Last updated: November 2026 · Reading time: about 12 minutes
What this guide covers
Why an AI-caused loss is a different trigger
Cyber policies are built around a specific insuring clause: unauthorized access to, or use of, a computer system. The event that triggers cover is a security failure — someone got in who should not have, or data left the system without authorisation.
An AI-caused loss usually does not fit that clause. Consider the most common scenarios a US business now faces:
- An AI customer service agent gives a customer incorrect advice, and the customer suffers a financial loss
- An AI agent with payment permissions initiates a transfer that was not authorised by a human
- A model summarising documents omits a critical clause, and the business relies on the summary
- An AI system with legitimate access to a database exposes records because of a prompt injection
In each case, the system did what it was authorised to do. It accessed data it had permission to access, and it executed a function it was configured to execute. There was no breach in the cyber sense. The failure was in performance, design, or oversight — which is the territory of professional liability (E&O) or tech E&O, not cyber.
That distinction is where most businesses get it wrong. They assume "AI" means "cyber" because both involve computers. In insurance terms, they are different triggers, different grants, and different exclusions .
The four scenarios and where each one lands
Mapping the scenarios against the coverage grants is the fastest way to see the gap.
| Scenario | What the loss actually is | Which policy responds |
|---|---|---|
| Prompt injection leaks customer records | Unauthorised access and data disclosure — a security failure | Cyber, under the privacy and network security grant. This is the cleanest AI-caused cyber claim . |
| AI agent initiates an unauthorised payment | Funds transfer fraud or social engineering — a financial transfer that should not have happened | Cyber, but almost always under the funds transfer fraud sublimit, often around $250,000 . |
| AI assistant gives wrong advice, client loses money | Professional services failure — negligent advice or a defective deliverable | Tech E&O or professional liability. A cyber policy will usually not respond, because there was no security failure . |
| Model hallucinates a clause in a contract summary | Errors and omissions in a professional work product | E&O, if the firm carries it. Neither cyber nor general liability is designed for this . |
Two of the four are cyber claims. Two are E&O claims. A business that carries only cyber cover has a 50% gap on this list — and the half it is missing is the half that grows as AI is used for more consequential work.
The silent AI problem
The market has not settled on a standard approach to AI. Carriers are currently doing one of three things, and the differences matter more than the headline wording.
| Approach | What it looks like in the policy | What it means for a claim |
|---|---|---|
| Explicit AI exclusion | An endorsement listing "loss arising out of or relating to artificial intelligence" as excluded, sometimes with carve-backs | The claim is denied unless it falls within a carve-back. The direction of travel in the market is toward explicit exclusion rather than silent coverage . |
| AI endorsement | A named AI coverage grant, usually with a sublimit and conditions on governance and human oversight | The claim can be paid, but only within the endorsement's terms and limit — often well below the policy aggregate . |
| Silent | No mention of AI either way | The claim is argued on the general insuring clause and exclusions. This is the riskiest position, because it depends on how a court or adjuster reads "unauthorized access" against an AI-caused event . |
Silence is not neutrality. A silent policy is one where the answer is decided after the loss, by someone other than you. For a business deploying AI in customer-facing or payment-facing roles, silence is the position to avoid.
The one AI loss that is clearly covered — and its sublimit
There is one AI-caused loss that maps cleanly onto existing cyber cover: deepfake-driven funds transfer fraud.
The mechanism is straightforward. An attacker uses an AI-generated voice clone or video deepfake to impersonate a CEO, a supplier, or a known contact, and convinces an employee to authorise a payment. The loss is a fraudulent transfer. That is the same insuring event as a traditional business email compromise — the delivery method is AI, but the loss category is funds transfer fraud.
The coverage exists. The limit is the problem. Funds transfer fraud and social engineering are almost always sublimited, typically around $250,000 on a small business policy . And the frequency is rising: BEC accounted for $3.05 billion in FBI-reported losses in 2025 across 24,768 complaints, and AI voice cloning is making the attacks more convincing and more successful .
So the deepfake scenario has a different answer from the other AI scenarios: it is covered, but likely capped well below the loss. A $400,000 deepfake-driven transfer on a policy with a $250,000 sublimit leaves a $150,000 gap — and the insured still has to prove the loss falls within the social engineering grant rather than the general fraud exclusion.
The agent problem: who owns the loss
The hardest AI-caused loss to place is one caused by an autonomous agent — a system that takes actions rather than producing text. An agent that initiates a payment, sends an email, files a document, or approves a transaction sits at the intersection of cyber, E&O, and a question the insurance market has not fully answered: who is the actor?
If a human approved the payment, the loss is a human error or a fraud. If an agent approved it, the loss is a system failure — but the system was authorised to act. This is the same accountability question raised in our guide to who owns the AI agent in your law firm, and it applies to any business deploying agents with transactional permissions.
Three practical consequences for coverage:
- Cyber cover depends on how the loss is characterised. If the agent was manipulated by a prompt injection or a compromised credential, the loss may fall within the cyber grant. If the agent simply executed a bad instruction within its permissions, it may not.
- E&O cover depends on whether the agent's output is a professional service. An agent that drafts advice or produces a deliverable sits closer to professional liability. An agent that moves money sits closer to cyber — or to crime and fidelity cover, if the business carries it.
- Human oversight is now an underwriting condition. Carriers are asking whether agents operate with human-in-the-loop approval, what permission boundaries apply, and whether agent actions are logged. A business that cannot answer those questions is a harder risk to place.
What underwriters now ask about AI
The underwriting conversation has moved quickly. Carriers are no longer asking only about security controls — they are asking about AI governance, and the answers affect both price and eligibility.
| Question | Why the carrier is asking |
|---|---|
| Where is AI used in your business? | To map the exposure to the right coverage grant — cyber, E&O, or both . |
| Which AI systems can take actions, not just produce text? | Agentic systems carry a different loss profile from chatbots. Carriers want to know which systems have transactional or communication permissions. |
| What human oversight applies before an AI action is executed? | Human-in-the-loop is a risk-reduction factor. It is also increasingly a coverage condition. |
| How are AI actions logged and reviewed? | Without a log, a claim cannot be proved. Carriers ask because the absence of logging is a claims-handling risk. |
| Do you have an AI use policy? | Governance documentation is becoming a prerequisite in the same way MFA and EDR became prerequisites for cyber . |
The pattern matches the earlier shift in cyber underwriting. MFA, EDR, tested backups and patch management moved from discount levers to mandatory prerequisites, and 73% of small businesses now fail their cyber insurance assessment as a result . AI governance is on the same trajectory. The businesses that document their AI use, permissions and oversight now will be the ones that can still buy cover when the market hardens.
What to ask your broker before renewal
Six questions, in this order. The first three determine whether the policy pays on an AI-caused loss at all.
- Is there an AI exclusion or an AI endorsement in this policy? Ask for the endorsement wording, not the summary. Confirm whether it is an exclusion, a grant, or silence .
- Which of my AI use cases fall inside the grant? Walk the broker through each system — customer service, document summarisation, payment agents, coding assistants — and ask for a written answer on each.
- What does the funds transfer fraud sublimit actually cover? Confirm whether a deepfake-driven transfer falls within the social engineering grant, and whether the sublimit applies to the transfer only or the full incident .
- Do I carry tech E&O or professional liability alongside cyber? If the answer is no, the AI advice and deliverable scenarios are uninsured. That is the gap most businesses discover only after a loss.
- What governance documentation does the carrier want? An AI use policy, an inventory of AI systems, permission boundaries for agents, and logging. Producing these before renewal improves both eligibility and price.
- What happens if an agent acts without human approval? Ask the broker to confirm in writing whether that loss is covered, and under which grant. If the answer is unclear, treat the agent as uninsured and configure human approval before it acts.
If any of the first three produces an unclear answer, do not renew on the assumption that "cyber covers AI." The policy may be silent, may exclude, or may cover only the narrow deepfake scenario — and the difference is only visible in the endorsement wording.
Frequently asked questions
Does my cyber policy cover an AI-caused loss?
It depends on the trigger and the wording. If the loss involved unauthorised access — such as a prompt injection that leaked data — it likely falls within the cyber grant. If the loss was a performance failure, such as bad advice or a hallucinated summary, it is closer to E&O and the cyber policy will usually not respond . If the policy contains an explicit AI exclusion, the claim may be denied even where it would otherwise fit the cyber grant .
What is a silent AI policy?
A policy that neither excludes nor expressly covers AI-related losses. The claim is then argued on the general insuring clause and exclusions, which means the answer is decided after the loss rather than at the point of purchase. For a business deploying AI in payment or customer-facing roles, silence is the position to avoid .
Is a deepfake-driven payment covered by cyber insurance?
Usually yes, because it maps onto the funds transfer fraud or social engineering grant — the same grant that responds to traditional business email compromise. But that grant is almost always sublimited, typically around $250,000, so a larger loss leaves a gap . BEC accounted for $3.05 billion in FBI-reported losses in 2025, and AI voice cloning is making the attacks more convincing .
Do I need tech E&O as well as cyber insurance?
If your business provides advice, produces deliverables, or deploys AI that customers rely on, yes. Cyber responds to security failures; E&O responds to professional services failures. An AI assistant that gives wrong advice and causes a client loss sits in E&O territory, and a cyber-only programme leaves that uninsured .
Who is liable when an AI agent moves money incorrectly?
That is the open question the insurance market has not fully resolved. If the agent was manipulated by an attacker, the loss may be characterised as a cyber event. If the agent simply executed a bad instruction within its permissions, it may fall to E&O — or to nothing, if the business carries neither. The practical answer is to configure human approval for any agent with transactional permissions, and to confirm in writing with the broker which grant responds .
Will an AI exclusion deny a claim even if the AI was only part of the workflow?
It can. An exclusion for "loss arising out of AI" may be argued to apply where AI was part of the process that produced the loss, even peripherally. That is why the endorsement wording matters more than the summary. Ask the broker to confirm in writing which of your AI use cases fall inside and outside the exclusion .
What documentation do carriers now want for AI exposure?
An inventory of AI systems, clarity on which systems can take actions rather than produce text, human oversight and approval rules, logging and review of AI actions, and a documented AI use policy. This mirrors the earlier shift on MFA and EDR, which moved from discount levers to mandatory prerequisites .
Related reading on Dedon Tech: cyber insurance for US small businesses: what 2026 policies actually cover · AI in insurance: what it actually does to your premium and claim · who owns the AI agent in your law firm · InsurTech vs traditional insurance · full site map
0 Comments: