What is AI liability insurance, and why is it suddenly a board-level question?

AI liability insurance is coverage built specifically for losses caused by an AI system's own decisions: a hallucinated answer that costs a customer money, a biased screening model that triggers a discrimination claim, an autonomous agent that approves a bad transaction. It exists because standard commercial policies were written for humans making mistakes, not software making them at scale. Search interest in the term is up roughly 240% year over year as of mid-2026, according to DataForSEO trend data, which tells you this stopped being a legal-team curiosity and became something CFOs are Googling on a Tuesday.

Two launches in the past 18 months explain the timing. In April 2025, Armilla Insurance Services rolled out an AI liability policy underwritten by Lloyd's of London, aimed at companies deploying third-party AI models where something goes wrong with the model's output rather than a data breach (Hunton Andrews Kurth's analysis). Then in March 2026, Munich Re's HSB unit went further downmarket, launching AI Liability Insurance built specifically for small and mid-sized businesses. That second one matters most for the reader of this post. It's a signal that carriers now consider a $10M or $30M revenue company deploying AI agents a real enough risk pool to underwrite, not just a niche for Fortune 500 legal departments.

Deloitte's research puts a number on where this is heading: roughly $4.7 billion in annual global AI insurance premiums by 2032, growing at a steady compounded rate from a near-zero base. Every article covering this space cites that figure because it's the only market-size estimate anyone has bothered to produce. Treat it as a direction, not a forecast you should plan a budget around.

Why your existing General Liability and Tech E&O policies probably don't cover this

They probably don't, and most operators find out the hard way. Insurers call this "silent AI" risk: AI exposure sitting inside a policy that was never written with AI in mind, because the underwriter who priced your General Liability or Tech Errors & Omissions policy three years ago wasn't thinking about a chatbot giving a customer bad financial advice or an agent auto-approving a claim it shouldn't have. Gallagher, one of the larger commercial brokers, has been warning clients about exactly this gap: AI risk that's technically uninsured because nobody wrote it into the policy language either way.

Carriers are closing that gap from their side, and not in the direction you'd want. Verisk, the company that drafts a lot of the standard policy language the commercial insurance industry uses, has introduced two endorsements, CG447 and CG448, that explicitly exclude AI-caused losses from standard General Liability coverage. If your insurer has adopted one of these endorsements at renewal, and a growing number have, you may have gone from "probably not covered, ambiguous" to "definitely not covered, in writing" without anyone flagging it to you. Atlas Insurance Agency's broker guidance covers the mechanics of how these exclusions get added to standard commercial policies, often at renewal, often without much fanfare.

The practical takeaway: if you've deployed any AI system that touches customers, money, or decisions with legal weight, and you haven't specifically asked your broker whether your current GL or Tech E&O policy has an AI exclusion endorsement, assume it does or soon will.

What AI liability insurance actually covers (and doesn't)

The policies emerging from Armilla, HSB, and similar carriers generally cover four categories: model underperformance or hallucination that causes a third party financial loss, IP disputes arising from AI-generated content, discrimination or bias claims tied to an automated decision (hiring, lending, claims triage), and third-party losses where your AI system's output was the proximate cause. That's meaningfully different from what cyber insurance and Tech E&O cover.

Cyber insurance is about unauthorized access: someone breaching your systems, stealing data, or holding it ransom. Tech E&O covers professional negligence in delivering a technology service. Neither was designed to answer "our AI agent made a decision that was wrong, and that decision cost someone money." AI liability insurance is the product built to answer that specific question, and it's why the emerging phrase "AI errors and omissions insurance" or "AI E&O" is showing up inside Google's AI Overviews for this topic even though it doesn't have independent search volume yet. It's a named category the market is still settling on.

What none of these policies cover, at least not yet: the cost of fixing a broken AI system, reputational damage that doesn't translate into a quantifiable third-party claim, or losses from AI systems your company never disclosed to the underwriter during the application. That last exclusion is the one that bites hardest, because it means the accuracy of what you tell your broker about your AI stack during underwriting is itself a risk-management decision.

Who is liable when an AI agent makes a mistake?

The honest answer is: it depends entirely on who built and owns the system, and that's the question nobody selling you an AI liability policy wants to spend much time on. There are roughly four parties who can end up holding the bag: your company (as the deployer), the SaaS vendor whose AI feature you turned on, the underlying model provider (OpenAI, Anthropic, whoever), and, if you commissioned custom work, the firm that built your agent.

If you bought a SaaS tool with an "AI-powered" feature bolted on, read the terms of service before you assume the vendor is on the hook. Most SaaS contracts push liability for AI-generated output back onto the customer, buried in a paragraph about "AI features are provided as-is" or "customer is responsible for reviewing AI-generated content before acting on it." You're renting the tool, but you're often still holding the liability, which is a worse deal than it sounds like at signing.

If you commissioned a custom-built agent from a development firm, the contract you signed determines who owns the failure mode, not just who owns the code. This is exactly the kind of clause our post on what to ask before you hire an AI agent development company walks through: what happens when the system produces a bad output in production, who's contractually responsible for monitoring it, and whether the vendor sticks around past go-live or hands you a black box and an invoice.

If you built and own the system yourself, or commissioned it under a contract where you own the IP and the vendor transfers full responsibility for what was built, the liability chain is short: it's you, full stop. That sounds like more exposure, and in one sense it is. But it also means you're the only party an underwriter needs to evaluate, which, as the next section covers, tends to produce cleaner and sometimes cheaper coverage than a liability chain with three vendors' terms of service tangled into it.

How underwriters actually evaluate AI risk before writing a policy

Underwriters are increasingly using the NIST AI Risk Management Framework as their due-diligence checklist, whether or not they cite it by name. NIST's AI RMF 1.0, released in January 2023, organizes AI governance around four functions: govern, map, measure, and manage. In an underwriting conversation, that translates into a fairly specific set of questions, and if you've never had to answer them before, it's worth rehearsing them now rather than during an application.

• Do you log every AI decision with enough detail to reconstruct why it made a specific call, and for how long is that log retained?

• Where are the human-in-the-loop gates? Does a person review outputs before they reach a customer, or does the agent act autonomously on anything above a certain dollar threshold or risk category?

• What data trained or fine-tuned the model, and do you have documentation of where that data came from and how it was handled?

• Is there a model card or equivalent document describing what the system is supposed to do, its known failure modes, and its performance boundaries?

• What's your incident response process when the AI system produces something wrong: how fast do you catch it, and how fast do you fix it?

Companies that can answer these cleanly get better terms and, in some cases, coverage they wouldn't get otherwise. Companies that can't tend to get quoted a higher premium, a narrower scope, or told to come back once they've built the controls. This is the same governance work covered in our piece on AI agent governance and what enterprise teams actually need to control, and it's not a coincidence that the controls a good governance program builds are almost identical to the controls an underwriter asks about. They're solving the same problem from different sides of the table.

Build vs. buy: how your ownership model changes your exposure

This is the section every insurer's blog post skips, because it's not their problem to solve, it's yours, and it happens before the policy conversation even starts. The way you acquired your AI system, not just what it does, determines both your actual exposure and what an underwriter is going to ask for.

A rented SaaS tool with an AI feature is the least controllable option from an underwriting standpoint. You don't have visibility into the model, the training data, or the vendor's own incident history. You're underwriting a black box you don't control, and the vendor's terms of service have likely already assigned much of the liability back to you anyway. Underwriters know this, and it shows up in pricing.

A custom-built system where a third-party firm did the work but you own the IP outright puts you in a stronger position, because you can produce documentation: training data lineage, logging, model versioning, the guardrails that constrain what the system can do on its own. Our guide on production LLM guardrails covers the specific technical controls (input validation, output filtering, escalation thresholds) that reduce both the frequency and severity of the failure modes AI liability policies are actually underwriting against. Fewer bad outputs, smaller claims, better terms.

The strongest position, and the one we build for regulated clients, is a self-hosted open-source model running on the client's own infrastructure with zero data retention. This closes off entire categories of exposure an underwriter would otherwise ask about: there's no third-party model provider whose terms of service you're exposed to, no question of where your data went after an API call, no dependency on a vendor's uptime or policy changes affecting your risk profile. For a healthcare, financial services, or life sciences company where the underlying data is the liability, this isn't an insurance nicety, it's the difference between an underwriter asking three follow-up questions or thirty. If your industry already has its own compliance layer on top of this, our piece on AI agents in financial services and what changes when compliance is non-negotiable covers how those two lenses, insurance risk and regulatory compliance, tend to point at the same architectural decisions.

A pre-deployment checklist to reduce insurable risk before you shop for a policy

Bring this list to your broker conversation and your AI vendor conversation, in that order, before you sign anything.

• Inventory every AI system touching customers, money, or legally consequential decisions, including "invisible" AI features inside SaaS tools you already pay for.

• Pull the terms of service on every one of those tools and find the liability clause. Most say something close to "customer is responsible for reviewing AI-generated output." Know which ones do.

• Ask your current commercial insurance broker directly whether your General Liability policy has adopted a Verisk CG447 or CG448 exclusion endorsement, or an equivalent, at your last renewal.

• For any custom-built system, confirm in writing who owns the IP, who owns the failure mode, and what the vendor's obligations are after go-live, not just at delivery.

• Build (or ask your vendor to build) logging sufficient to reconstruct any AI decision after the fact. If you can't answer "why did it do that" for a specific case, you're not ready for an underwriting conversation.

• Put a human-in-the-loop gate on anything above a defined risk or dollar threshold, and document where those gates sit.

• Decide, honestly, whether your data sensitivity justifies self-hosting instead of routing sensitive data through a third-party API. This is a build decision that happens well before the insurance decision.

None of this replaces a policy. It changes what that policy costs and what it actually covers, because underwriters price certainty, and a company that can answer every item on this list looks like a much better bet than one that can't.

If you're working through this decision, this is close to what our Discovery phase maps out before we build anything, and it's a natural conversation to have before you either buy a policy or sign another AI vendor contract; you can see how we approach it on our enterprise AI page.

Frequently asked questions

Does general liability insurance cover mistakes made by AI?

Usually not, and increasingly the answer is an explicit no. Standard commercial GL policies were written before AI-caused losses were a distinct category, and insurers have started adding exclusion endorsements, notably Verisk's CG447 and CG448, that specifically remove AI-related losses from coverage. Ask your broker whether your policy has adopted one at your last renewal.

What does AI liability insurance actually cover that a normal policy doesn't?

AI liability policies cover losses caused by an AI system's own output or decisions: hallucinated information that causes financial harm, biased automated decisions, IP disputes over AI-generated content, and third-party losses where the AI's output was the direct cause. Standard GL and cyber policies weren't built for any of these scenarios specifically.

Who is liable when an AI agent makes a mistake, the company, the vendor, or the model provider?

It depends on ownership. If you're using a SaaS tool's AI feature, check the terms of service, most push liability back to the customer. If a firm built a custom agent for you, your contract determines who's responsible. If you own the system outright, the liability sits with you, but you also have the cleanest path to demonstrating good governance to an underwriter.

Do I need AI insurance if I'm only using tools like ChatGPT or a vendor's AI features, not building my own?

Possibly, and don't assume the vendor has you covered. Most vendor terms of service explicitly disclaim liability for AI-generated output and place responsibility on the customer to review it before acting. If that AI feature touches customer-facing decisions or money, it's worth the same insurance conversation as a custom-built system.

What's the difference between AI liability insurance, tech E&O, and cyber insurance?

Cyber insurance covers unauthorized access, breaches, and data theft. Tech E&O covers professional negligence in delivering a technology service. AI liability insurance covers harm caused by an AI system's own decisions or output, a category neither of the other two was designed to address, which is why insurers are launching it as a distinct product rather than an add-on rider.