AI insurance risks are reshaping coverage faster than most programs can keep up. This guide shows retail agents how AI creates new exposures across cyber, tech E&O, and professional liability, whether traditional cyber policies respond, which generative-AI exclusions are hitting 2026 renewals, and the new affirmative AI products now filling the coverage gap.
CVI · Coverage Advisory for Retail Agents
As businesses adopt AI, the exposures are landing in the gaps between cyber, technology E&O, and professional liability — and insurers are quietly rewriting the forms. Here’s where coverage may apply, where new exclusions are emerging, and what the new AI-specific products actually do.
Your clients are using AI whether or not it’s on any application you’ve filed. It’s in their customer-service chatbots, their marketing copy, their hiring workflows, their contract-review tools, and increasingly baked into software they didn’t choose to make “AI-powered.” Surveys now put small-business AI adoption around three-quarters of the market. The exposure is already there. The question is whether their insurance program knows about it.
For retail agents, this is the moment to get ahead of a shift that looks a lot like the “silent cyber” era of the late 2010s — when data-breach losses floated ambiguously under CGL and property forms until insurers slammed the door with exclusions and stood up a whole new market to replace what they took away. AI is on the same arc, only faster. This guide walks through how AI creates new risk, whether traditional cyber responds, which exclusions are arriving on 2026 renewals, and the affirmative AI products now available to fill the gap.
Full disclosure, appropriate to the subject: this article was researched and drafted with the help of the very technology it’s about. Which is either reassuring or precisely the point.
Key takeaways
- AI risk splits in two. AI as a weapon (deepfake fraud, AI-enhanced phishing) largely still falls to cyber. AI as the source of a loss (a client’s own chatbot or model output) generally does not — and that’s the gap catching people off guard.
- The exclusions are here. New ISO endorsements (CG 40 47 and CG 40 48) took effect January 2026, letting carriers exclude generative-AI claims from standard CGL. At least one carrier has introduced an “Absolute AI” exclusion on specialty lines.
- “Silent AI” is being removed at renewal. Coverage that used to respond by accident is being expressly allocated or quietly deleted from cyber and tech E&O forms — on faster timelines than silent cyber moved.
- A real affirmative-AI market now exists. Standalone and endorsement products from Armilla (Lloyd’s), Munich Re, Testudo, Counterpart, Hiscox, and Coalition can cover hallucinations, model failure, and AI output liability that legacy policies won’t.
How AI creates new risk
The single most useful thing an agent can do is stop treating “AI risk” as one thing. It’s two, and they land in completely different parts of the insurance tower.
1. AI as the weapon — attacks made faster, cheaper, and more convincing
This is the exposure most people picture, and it’s real. Generative tools have industrialized social engineering. Deepfake audio and video now defeat the “I recognized the CEO’s voice” control that used to backstop wire approvals — one multinational’s finance staff were deceived into transferring roughly $25 million after a video call in which every participant except the victim was an AI-generated fake. AI also writes cleaner phishing emails at scale, clones writing styles for business email compromise, and helps less-skilled actors probe for vulnerabilities. The threat is the same category your clients already faced — fraud, intrusion, extortion — just supercharged.
2. AI as the source of the loss — your client’s own system causes the harm
This is the newer, thornier exposure, and the one most likely to fall into a coverage gap. Here the loss doesn’t come from an attacker — it comes from the client’s own AI doing something wrong:
- Hallucination and bad output. An airline was held to a refund policy its customer-service chatbot simply invented; a tribunal rejected the argument that the bot was a separate entity the company wasn’t responsible for. When a client’s AI states something false and a customer relies on it, that’s a liability the client owns.
- Reputational and defamation harm. A solar-installation company sued Google after its AI Overviews feature allegedly stated the firm was facing legal action that didn’t exist, costing it business.
- IP and copyright infringement. AI-generated marketing content or code can reproduce protected material; the wave of training-data and output-infringement litigation is squarely a business exposure now.
- Bias and discrimination. AI hiring, underwriting, and pricing tools can produce discriminatory outcomes that surface as EPLI or regulatory claims.
- Model failure, drift, and agentic error. As “agentic” AI takes real actions, an autonomous system that misfires — books the wrong thing, sends the wrong data, executes a bad instruction — creates a loss no one can cleanly trace to a human decision.
- Data leakage. An employee pasting client PII or trade secrets into a public model is a disclosure event that may or may not trip a privacy trigger.
Why the split matters: the reason underwriters treat category 2 so cautiously is that nobody — including the insurer — can reliably reconstruct how an AI reached its output. That’s exactly the information a carrier needs to assign fault and price a policy. Some insurers are declining to write AI-output coverage at all; others require proof of red-teaming and documented governance first.
Is it covered under traditional cyber?
The honest answer is: partly, and less than your clients assume. A well-structured cyber policy is built around unauthorized access, privacy breach, network security, and extortion. So the weapon side of AI risk — a deepfake-enabled intrusion, AI-written ransomware, an AI-crafted phishing attack that leads to a breach — generally still looks like a covered cyber event, because the underlying peril hasn’t changed.
The source side is where cyber falls down. A hallucinating chatbot, an infringing AI-generated ad, a discriminatory algorithm, or a model that simply underperforms is not a network-security or privacy failure — it’s closer to a professional services error. That’s technology E&O or professional liability territory, not cyber. And even there, the fit is imperfect: standard tech E&O often excludes custom models, and the contractual liability created by an AI vendor’s indemnification clause typically sits behind a contractual-liability exclusion. Cyber forms, meanwhile, vary enormously and frequently carry electronic-data, bodily-injury, and intentional-act limitations that an AI claim can trip.
The practical result isn’t clean coverage or a clean denial — it’s fragmentation. An AI claim can fall between cyber, tech E&O, D&O, and EPLI, subject to competing exclusions across the tower, with carriers arguing about which policy (if any) owns it. That ambiguity is precisely what the market is now moving to eliminate — usually by taking coverage away.
The exclusions arriving now
This is the part agents need on their radar for every 2026 renewal, because it’s happening quietly and by endorsement:
- ISO generative-AI CGL exclusions (CG 40 47 and CG 40 48). Effective January 2026, these let carriers exclude generative-AI claims from standard commercial general liability — reaching bodily injury, property damage, and personal and advertising injury “arising out of” generative AI. Since ISO forms underpin the large majority of U.S. property-casualty policies, expect to see these attach at renewal, often without fanfare.
- “Absolute AI” exclusions on specialty lines. At least one carrier has introduced broad AI exclusion language that may reach any claim involving the use, deployment, development, integration, or failure of AI — not just a rogue chatbot. Read the trigger words carefully; “arising out of AI” is doing a lot of work.
- Silent AI being deleted from cyber and tech E&O. Forms that previously responded to AI losses “silently” are being revised ahead of renewals to either affirmatively grant or expressly exclude AI risk. The Lloyd’s and London markets are moving on this, mirroring the silent-cyber cleanup after NotPetya — only faster.
The renewal trap, already documented: a business’s AI chatbot promised a customer something the company never approved; the customer filed a claim; the carrier denied it because an AI exclusion had been added at the last renewal and nobody noticed. That’s not hypothetical — it’s the pattern agents are being warned to catch before binding.
Have a client leaning hard on AI? Let’s pressure-test their tower for the gap before their carrier does.
Call Steve: (818) 974-8117Which policy responds? The coverage grid
Here’s the honest lay of the land across the three lines most likely to be in play. “Maybe” means it depends on specific form language, endorsements, and whether silent AI has already been removed — which is exactly where a specialist reading the form earns the placement.
| The AI-driven event | Cyber | Tech E&O | Affirmative AI |
|---|---|---|---|
| Deepfake / AI-enhanced attack that breaches the network | Likely | No | Maybe |
| AI-generated ransomware or phishing leading to a breach | Likely | No | No* |
| Client’s chatbot gives wrong info the client is held to | Unlikely | Maybe | Yes |
| AI output harms a third party financially (hallucination) | No | Maybe | Yes |
| AI-generated content infringes copyright / IP | No | Maybe | Yes |
| AI hiring/pricing tool produces a discriminatory outcome | No | No | Maybe |
| Employee leaks PII into a public AI model | Maybe | No | Maybe |
| Model drift / underperformance breaches a client contract | No | No | Yes |
*Attack-driven losses are the cyber policy’s job; standalone AI products focus on the insured’s own AI performance, not third-party attacks. Every “Maybe” turns on form language and whether AI exclusions have attached.
The new AI insurance products
The good news for agents with AI-exposed accounts: the affirmative market has grown from nothing to a genuine set of options in barely a year. It’s still small — but it’s real, and it’s where the coverage the standard market is deleting has resurfaced.
- Armilla AI (Lloyd’s coverholder). The first standalone AI liability policy written at Lloyd’s, backed by syndicates including Chaucer, with limits expanded to $25M or more per organization. It affirmatively covers AI underperformance — hallucinations, model drift, errors, output liability, agent mistakes, and data leakage — and pairs coverage with independent AI system certification. Its “Vanguard AI” structure coordinates cyber and technology E&O with the standalone AI grant so the pieces don’t fight each other.
- Munich Re (aiSure). A performance-warranty pioneer (writing AI guarantees since 2018), now offered at meaningful limits, with an aiSure-General Liability variant aimed at bias and discrimination exposure — most relevant for AI developers.
- Testudo. A Lloyd’s Lab graduate focused on enterprises deploying AI, with limits up to roughly $9M.
- Counterpart & Hiscox. Counterpart added affirmative AI coverage to miscellaneous professional liability with a tech E&O agreement; Hiscox offers affirmative AI cover within its technology professional-indemnity book.
- Coalition. Added deepfake-related coverage to its cyber policies in late 2025 and offers AI coverage within its cyber program — most useful for the cyber-adjacent AI exposures above.
- Google Cloud Risk Protection Program. An affirmative AI endorsement offered with Beazley, Chubb, and Munich Re for Google Cloud customers.
Two structural notes worth passing to clients. First, these products increasingly come with an underwriting precondition: documented governance, risk assessments, and red-teaming evidence before a carrier will extend cover. Second, enterprise buyers are already being asked to show AI insurance — proof of AI coverage is starting to appear in master service agreements and procurement reviews as a condition of the deal. Coverage is becoming a sales requirement, not just a risk-management one.
What this means for retail agents
You don’t need to become an AI underwriter. You need a short discipline for every account that touches AI:
- Ask the question at renewal. Is an AI exclusion attaching to this CGL, cyber, or tech E&O form? Get it in writing. The denial-after-quiet-endorsement pattern is the one to prevent.
- Identify which bucket the client is in. Are they mostly exposed to AI-as-weapon (keep cyber strong) or AI-as-source (they need something cyber won’t give them)? Most meaningful adopters are now both.
- Document governance early. Clients who can show an AI use-policy, human-in-the-loop controls, and vendor due diligence are the ones the affirmative market will write — and price better.
- Don’t rely on silent coverage. “It might respond” is not a coverage position you want to explain to a client after a denial. Where the exposure is real, place affirmative cover.
- Partner with a specialist for placement. The affirmative AI market lives in the surplus lines and Lloyd’s/London space — exactly where a standard-market agent’s carriers can’t reach. That’s the referral.
CVI places the hard-to-fit risks your standard markets are now excluding. Bring us your AI-exposed accounts.
Email Steve: steve@cvins.comFrequently asked questions
Does a standard cyber policy cover AI risk?
Partly. Cyber generally still responds to AI-enabled attacks — deepfake fraud, AI-written phishing, or ransomware that leads to a breach — because the underlying peril is unchanged. It generally does not respond to losses caused by the insured’s own AI output, such as a hallucinating chatbot or an infringing AI-generated ad, which look more like professional-services errors than network-security failures.
What are the new ISO AI exclusions?
Two ISO endorsements, CG 40 47 and CG 40 48, took effect in January 2026 and allow carriers to exclude generative-AI claims from standard commercial general liability policies — reaching bodily injury, property damage, and personal and advertising injury arising out of generative AI. Because ISO forms underpin most U.S. property-casualty policies, expect them to attach at renewal.
What is “silent AI”?
By analogy to “silent cyber,” it refers to AI-caused losses falling under policies (CGL, D&O, E&O, EPLI, crime, media) that never contemplated AI when they were written. Insurers are now resolving that ambiguity — either affirmatively granting AI coverage or, more commonly, excluding it — which means coverage that used to respond by accident is disappearing at renewal.
If my client’s AI chatbot gives a customer bad information, is that covered?
Usually not under cyber. A chatbot that states something false and is relied upon creates a liability the business owns — as one airline learned when it was held to a refund policy its bot invented. That exposure fits a technology E&O or affirmative AI policy, not a standard cyber form, and even tech E&O may exclude it depending on wording.
Is AI-generated copyright infringement insurable?
It can be, but rarely under cyber. IP and copyright exposure from AI-generated content is generally addressed through media liability or affirmative AI products that expressly grant output liability. Standard cyber and many tech E&O forms exclude or limit IP claims.
What is affirmative AI insurance?
Coverage that names AI risk explicitly and grants it, rather than leaving it to be argued under a legacy policy. Products from Armilla (Lloyd’s), Munich Re, Testudo, Counterpart, Hiscox, and others affirmatively cover AI-specific failure modes such as hallucinations, model error, output liability, and underperformance — giving policyholders certainty instead of a coverage fight.
How much AI liability coverage is available?
Standalone limits currently run up to about $25M per organization from the leading Lloyd’s-backed writer, with other markets offering smaller towers. Some carriers cap or exclude systemic scenarios — losses from a single flawed model hitting many insureds at once — to guard against accumulation risk.
Do carriers require anything before writing AI coverage?
Increasingly, yes. Expect underwriting preconditions: an AI use-policy, human-in-the-loop controls, vendor due diligence, documented risk assessments, and in some cases red-teaming evidence. Some writers pair coverage with independent certification of the AI system itself.
Where does AI risk hit besides cyber and E&O?
D&O and EPLI both have exposure. D&O risk arises from management decisions — did leadership approve the use case, disclose material AI risks, and oversee vendors? EPLI risk arises when AI hiring or evaluation tools produce discriminatory outcomes. Agents should look at the whole tower, not just the technology lines.
Why place AI coverage through a surplus lines broker?
The affirmative AI market lives largely in the surplus lines and Lloyd’s/London space, which standard-market agents typically can’t access directly. A specialist wholesale broker like CVI can reach the carriers and syndicates writing these forms and structure them so cyber, tech E&O, and AI coverage work together rather than leaving gaps.
Saying yes when others say no — including on the emerging risks nobody’s forms are ready for. Let’s talk through your AI-exposed accounts.
(818) 974-8117 steve@cvins.comSteve McClure is Principal Broker at CVI, specializing in surplus lines and hard-to-place commercial risk. CA License #0G58010. This article is general information for insurance professionals, not a coverage opinion or legal advice; policy terms, exclusions, and conditions govern any actual claim, and AI coverage forms are evolving rapidly. Product references reflect public 2025–2026 market reporting (ISO, Lloyd’s, Munich Re, Armilla, Coalition, and industry press) and are not endorsements. Verify current form language and availability before binding.
Leave a Reply