When an AI agent goes wrong, the first question is who pays, and the natural assumption is that one of your standard policies will respond.
An AI agent failure does not sit inside any single policy. Coverage is a patchwork across cyber, technology errors and omissions (tech E&O), directors and officers (D&O) and employment practices liability insurance (EPLI). Gap risk is the distance between what a loss is and what each of your policies was designed to pay. AI agent failures create exactly that kind of mismatch.
For the wider picture, see the full who-pays overview and the exclusions that created this gap.
Cyber vs Tech E&O vs D&O: which coverage line responds to an AI agent failure?
Claim characterisation decides which line pays. No traditional policy is written for AI agent errors, so cyber, tech E&O, D&O and EPLI each respond to a slice of it. Cyber responds when the agent is the vector of a security incident, like a prompt injection that exfiltrates your data. A biased recommendation is different: no breach occurred, so it does not trigger cyber. Tech E&O is for when the system you deployed failed your client, but its trigger is professional error, and autonomous behaviour rarely matches that. D&O answers governance and securities claims, like an “AI washing” allegation, and a discriminating hiring agent drags in EPLI.
What is “gap risk” in insurance coverage for AI-related losses?
Gap risk is the possibility that no policy in your tower responds to an AI loss at all. An agent failure can implicate cyber, tech E&O, D&O and EPLI, each with its own trigger, exclusion and limit, so the loss can fall between them. Insurers are narrowing each line independently, and the combined effect is fragmentation.
AI losses fall between the layers of your tower, your stack of insurance policies, because no layer is written to respond to an agent acting on its own. Policies written before generative AI are “silent AI”, so coverage is argued out after the loss. One in five commercial insurers reported an AI-related loss in 2025, and only about half were fully covered.
How do I assess whether my existing cyber and E&O policies still respond to AI agent actions?
Read the exclusions, definitions and sublimits (lower caps on AI losses inside a broader limit) added at renewal, then map each plausible failure to a trigger. Coverage that silently existed yesterday is being removed at renewal, because most policies predate generative AI.
ISO is the organisation that writes the standard insurance forms most US policies are built on. ISO’s generative-AI exclusions (CG 40 47, CG 40 48 and CG 35 08), three forms effective January 2026, remove bodily injury, property damage and personal-and-advertising injury from general liability. W. R. Berkley introduced an “absolute” AI exclusion across D&O, E&O and fiduciary liability (the cover for the people who run a company’s employee benefit plans), and one law firm called its breadth impossible to overstate. A sublimit caps AI losses inside a broader limit, while a carve-back restores named claims an exclusion removed. Wrong integration steps are a tech E&O problem, a compromised agent session falls to cyber, and an investor who alleges the model was overstated brings a D&O claim.
How should I approach a coordinated policy review across cyber, Tech E&O, D&O, and EPLI?
A review reads every policy for AI definitions, exclusions, carve-backs and sublimits, then maps each trigger to the agent’s failure modes. Coverage is split across lines, so the gaps sit between the policies, and a review of one policy misses them. Law firms now advise this well before renewal.
Responding to the exclusion wave, the review asks whether an exclusion in one line pushes a claim onto another, and whether the excess policies that sit above the primary policy and pay after its limit is exhausted follow the primary policy’s wording. Claims-made policies respond only to claims reported in the period, so audit-log retention and prompt notice matter. An agent treated as an employee whose actions cause workplace harm brings EPLI in, and the review flags that area as unsettled.
Standalone AI liability insurance vs vendor indemnity: which actually protects the deployer?
Neither standalone AI liability insurance nor vendor indemnity protects the deployer on its own, and the combination usually does. Insurance is a third-party promise backed by capital, while indemnity is the vendor’s own promise, differing on credit risk, caps and triggers.
Vendor indemnity is thin in practice. OpenAI’s terms cap liability at twelve months of fees, and an enterprise paying $100,000 a year faces a $100,000 ceiling. An indemnity capped that way is worth little against a claim the size of Bartz v. Anthropic or Concord Music Group v. Anthropic, two copyright suits against Anthropic. Standalone AI liability insurance triggers on AI behaviour such as underperformance, hallucinations, inaccuracies and bias, with limits far below the exposure the biggest AI labs like OpenAI and Anthropic face. And the contract clauses decide a lot before insurance is reached.
Where can I find standalone AI liability insurance products?
Armilla, Testudo, HSB/Munich Re, Corgi and AIUC make up a small specialty market, and no product is a general-purpose “AI policy”. Armilla, backed by Lloyd’s syndicates Chaucer and Axis, sells coverage for AI underperformance, hallucinations and inaccuracies up to $25 million per insured. Munich Re’s aiSure has used threshold-based performance triggers since 2018, and HSB writes AI liability for small businesses.
Testudo launched at $10 million with Apollo, Atrium and QBE capacity, meaning those insurers underwrite the risk. Corgi runs an integrated carrier platform, underwriting through its own carrier and partner carriers, and AIUC writes up to $50 million. Deloitte projects AI-related premiums of about $4.7 billion a year by 2032. These products sit alongside the contract clauses that decide who pays.
Captive and difference-in-conditions structures vs commercial market AI coverage?
When the commercial market will not write AI risk, alternative risk transfer takes over. A captive, a subsidiary you set up to insure your own risk, carries the frequency layer of small, recurring losses. A difference-in-conditions (DIC) policy carries the severity, the rare but expensive losses, and fills the gaps the layers below leave.
A captive is a balance sheet partition: you capitalise a subsidiary, retain the first losses, and pass part to reinsurers. The self-insured retention is that first-loss layer, and AI losses often sit below the attachment point, the amount of loss at which insurance starts to pay. A DIC policy is written to respond when the policy underneath it is narrower than the DIC’s own terms, because the underlying policy has excluded a type of claim. Reinsurance backs this through quota-share and excess-of-loss structures, where a reinsurer takes a set share of each policy or pays only the losses above an agreed amount. The constraint is accumulation risk, where one model failure triggers thousands of correlated losses across unrelated policyholders.
Claim characterisation, how the loss is framed as a breach, a professional error or a governance failure, decides which insurance pays. No single policy covers the whole loss.
Standalone AI liability and vendor indemnity protect the deployer differently. When the commercial market refuses the risk, captives carrying the frequency of small losses and DIC or reinsurance carrying the severity of large ones take over where no single policy will respond. The answer is a patchwork, and the full who-pays picture explains each piece.
Frequently Asked Questions
Is there a single “AI insurance” policy that covers everything an AI agent can do?
No. There is no general-purpose “AI policy” that responds to every AI agent failure. The honest answer today is a patchwork across cyber, tech E&O, D&O and EPLI, each with its own trigger, exclusion and limit. The standalone market only offers narrow, boutique products written for a specific slice of the exposure, such as hallucination or algorithmic bias, not one policy that owns the whole loss.
Does using AI agents void my existing cyber or tech E&O policy?
Not usually. Using AI agents does not typically void a policy outright. What happens is more subtle: insurers remove the ambiguous “silent AI” coverage at renewal by adding exclusions and sublimits. Your policy stays in force but responds to less than you assume. The real risk is a renewal that quietly narrows cover, not a policy being cancelled the day you deploy an agent.
What should I do in the first hours after an AI agent causes a loss?
Preserve evidence and notify promptly. Capture the audit logs, prompts and outputs before they are overwritten, because claims-made-and-reported policies attach to when you first knew of the incident. Then give notice to every carrier that could conceivably respond, cyber, tech E&O and D&O. Because claim characterisation is contested, early notice to all lines protects your position rather than betting early on one policy.
Does cyber insurance cover a deepfake or prompt-injection funds transfer?
Only sometimes. A prompt injection or deepfake funds-transfer fraud is a security-vector event, so cyber is the natural line, but standard cyber policies often require a social-engineering or funds-transfer-fraud endorsement to respond, and some fall to a separate crime policy instead. This is exactly a claim-characterisation problem: the same loss can be framed as a breach, a fraud or a professional error, and the framing decides who pays.
What happens if an AI agent causes physical injury or property damage?
You will usually fall into a gap, because cyber and tech E&O policies generally exclude or do not cover bodily injury and property damage from an AI agent. Physical harm pushes the claim toward commercial general liability or product liability, where the new ISO generative-AI exclusions may now cut it out. The distinction between physical and financial loss is underdeveloped in most commentary, so treat it as a flagged gap rather than a resolved answer.
My policies were written before generative AI existed. Are AI losses still covered?
They are silent, which means neither covered nor excluded, so the answer is resolved by argument after a loss rather than in advance. A pre-2023 cyber or E&O policy has no AI wording, so an insurer might pay, deny or compromise depending on how the claim is framed. This “silent AI” ambiguity is exactly what insurers are ending at renewal, so do not assume your older policy quietly covers what newer ones now exclude.
Do I have to tell my insurer I’m using AI agents at renewal?
Yes, and failing to do so can be worse than a higher premium. If AI use is material to the risk, non-disclosure can let an insurer reduce or deny a claim later. Insurers are now asking about AI deployment and adding exclusions or affirmative cover in response, so disclose what you actually run. The honest position at renewal is the difference between buying the cover you think you have and discovering the gap only after a loss.
Who decides how my claim gets characterised, me or the insurer?
Both, which is why framing is contested. You present the loss through the line you believe responds, and the insurer argues for the trigger, exclusion and limit in the wording it wrote. The same facts can be framed as a data breach, a professional error or a governance failure. Because each mapping lands on a different policy with a different limit, the characterisation fight is where the money is won or lost, not a formality after the event.
If OpenAI or Anthropic indemnifies me, do I still need my own insurance?
Usually yes. Vendor indemnity is the vendor’s own promise, not third-party capital, and AI vendors cap liability at around 12 months of fees while waiving consequential damages. That pushes the largest part of the loss back onto you. An indemnity also depends on the vendor’s credit and willingness to pay after the fact. Your own insurance is a promise backed by capital and an affirmative trigger, so the two work as a combination, not a substitute.
Can a small business buy AI liability cover, or is it only for large tech companies?
Yes, small businesses can. HSB, backed by Munich Re, writes AI liability insurance specifically for small businesses as an add-on for AI users, not just developers. The broader standalone market, Armilla, Testudo, Corgi and the AIUC, skews toward builders and larger insureds, but the exposure for a small business using an agent is real. Start with the user-side products and read the limits, which are low across the whole market.
Is standalone AI insurance worth buying when the limits are so low?
Only as a frequency layer, not a severity backstop. Limits top out around $25 million to $50 million per insured, far below frontier-lab litigation exposure, so standalone AI cover responds to underperformance, hallucination and bias claims rather than a catastrophic loss. Treat it as filling the specific gap your standard lines leave, and carry severity through higher layers, captives or DIC instead of through this product alone.
What’s the difference between an AI performance warranty and an insurance policy?
A performance warranty triggers on a defined metric, such as accuracy or error-rate thresholds, and pays if the model fails to meet it. An insurance policy triggers on a liability claim against you, after someone alleges harm. They are different instruments that are often bundled or confused, and each has different limits and conditions. The warranty guarantees the model’s performance; insurance responds to the legal claim that follows the failure.