer and legal advisers. This placement is a research-service invitation, not insurance broking, a quote or a coverage decision.
Methodology, Corrections and Commercial Disclosure
Scope: four providers’ official public descriptions were reviewed on 5 November 2026. The comparison records their published approaches and the unanswered purchasing questions. It is not an exhaustive market map, independent policy audit or recommendation of the best insurer.
Research not performed: no issued policy was assessed, quote obtained, claim filed or claims outcome tested. DN did not verify individual purchase eligibility or determine applicable law. Public descriptions can change and may differ by region. CFC’s regional site can redirect visitors; retain the dated source and confirm the current local wording.
Original files: ten hypothetical scenarios contain forty DN-authored questions and preparation prompts. The JSON and CSV expose the same questions as the article. Every scenario retains coverageVerdict: undetermined. The brief starts blank; null fields mean information has not been supplied. These are editorial research assets, not underwriting decisions or executed agent benchmarks.
Falsification: identify a provider statement attributed inaccurately, a mismatch between the visible and downloaded map, or a material omission in the preparation questions. A real claim outcome cannot be evaluated from the map alone. Send corrections with the source and specific issue.
Commercial context and reuse: provider links lead to official reference materials; inclusion is not a rating. Research Desk is DN’s paid service. Commercial considerations do not determine the questions or findings. DN-authored research files are CC BY 4.0 with attribution and limitations; provider documents retain their terms. Initial release: v0.1, 5 November 2026. Material changes require a new version.
Primary Sources
- Munich Re: aiSure, reviewed 5 November 2026.
- Armilla AI: affirmative AI liability public description, reviewed 5 November 2026.
- CFC: affirmative AI coverage programme, 25 June 2026.
- CFC: illustrative policy-response examples, 30 September 2026.
- iTOO: aiSure product page and FAQs, reviewed 5 November 2026.
- DN provider-description review and original question map.
Frequently Asked Questions
What is AI agent insurance?
The phrase can refer to different insurance approaches involving AI-enabled systems, including liability, performance-related promises and financial-loss cover. The applicable policy, schedule, endorsements, insured activities and facts determine what is actually addressed.
Does an AI insurance label mean every agent mistake is covered?
No. A public product description is not a coverage decision. Ask a licensed broker or insurer to assess the actual wording, insured entity, deployment, loss, applicable jurisdiction and conditions.
Does cyber insurance only matter when an attacker is involved?
Do not assume that rule. Some wordings may address specified system failures as well as malicious events. Ask how the actual event and cause are treated; the relevant CFC examples are illustrative and conditional.
What is the difference between my loss and a customer claim?
They raise different coverage questions. Identify who suffered the loss, who is insured, who is alleging liability and which insuring clause or agreed performance trigger is being considered. Do not treat those roles as interchangeable.
Is a vendor performance warranty the same as my own insurance policy?
Not necessarily. Ask who holds the policy, who owes the warranty, who can claim or receive payment, what metric triggers the arrangement and which terms, exclusions and limits apply.
Does DN recommend one insurer as the best?
No. This is a dated review of four official public descriptions and an original question map. DN did not obtain quotes, assess issued policies, verify individual purchase eligibility or test claims outcomes.
Will AI trading losses or a mistaken payment be reimbursed?
This article cannot determine that. Ask about the actual activities, cause of loss, authorization facts, applicable provisions and exclusions. Do not infer coverage for financial trading, fraud or money movement from generic AI wording.
Are regulatory fines always insurable?
No blanket conclusion is appropriate. Ask qualified advisers which penalties, if any, are legally insurable and included under the relevant wording and jurisdiction. Investigation and defense costs require their own review.
What should I prepare for a broker discussion?
Prepare a deployment inventory, permissions, contracts, loss scenarios, relevant test and KPI evidence, incident records and existing policy references. Use secure, redacted references rather than publishing sensitive information in a public brief.
What does the DN question map prove?
It supplies ten illustrative scenarios, forty original questions and evidence-preparation prompts. Coverage remains undetermined throughout. It does not create insurance, certify insurability, submit an enquiry or replace professional advice.
Evidence cut-off: 5 November 2026. DN Agent Insurance Question Map v0.1. Public descriptions and preparation questions do not establish actual insurance coverage.
About the Author
