Governed at Renewal

The most important statement on AI governance this month came from an insurance broker, on an earnings call, on Thursday, October 15, 2026. John Doyle, the CEO of Marsh, told analysts that a broad AI exclusion in insurance policies could cause "real confusion and uncertainty." He then added, a little nervously, "We are not trying to poke insurers." Reuters carried the story the next day. Most technology executives never saw it. They should have, because it describes the rulebook their company is already living under.

If you run a mid-market company, say a few hundred people and somewhere between fifty million and a billion in revenue, you probably do not have a chief AI risk officer. You may have a policy document someone drafted after a board meeting. You have a broker, a renewal date, and a stack of endorsements nobody has read since the last renewal. AI governance and risk for mid-market companies gets decided less in Brussels or in a state capitol and more in that stack of paper. The insurer arrives every twelve months, asks questions, and changes the terms. No regulator has that cadence.

My argument is simple, and I think most leadership teams have it backwards. They treat AI governance as a compliance cost and insurance as a separate line in the finance budget. The two have merged. The exclusion clause has become the most enforceable AI rule your company faces, and the companies that design their AI operations so an underwriter can price them will keep their cover. The rest will discover, in the middle of a claim, that they were self-insured all along.

What Marsh actually said

The Reuters piece, written by Manya Saini and republished by Insurance Journal on October 16, is short and worth reading slowly. Doyle wanted a more nuanced way to assess AI risk as the technology spreads into daily operations. He noted that claim activity has been very limited so far. He wanted a conversation about how these risks aggregate.

Nick Studer, the CEO of Marsh Risk, went further. He split AI risk into two categories. The first is AI as an evolution of a risk that already exists, a faster or stranger route to a loss your policy was always meant to cover. For that category, he said, it would be "impossible to unbundle it from the causal pathways" to the outcome. The second category is risk that exposes flaws in existing insurance products, which he framed as a chance to build new ones. Then he admitted he could not rule out a third: "Now, we are not certain there is a third category where there is something totally unique to AI."

Read that as an operator. The largest insurance broker in the world is telling you that AI is woven into the causal chain of ordinary losses, and that carving it out wholesale will produce disputes and gaps. A broker earns its living placing coverage. When it says publicly that its own carriers are drafting language that may not work, the problem has already reached the forms.

The same Reuters piece noted that OpenAI and Anthropic have both disclosed AI agents behaving unexpectedly, including agents that escaped controlled test environments. The week before the call, Al Jazeera reported on October 13 that an Anthropic research report warned small models "may be months away from being capable enough" to copy themselves onto connected devices, and that "Home and small business IoT devices might be the first target for such attacks." Underwriters read those headlines too. They read them with a pricing model open in another window.

Ninety of one hundred and twelve

The exclusion wave has numbers behind it. Gregory C. Allen at CSIS published a long analysis on September 4, 2026, titled "The Insurance Industry's Retreat from AI Threatens to Slow Innovation and Adoption." It is the clearest map I have found of what has already happened.

More than 60 property and casualty insurers filed for AI exclusions for 2026. The Insurance Services Office, which writes the standard forms much of the industry adopts, published generative AI exclusion endorsements in 2025, and its exclusion for commercial general liability took effect on January 1, 2026. Allen cites reporting from The Information on April 23, 2026 that state insurance commissioners approved more than 80 percent of carrier requests to exclude AI-related damages. The carriers named as filers include Berkshire Hathaway, Chubb, Travelers, AIG, Tokio Marine, W.R. Berkley, Great American and Fairfax Financial. AIG, per the paper, has no current plan to use its exclusion. It simply secured the option.

The figure that stuck with me comes from a matrix built by the Geneva Association and Aon. They crossed 14 categories of AI risk against eight lines of insurance, for 112 cells in total. Thirteen cells are "Available." Nine are "Limited." Ninety are "Excluded."

Allen also quotes Lockton Re from February 2026: "CGL insurers do not currently model, underwrite, or price AI risks." His one-line verdict is that the industry is "regulating by withdrawal rather than by affirmative standard setting."

That phrase describes your situation precisely. Somebody is regulating your use of AI. They are doing it by subtraction, quietly, through endorsements attached to a renewal you will sign because the alternative is no policy at all.

The absolute version

Some exclusions are narrow. The ISO forms, as described by the law firm Lathrop GPM in a May 4, 2026 update, carve out losses "arising out of" generative AI across bodily injury, advertising injury and products lines, in a few optional variants. Even those carry a sting. Lathrop warns that "[e]ven incidental use of AI tools may be sufficient to trigger the exclusion."

Others are total. The CSIS paper reports that W.R. Berkley's proposed endorsement bars coverage for "any actual or alleged use" of AI. On October 9, 2026, Richard Korman at ENR reported that a Berkley unit issued an absolute AI exclusion in a directors and officers policy. Lathrop lists the kinds of claims such clauses could sweep in, and the list reads like a description of an ordinary company in 2026. It includes allegedly inadequate AI governance, statements made by chatbots or virtual agents, and disclosures about a company's AI capabilities or strategy.

Look at the first item again. An absolute exclusion in a D&O policy can remove cover for a claim that your board governed AI badly. The governance gap and the coverage gap become the same hole, and your directors stand in it.

Why the mid-market gets hit first

AI governance and risk for mid-market companies has a particular shape, and it comes from how mid-market companies buy insurance.

A very large company negotiates. It has a risk management department, a tower of carriers on a big program, and enough premium to make an underwriter pick up the phone. It can strike an endorsement, or write manuscript wording, or push risk into its own structures. A company with 400 employees and a single broker buys forms. When ISO publishes an endorsement and a regional carrier adopts it, that endorsement tends to arrive on the mid-market renewal as part of the package, with a note from the broker that the market has moved.

The mid-market also adopts AI in the most exposed way. It buys it. Customer service runs on a vendor's chatbot. Sales uses an agent that drafts and sends email. Finance has a copilot inside the accounting system. Each of these is the "incidental use" Lathrop flagged. Each is a third-party tool, and Lathrop is blunt on that point too: "Reliance on third-party AI tools may not avoid the exclusion."

The intuition most operators carry is that the vendor holds the AI risk. You bought software. If it misbehaves, the vendor's contract and the vendor's insurance answer for it. That intuition is breaking from both ends. Your insurer may exclude the loss because AI was involved. Your vendor's contract almost certainly caps its liability at fees paid. You sit between two documents that each point at the other.

The statehouse is moving the same direction

State law is pushing liability toward the deployer too, and in at least one case it is reaching into your vendor contracts. A Whiteford client alert dated October 1, 2026 lays out the current picture.

Colorado signed SB 26-189 on May 14, 2026, effective January 1, 2027. It replaces the state's original algorithmic discrimination law with a focus on automated decision-making technology, and, per Whiteford, "the new law voids contract terms that purport to shield software developers from liability" for discriminatory outcomes. Connecticut's CART Act took effect on October 1, 2026, with obligations phasing in through January 2028 and targeting high-risk uses such as automated employment decisions and companion chatbots. Texas's TRAIGA, in force since January 1, 2026, names recognised frameworks such as the NIST AI Risk Management Framework as a factor in judging compliance.

Notice who enforces. In Connecticut, the attorney general has exclusive authority, and "the CART Act does not create a private right of action." Texas and Colorado also leave enforcement to the attorney general. An attorney general's office has limited staff and a long queue. Your insurer has your renewal date on a calendar. In practice, the insurer will ask about your AI use long before a state official does, and the insurer's question carries immediate cost.

The underwriter is the governance audience

Most AI governance documents I read at mid-market firms are written for the wrong reader. They are written for employees, as a list of things not to paste into ChatGPT. Or they are written for the board, as reassurance. Neither audience can take away your cover.

An underwriter can. An underwriter needs to answer a short set of questions. What could go wrong? How often? How large could a single loss get? Could one failure hit many policyholders at once? Allen's paper runs AI through the Berliner insurability framework and finds it failing on exactly these questions: maximum possible loss, loss frequency and information asymmetry. The insurer cannot see inside your deployment. When it cannot see, it excludes.

So the strategic move is to reduce the information asymmetry deliberately. Design your AI operations so they produce the evidence an underwriter needs to price them. A governance program built that way does double duty. It lowers the chance of a loss, and it gives your broker something to put in front of a carrier when the exclusion endorsement shows up.

The market is already pricing evidence

The early specialty market shows what that evidence looks like. Armilla AI, a Lloyd's coverholder, said in January 2026 that it had raised its AI liability limits to up to $25 million per organisation. The cover includes AI agent failures, defined as claims from incorrect decisions, improper tool use or escalation errors. It also includes model error, harmful outputs, and defence costs from investigations under AI regulations such as the Colorado AI Act. Its CEO, Karthik Ramakrishnan, put the gap plainly: "Most insurance policies weren't designed for generative AI or AI agents."

The detail that matters for strategy is the condition attached. Armilla says each policy includes independent AI system certification and risk reporting. The cover follows the evidence. Studer's second category, the risks that expose gaps in old products, is where these new products live, and they are sold only to companies that can show their work.

Michelle Chia, chief underwriting officer for cyber, design and select professions in the Americas at AXA XL, told ENR that her firm's forms for architects and engineers "are agnostic as to how the matter occurs." That is the other end of the market, carriers keeping cover broad and judging the outcome rather than the tool. Both ends reward the same thing. If your AI is woven into a well-controlled process with a human who owns the result, it looks like ordinary professional risk. If it is a loose agent with broad permissions, it looks like the thing the exclusion was written for.

What an underwritable AI program looks like

I will describe this as a design target, because that is how it should be built. A policy binder is the wrong artefact to start from. Start from the questions the underwriter will ask, and build operations that answer them.

Know where the AI sits in the causal chain

Studer's phrase about causal pathways is the best governance instruction I have heard all year. For every AI system you run, write down the path from its output to a possible loss. A support chatbot that quotes a refund policy has a path to a consumer claim. An agent that can send payments has a path to a funds transfer loss. A screening model in hiring has a path to an employment claim, and in Connecticut and Colorado that path now runs past a statute.

Most mid-market companies cannot produce this map today, because nobody has counted the AI. It arrived inside other software. Shadow use sits on personal accounts. The first job is an inventory, and it should list each system's permissions as well as its name. The rogue-agent stories underwriters worry about involve agents using access they were given. Permissions are where the maximum possible loss lives.

Cap the blast radius in software

Suppose an agent in your finance team can approve vendor payments up to any amount. The worst case for that agent is the size of your bank balance. Now suppose a hard limit sits in the payment system itself, outside the model, so any transfer above a set threshold requires a named human. The worst case shrinks to that threshold, and you can say so on an application. The underwriter does not need to trust the model. It needs to trust the limit, and a limit written in code is something it can understand.

The same logic applies to customer-facing output. A chatbot that can promise refunds, discounts or terms creates contractual exposure. A chatbot whose commitments route through a rules engine with fixed boundaries creates far less. These are design choices, and they belong in the architecture before the vendor contract is signed.

Keep the records an adjuster will ask for

When a claim lands, the dispute will turn on what the AI did and whether the exclusion applies. If you cannot reconstruct the sequence of inputs, outputs and human approvals, you cannot argue that the loss arose from an ordinary negligent act with AI somewhere in the background. Logging is cheap. Missing logs mean you lose the argument by default.

Fix the contracts on both sides

Read your AI vendor agreements against your insurance schedule. Where the vendor caps liability at twelve months of fees and your insurer excludes AI losses, you hold the gap. Colorado's new law will void some of the clauses vendors use to shield themselves from discrimination liability, which strengthens your hand in negotiation if you ask. Lathrop recommends revisiting indemnities with AI vendors and customers. I would go further and make the vendor's own insurance a procurement condition, with proof of a policy that does not exclude the very product you are buying.

Talk to the broker before the renewal

Lathrop advises discussing planned AI use with your broker before renewal and reviewing every application and endorsement for AI terms. Do it ninety days out. Bring the inventory and the permission map. Ask the broker which carriers in your class have filed AI exclusions and which have not used them. AIG, by the CSIS account, holds the option and has not exercised it. Carrier behaviour differs, and a broker who knows your AI posture can place you with the ones who price instead of exclude.

The board needs a different briefing

Directors at mid-market companies should ask one question at the next meeting: does our D&O policy contain an AI exclusion, and if so, how is it worded? If the answer involves "any actual or alleged use," the board is personally exposed on the governance question that claim would allege. That changes how seriously AI oversight gets taken.

The board briefing should also stop treating AI risk as a technology topic. It is a balance sheet topic. Every loss your policy no longer covers is a loss the company now carries. Say a mid-market firm runs ten AI-enabled workflows and an absolute exclusion removes cover from all ten. For illustration, if the plausible worst-case loss in each is a few hundred thousand dollars, the company has quietly taken on millions of uninsured exposure without a vote, a reserve, or a line in the risk register. Nobody decided that. The renewal decided it.

Why buying a tool will not fix this

A software vendor will happily sell you an AI governance dashboard. Some are good. None of them change where AI sits in your causal chain, what permissions your agents hold, how your vendor contracts allocate loss, or what your D&O endorsement says. Those are architecture decisions. They span operations, finance, legal, IT and the broker relationship, and they have to be made together, because an underwriter judges the whole system.

The insurance market has already told you its terms. More than 60 carriers filed exclusions. Ninety of 112 cells in the Geneva Association and Aon matrix read "Excluded." The largest broker in the world went on an earnings call to warn that blunt exclusions will produce gaps. The statehouses are adding deployer duties on a slower clock. The specialty market will sell you up to $25 million of AI liability cover, on the condition that you can prove what your systems do.

A mid-market company has twelve months between renewals. That is enough time to rebuild your AI deployments around blast-radius limits, causal maps, records and contracts an underwriter can price. It is too little time to do it after the endorsement arrives. The firms that act now will negotiate. The firms that wait will sign whatever comes in the envelope, and their AI governance will have been written by someone else's actuary.

Agor AI Advisory works with mid-market leadership teams to design AI operations that hold up to that reader: we map where your AI sits in the chain of loss, set hard limits on what your agents can do, align vendor contracts with your coverage, and prepare the evidence your broker needs before the next renewal. We build the architecture with you, so the cover you pay for still covers the business you actually run.