Photo by Amina Atar on Unsplash
The Endorsement Nobody Sent a Letter About
It's renewal week. The broker forwards a 40-page general liability quote, the premium is roughly flat, and the cover note says there are no material changes. On page 31 sits a new three-paragraph endorsement removing coverage for any claim arising out of the use of artificial intelligence. Nothing about it shows up on the premium line. As of July 28, 2026, that is the shape this problem is taking for most small businesses — not a price shock, a quiet subtraction.
According to refresh, which flagged the trend in commercial lines, carriers have been folding artificial-intelligence exclusions into ordinary business policies rather than repricing them. That distinction matters more than the headline does. A premium increase is visible; you argue about it, you shop it, you budget for it. An exclusion is invisible until the day you file a claim, and by then the argument is over.
The useful question is not whether AI exclusions exist. It's which of the three flavors landed on your policy — because they are not close to equivalent, and most brokers will not volunteer the difference.
Why Carriers Moved Before the Losses Showed Up
The obvious reading is that insurers are watching AI claims pour in and slamming the door. The evidence does not really support that, and it's worth saying plainly: the surface story gets the causation backwards.
Underwriting depends on loss history. An insurer prices commercial auto because it has decades of crash frequency and severity data — how often, how much. Generative AI deployed at scale in ordinary businesses has almost no comparable loss record, and the claims that do exist are still working through litigation, which means nobody knows yet what a typical payout looks like. You cannot build a rate for a peril whose severity curve hasn't been drawn. Faced with that, an underwriter has two honest options: charge a large uncertainty margin, or exclude and wait. Excluding is cheaper for everyone in the short run, which is precisely why it happened first.
The second-order concern is the one that actually keeps reinsurers awake — correlation. Ordinary liability risk is diversified: one restaurant's slip-and-fall has nothing to do with another's. AI risk isn't shaped like that. If thousands of policyholders run the same third-party model and that model has a systematic defect, the losses arrive together, on the same day, from the same root cause. This is the aggregation problem that made "silent cyber" such an expensive lesson for the market in the 2010s, when policies that never mentioned hacking ended up paying for it anyway. Carriers learned that lesson at real cost and are now applying it early. Silent AI is the same movie with a new cast.
A fair skeptic pushes back here: isn't this just risk-averse carriers offloading a normal cost of doing business onto the customer? Partly, yes. But the criticism that lands isn't that exclusions exist — it's how they're drafted. Some of the broad wording captures "any claim arising out of" AI, and in an economy where spell-check, routing software, fraud screening, and customer-service chat all involve machine learning, "arising out of" is an enormous phrase. A bakery that uses an AI-powered scheduling tool has no meaningful AI exposure. Under a broadly worded exclusion, a coverage lawyer could still try to use that tool to contest an unrelated claim. That is the real problem: not risk selection, but wording that reaches past the risk it was written for.
The Three Versions, and Who Actually Loses Under Each
Here is the comparison no single renewal document will lay out for you. Three drafting approaches are circulating, and your exposure depends entirely on which one you got.
The absolute exclusion. All AI-connected claims are out, with no exceptions. This is the worst outcome for a business with only incidental AI use, because it converts an ordinary claim — a client sues over bad advice, a customer over a defective product — into a coverage fight the moment the defense discovers software in the chain of events. The party who loses here is the low-risk business, which gives up broad policy coverage in exchange for removing a risk it never had.
The exclusion with a carve-back. AI is excluded, except for incidental or embedded use — the AI features inside the software you licensed, where you neither built nor trained the model. This is the version most small businesses should be fighting for, and it costs nothing but a conversation. Under a carve-back, the bakery keeps its coverage and the AI-development shop still doesn't; the exclusion does the job it was written to do and stops there.
The affirmative AI endorsement. Coverage is granted back explicitly, usually with its own sublimit (a cap that applies only to that category of claim, lower than your overall policy limit) and its own conditions — typically human review of AI output, documented testing, and disclosure to customers. Specialist markets, including syndicates at Lloyd's of London, have built products in this space aimed at model performance failure. This is the right answer only for businesses whose product IS the AI output.
Put those side by side and the decision rule falls out. If AI is something your vendors sell you, you need a carve-back, not a new policy. If AI is something you sell your clients, a carve-back won't save you and a standalone endorsement is the real conversation. The expensive mistake is a business in the first category being sold a product built for the second — and the equally expensive mistake is a business in the second category assuming its existing errors-and-omissions cover still responds.
The Cheaper Move Most Owners Never Make
Before shopping for anything new, three steps cost nothing.
First, request the endorsement schedule — the list of every form number attached to your policy — and search the actual forms for "artificial intelligence," "machine learning," and "algorithm." Not the summary. The forms. An exclusion that was added at renewal will be there, and roughly nobody reads that far.
Second, ask your broker in writing for a carve-back for incidental and embedded AI use. This is a negotiable term, not a law of nature, and asking is free. Get the answer in writing either way, because a broker's written refusal is itself useful documentation.
Third — and this is the step that most often makes a new policy unnecessary — read the indemnity clause in your AI vendor's contract. If a licensed tool causes the loss, the vendor's indemnity and its own coverage may be the first line of response, ahead of anything you buy. Contractual risk transfer is almost always cheaper than premium, and it's the tool small businesses reach for last instead of first.
Two footnotes worth knowing. Carriers are deploying AI on their own side of the table too, in automated claims management and risk assessment — which means the same technology being excluded from your policy is increasingly deciding how your claim gets triaged. And the compliance backdrop is shifting underneath all of this; the state-level obligations that Smart AI Trends sorted from the noise in its rundown of confirmed U.S. AI rules will eventually feed into how underwriters score applicants. Documented AI governance is starting to look like a rating factor, not just a legal checkbox.
Frequently Asked Questions
Does my business policy cover AI errors if it doesn't mention AI at all?
Silence is not the same as coverage. A policy that never mentions AI may still respond, but that ambiguity is exactly what the industry called "silent cyber" — and it usually gets resolved in litigation rather than at the claims desk. An express carve-back is worth far more than an absent exclusion.
Do I need a separate AI liability policy for a small business in 2026?
For most businesses that merely use AI features inside licensed software, no — a carve-back on the existing policy plus a solid vendor indemnity typically does more for less. Standalone AI cover is built for businesses that develop, train, or resell AI output, where the model itself is the product.
Can my insurer add an AI exclusion at renewal without telling me?
Notice requirements vary by state and by line of business, but a renewal quote is legally a new offer, and a materially narrower form can appear inside one. The practical defense is comparing endorsement schedules year over year — an insurance comparison at the form level, not the premium level.
Bottom Line
- As of July 28, 2026, AI exclusions are being added to standard commercial policies quietly, with little or no change to premium — making them easy to miss at renewal.
- Three drafting styles are circulating: absolute exclusion, exclusion with an incidental-use carve-back, and an affirmative AI endorsement with its own sublimit. They are not interchangeable.
- Businesses that only use embedded AI features should negotiate a carve-back — free to ask — rather than buy a new policy built for AI developers.
- A vendor's contractual indemnity is often the cheapest first line of response and is routinely checked last.
Our analysis: the current wave of broad exclusions reads as a placeholder, not a settled position. Insurers exclude what they cannot yet price, and they narrow those exclusions once loss data arrives and a competitor starts winning accounts by offering coverage. The more likely outcome over the next few renewal cycles is a normalizing market — carve-backs becoming standard, affirmative endorsements getting cheaper, and today's absolute exclusions looking as dated as the first blanket cyber exclusions do now. On balance, the businesses that get hurt in the meantime won't be the ones with real AI exposure. They'll be the ones with almost none, who never read page 31.
Explore Our Network
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute insurance advice. Policy wording, availability, and regulatory treatment vary by carrier and by state. Always consult a licensed insurance agent or coverage counsel before making decisions about your policy coverage. Research based on publicly available sources current as of July 28, 2026.
No comments:
Post a Comment