Insurable

Matthew Grimes

Last week, the Insurance Information Institute counted something most companies haven't noticed. More than 60 property and casualty insurance groups have filed to add some form of AI exclusion to their policies.

It's been building all year. Since January, insurers have had standard wording from ISO, the group that writes the industry's common policy forms, to strike generative AI out of general liability coverage. Berkley has an absolute AI exclusion across its management and professional liability policies. In April, the Financial Times reported that QBE and Beazley were moving to cap some AI-related cyber losses at about 10% of a policy's value.

Most people read this as insurers running from AI. I read it differently. Insurers are telling us exactly what they can't see. As Erike Young, an insurance risk specialist, told Global Finance: "The exposure existed before the exclusion; the exclusion simply made it visible."

Priced

Insurance is how an economy decides a risk is acceptable. An underwriter looks at a building, a ship or a factory and puts a number on what could go wrong. Once a risk has a price, a company can take it on. If nobody can price it, the company has to carry it alone, or walk away.

AI is falling on the wrong side of that line. Lawsuits involving AI rose 978% from 2021 to 2025, according to the broker Gallagher, and 137% in the last of those years alone. So why would an industry built on taking risks turn this one down? Because insurers can see the claims coming, but they can't see what the AI actually did. You can't price what you can't see, so they exclude it.

Cyber insurance has the same problem at a larger scale. Munich Re estimates the whole world spent nearly $15 billion on cyber insurance last year. The FBI counted $20.9 billion in reported internet crime losses in the US alone.

World cyber insurance vs. US reported crime losses

Source: Munich Re, Apr 2026; FBI IC3, Apr 2026

World

~$15B

US only

$20.9B

The whole planet's cyber insurance market is smaller than one year of reported losses in one country. Most digital risk is still carried by the people it lands on.

Fire marks

We've been here before, and the people pricing the risk helped fix it.

In 1666, the Great Fire of London destroyed much of the city. In 1680, Nicholas Barbon opened the Fire Office, one of the first fire insurers. Insurers soon ran their own firefighting crews, hiring Thames watermen to put out fires in the buildings they covered. Each insured building carried a metal plaque called a fire mark, showing it was covered. In 1833, ten of those insurance crews merged into the force that became the London Fire Brigade.

Two centuries later, electricity posed the same question. After the 1893 World's Fair in Chicago, a young electrical engineer named William Henry Merrill started testing electrical equipment for fire risk, with support from fire insurance underwriters. His lab ran its first test in March 1894. It became Underwriters Laboratories, and its mark is still printed on electrical products today.

In both cases, insurers couldn't price a new risk, so they paid to measure it. Measurement made the risk insurable, and insurance made the technology ordinary.

The MFA question

Cyber has its own version of this story, and it's recent.

In 2021, ransomware claims surged. Marsh found that US cyber insurance prices rose 130% in the last quarter of that year. Then insurers started putting one question on their applications: do you use multi-factor authentication? One trade headline summed it up: "For cyber insurance, it's MFA or the highway."

US cyber insurance price increases, 2021

Source: Marsh Global Insurance Market Index, Q4 2021

50%

100%

0%

+96%

+130%

Q3 2021

Q4 2021

For many companies, the renewal form did what years of security advice hadn't. The insurer didn't need to understand every server. It needed one control it could check.

AI is waiting for its MFA question. Early trade accounts suggest what it will look like. According to a Cloud Security Alliance research note, insurers writing AI coverage want four things. They want an inventory of the AI systems a company runs and a tested way for a person to override them. They want one named executive who signs off. And they want proof the controls are enforced, through logs and identity systems.

Read that list again. Every item on it is a security record.

The record an underwriter can read

An underwriter can't read a model's mind. Nobody can, yet. But an underwriter can read a record: which AI systems exist, which identities they act under, what they touched, and who approved what they did.

Keeping that record is what cybersecurity already does. It has to meet two tests. First, nobody can quietly edit it after a claim is filed. Sign each entry as it's written, chain it to the one before, and let an outsider check the chain. Second, the record has to separate what the AI read from what a person decided. If an attacker can plant text in a log that an AI then acts on, the record shows a decision nobody made.

This is the part most companies miss. When a claim comes in, the adjuster asks the same questions an auditor would. What happened? On what evidence? Who decided? A company that can answer from a signed record has a claim. A company that can only offer a reconstruction has an argument.

What an underwriter can check

Coverage

AI inventory

Identity and access

Signed record

Named approver

That's the record we build at ARRTECH. Our AI reads the evidence and investigates; it doesn't block, change or approve anything. Our AI detection is available to select organizations, and our AI investigation is in early access. Every write to the log is signed and chained, and a standalone verifier checks it without us. Every approval is on record.

The prize

This matters most to companies that can't carry risk alone.

A global bank can self-insure its AI mistakes. A regional hospital, a 200-person manufacturer or a city government can't. For them, an uninsurable risk is a technology they don't adopt. Standalone AI liability policies already offer limits from $2 million to $50 million, and Munich Re expects cyber premiums to nearly double, to around $28 billion, by 2030. The capacity is coming. What it waits on is evidence.

What I'm not saying

I'm not saying insurers are wrong to exclude AI. Pricing what you can't measure is how insurers go broke. An exclusion is an honest answer.

I'm not saying a policy makes AI safe. Insurance moves the cost of a loss. It doesn't prevent one.

And the list of what insurers want comes from early trade reporting. Carriers haven't agreed on a common standard yet.

What would change my mind: if insurers start covering AI broadly without asking for any evidence, then the record doesn't matter to them, and I'm wrong.

What kind of world?

There's a world where AI stays uninsurable. The largest companies adopt it because they can absorb the losses. Everyone else waits, or goes without cover and hopes.

There's another world, where a company can show an underwriter what its AI did, on what evidence, and who approved it. In that world, AI gets a price like every risk before it. Hospitals, factories and city halls adopt it with confidence, because someone has checked and someone will pay if it goes wrong.

Three and a half centuries ago, a small metal plaque on a London house meant exactly that. AI needs its own fire mark.

Thanks for reading,
Matthew