Insurers Say AI Is Too Risky. Here Is What That Means for Your Business

The people whose entire job is pricing risk have looked at AI and started backing away. Major insurers, including Great American, Chubb, and W. R. Berkley, have asked US regulators for permission to exclude AI-related liabilities from standard corporate policies, with one underwriter describing AI model output as "too much of a black box" to underwrite.

That is worth pausing on. Insurers happily cover hurricanes, cyberattacks, and fleets of teenage drivers. They cover those things because they can model them. What they are saying about AI is not that it always fails. It is that when it fails, nobody can predict how, how often, or how expensively.

If your business is adopting AI tools, and in 2026 nearly every business is, this is not an abstract industry story. It is a preview of who holds the bag when an AI tool acting on your behalf gets something wrong. Increasingly, the answer is: you.

The failures insurers are looking at

The hesitation is grounded in incidents that have already happened, not hypotheticals. The same TechCrunch reporting collects three that should get any owner's attention:

• Google's AI Overviews falsely stated that a solar company was being sued by a state attorney general. The company filed a defamation suit seeking 110 million dollars.
• Air Canada's website chatbot invented a bereavement discount that did not exist. A tribunal ordered the airline to honor it. The company was responsible for what its bot said, full stop.
• Fraudsters used a digitally cloned likeness of a senior executive on a video call to trick an employee at Arup, a London engineering firm, into transferring 25 million dollars.

Notice the pattern. None of these are science-fiction failures. They are ordinary business functions, marketing content, customer service, payment approvals, where an AI system introduced a new way to be wrong. And in the Air Canada case, the legal system was clear that "the chatbot did it" is not a defense.

Why the black box framing matters

Traditional software is auditable. When an invoice calculates wrong, a developer can trace the exact line of code and tell you it will not happen again. Insurers, courts, and regulators are comfortable with that kind of causality.

Large AI models do not offer it. Their behavior emerges from billions of learned parameters, so nobody, including their makers, can fully explain why a specific output happened or guarantee it will not recur. That opacity breaks the chain insurers rely on to price a policy: understand the failure mode, estimate its frequency, set a premium.

It also creates what insurers fear most, correlated losses. A defect in one widely used model does not cause one claim. It causes thousands of simultaneous claims across every business that deployed it. That systemic exposure, more than any single incident, is what is driving the push for exclusions.

We wrote a companion piece on the research side of this problem, including OpenAI's work on models designed to be inspectable: what AI transparency means for your business software.

The regulatory floor is rising too

Insurance is only half the squeeze. The EU AI Act, the first comprehensive AI law, reaches full applicability on August 2, 2026, and imposes logging, documentation, and human-oversight requirements on higher-risk systems. In the US, state-level rules like New York's RAISE Act point the same direction: safety protocols, incident reporting, accountability that can be audited.

Most Utah small businesses will never fall under the strictest tiers of these laws. But the direction of travel is clear. The standard being set everywhere at once is: if you deploy AI, you should be able to show what it did, when, and under whose supervision. Businesses that can demonstrate that will find insurance, enterprise customers, and regulators far easier to deal with than businesses that cannot.

What this means practically for a small business

You do not need an AI ethics department. You need a handful of habits that keep AI failures small, visible, and recoverable:

1. Inventory your AI exposure. List every place AI already acts on your behalf: website chatbots, AI email replies, auto-generated content, AI features quietly added to your CRM or accounting tools. Most owners are surprised by the length of this list.
2. Read your policies and your vendor terms. Ask your broker directly whether AI-caused losses are covered, and check whether your AI vendors' terms push liability for output onto you. Most do.
3. Keep a human between AI and anything irreversible. Money movement, contract terms, pricing promises, medical or legal statements: AI can draft, a person approves.
4. Log everything. If an AI tool influences a decision, you want a record of the input, the output, and who approved it. When something goes wrong, that log is the difference between an explainable incident and an indefensible one.
5. Constrain the scope. A chatbot that answers from your actual documentation and says "I will connect you with our team" for everything else is dramatically safer than one improvising answers to any question.
6. Train your people for AI-enabled fraud. The Arup case worked because a video call looked convincing. Institute a simple rule now: any request to move money or change payment details gets verified through a second, known channel, no matter who appears to be asking.

Our point of view: automate the process, supervise the judgment

At Articulate Vision, most of the systems we build are workflow automations, and clients increasingly ask where AI fits into them. Our answer is an architecture rule we apply everywhere: deterministic logic for anything with consequences, AI at the edges, humans at the gates.

Deterministic automation, the kind that routes a lead, files an invoice, or fires an intake reminder by explicit rules, does the same thing every time. It is testable, auditable, and boring in exactly the way money-handling systems should be boring. In the healthcare analytics dashboard we are building for a residential treatment provider, that principle is non-negotiable: every number on screen traces to source records through logic a human can inspect, because clinical and billing decisions ride on it. AI has real value in that kind of system, but as an assistant layer, summarizing, drafting, flagging anomalies for review, never as the unauditable authority on what a metric says.

That is the standard we would suggest for any business, whether we build your systems or not. Let AI accelerate the work. Do not let it silently own decisions you would not delegate to a brand-new employee on their first day, because unlike that employee, it cannot yet explain its reasoning, and your insurer may not stand behind it.

The bottom line

Insurers declaring AI "too risky" is not a reason to avoid AI. It is a market signal about where unmanaged risk currently sits: with the businesses deploying it. The companies that thrive with AI over the next few years will be the ones that adopted it deliberately, with human oversight, audit trails, and automation architectures that fail loudly instead of silently.

If you want AI-augmented automation in your business without betting the business on a black box, that is exactly the kind of system we design.

Talk through an automation project