September, 2026
When Your AI Agent Makes a Mistake, Will Your Insurance Pay?
Synergy Group AI Team
If an AI agent at your company causes a loss this year, the law may not let you blame the AI, your insurer may decline to pay, and the main thing left to defend you is your own paperwork.
The last meeting about AI risk probably went like this: someone said “governance,” people nodded, someone promised a policy, and everyone went back to work. That conversation is out of date. Three things changed in 2026, and very few people selling AI have connected them for you. Together, they turn AI from an IT topic into a money problem that lands on your desk.
Start with insurance, because that’s where most companies are exposed without knowing it. Since January 1, new ISO endorsements (CG 40 47, CG 40 48, and CG 35 08) allow carriers to exclude losses from generative AI from standard general liability policies. Some go further. W. R. Berkley offers an “absolute” AI exclusion on its D&O, E&O and fiduciary lines. Your renewal can look almost identical to last year’s: the same broker, the same friendly email, the same premium. But the coverage underneath may have shrunk, and you won’t find out until you file a claim. So, check your renewal. Your coverage may have changed while your premium didn’t.
Then the law. California’s AB 316 (Civil Code 1714.46) also took effect January 1. It says a company that developed, modified or used AI can’t argue the AI caused the harm on its own. There used to be a quiet hope that when an autonomous system went wrong, responsibility would get blurry. That door is closing.
“But we’re not all in California.” Fair question, and the honest answer is that it depends. AB 316 matters most if you have California customers, employees, or operations, and your counsel should explain exactly how it applies to you. But I wouldn’t build a plan around a state line. The insurance exclusions are national and apply wherever you’re headquartered.
The losses are real, too. WitnessAI surveyed 300 executives at companies with more than 1,000 employees. 43% said AI incidents cost them $2 million or more in the past year. About 70% run or pilot autonomous agents, yet only 18% have every one inventoried and approved. And none of it requires a hacker. In April, a coding agent at PocketOS deleted the company’s production database and its backups within seconds. No breach, no rogue employee, just an agent doing something nobody expected with the access it had been given.
Now stack all three. A loss happens. The law says it’s yours. Your insurer points to an exclusion. You’re holding the bill, and nobody did anything that felt reckless at the time. A team ran a pilot, and an agent had a little more access than it needed.
That’s the squeeze, and the only place left to stand is your records. Legal commentary on AB 316 notes that governance and deployment records can support the remaining defenses, such as foreseeability and comparative fault. Insurance guidance tells buyers to bring an agent inventory and an audit trail to underwriting. The same record that helps you in court helps you at renewal. It’s your legal defense and your insurance application in one.
“Won’t records just give a plaintiff more to use against us?”
A smart question, and the one your general counsel will raise first. Records show what you knew and when, so a record that reveals a known problem you ignored can hurt you. But having no record doesn’t make the exposure disappear. It just means you can’t show that you took reasonable care, and you lose the defenses that depend on it. Build the record with counsel involved from the start, agree on what gets kept and for how long, and fix the problems it uncovers.
Whose problem is this?
Everyone in the C-suite has a piece of it, and none of them can pass it along.
The CFO’s worry is an uncovered loss and a nasty surprise at renewal. The first step is an awkward call to your broker asking exactly what changed in your AI coverage this year.
The general counsel’s worry is discovery with nothing to hand over. Your AI vendor isn’t in the courtroom, but your records are. When opposing counsel asks what your agents did, when, and who signed off, “we’re still working on that” is not an answer.
If you lead security or compliance, you probably saw this coming. Agents with too much access and no inventory is the exact situation you’ve been flagging, and you can’t govern agents you can’t count.
The CIO is the one the board will turn to and ask whether the company is exposed. You want a real answer, backed by an inventory and evidence, not a promise to look into it.
Because it’s shared, it tends to fall through the cracks. Pick one executive to own it, with security and IT supporting. Someone has to be able to say “this one is mine.”
What does a record that holds up look like?
Forget the policy PDF and the slide from the ethics workshop. It’s something you could put in front of an underwriter or a judge without flinching. It tells you which agents you have, including pilots and the ones a team spun up on the side. It shows what each can access and do today, not what someone intended when they drew the diagram. It shows who approved it and when. And it lets you reconstruct what an agent actually did after the fact, without depending on anyone’s memory.
Four questions before your next renewal
Sit down with your broker, your counsel, and whoever runs security, and get these answered:
Did our coverage change? Look for the new generative AI exclusions in your general liability, D&O, E&O, and fiduciary policies.
Can we count our agents? Every one running or in pilot, and whether each is approved.
Could we show what they did? If an agent caused a loss tomorrow, would there be a trail of its actions and approvals?
Who owns this? Name one person and set a date for the first report to the board.
If any of those makes you hesitate, that’s useful information. Better to find the gap now, on your own schedule, than after a claim is denied and a lawyer is waiting.
You own your AI agent’s mistakes. What’s still open is whether you can prove you exercised reasonable care. Start with the renewal, then build the record.
IMPORTANT NOTE: This article is for general information only and is not legal, insurance or financial advice. Coverage terms and the reach of laws such as California AB 316 vary by company and jurisdiction. Consult your own counsel and broker before making decisions.