Skim time: 6 minutes · Listening time: 5 minutes. Prefer to listen? Flip reads this week’s Brief here.

---

Over the last few weeks, several clients have asked me to talk to their management teams about AI risk. Too much of the insurance industry’s AI conversation starts and ends with governance.

That misses the point.

Governance is necessary. But it is not the business objective.

Insurance operators should be focused on three things:

1. Decision quality. Are we making better underwriting and claims decisions, or just making them faster?
2. Liability and loss exposure. What happens when AI gets it wrong, who bears the loss, and what insurance actually responds?
3. Operating economics and resilience. Are we improving the economics of the business, or adding cost and dependencies we cannot operate without?

Governance should support those outcomes. It should not become a substitute for them.

I wrote three essays for those management teams, one per question. Here’s the short version of each, with a link to the full essay in each section:

• The files nobody fights. Medicare Advantage patients appealed about one denial in nine. Four in five appeals overturned the denial, at least in part.
• The wording waiting for you. 41 insurance groups have filed to adopt AI exclusions. Your renewal is where the wording lands.
• The math after the demo. An illustrative $375,000 AI saving, and what’s left once you pay to use it.

---

The AI errors nobody challenges are the ones your board can’t see

Your AI governance is probably excellent, but that’s not the same as it working.

UnitedHealth, the country’s largest health insurer, is in discovery over nH Predict, an algorithm plaintiffs say it used to cut off post-acute care for Medicare Advantage patients against their doctors’ advice. UnitedHealth disputes the allegations and nothing has been proven, but the breach-of-contract and good-faith claims survived its motion to dismiss. The plaintiffs allege roughly 90% of appealed denials were reversed.

The 90% gets quoted far more often than the appeal rate. KFF, the health policy research group, found that Medicare Advantage enrollees appealed just 11.5% of denied prior-authorization requests in 2024, and 80.7% of those appeals ended with the denial partly or fully overturned. Those figures cover prior-authorization decisions generally, not just AI-assisted ones. A reversal isn’t proof the first decision was wrong, but the files nobody appealed stay a blind spot.

When someone does fight back, it gets expensive. USAA offered its own insured $10,000 on a $250,000 underinsured-motorist policy, with no AI involved, and tendered the limit only on the eve of trial. In January 2025 a Las Vegas jury returned a $114 million bad-faith verdict, $100 million of it punitive. USAA disagrees and has appealed, and the judge has already cut the punitive award to $63 million.

The challenged files tell you something about those decisions, but nothing about whether the unchallenged ones were right.

Read the full essay: The AI Errors Nobody Challenges Are the Ones Your Board Can’t See (every source linked inside)

So what?

The board question doesn’t wait for a verdict: what was management measuring while those decisions were being made?

Picture a regional carrier using an AI tool to recommend reserves and settlement offers on auto injury claims. Handling time falls and most claimants accept. The few who hire a lawyer or send a time-limit demand are the only place a low offer shows up, and they’re where the cost lives.

If the problem is systematic, the same faulty logic may affect an entire class of files. Ask how your ICPL (insurance company professional liability, the carrier’s own errors-and-omissions cover) would treat related allegations arising from that pattern: one claim with one retention and one shared limit, or multiple claims.

A faster dashboard measures production. Before the next deployment or expansion, ask management for four things:

1. A baseline. AI-assisted files compared with independently reviewed files or a control group.
2. Errors weighted by consequence. A formatting slip and a missed material exposure shouldn’t count the same.
3. Samples of files nobody challenged. Reopens and attorney letters should start an investigation, not stand in for one.
4. A record of each decision. What the tool recommended, which version produced it, and what the adjuster did with it. If a file turns into a bad-faith suit, that record may be central to reconstructing and defending the decision.

Underwriting gets the same test. A renewal nobody objected to hasn’t proven the price was right.

Governance establishes responsibilities and controls. Testing establishes whether the work deserves confidence.

Ask your board: Where would we see it if our AI-assisted decisions were worse, and what are we doing to find the errors nobody challenges?

Finding the error is half of it. The next question is who pays for it.

---

Who pays when your AI gets it wrong? Most firms are assuming someone else will

AI exclusions are moving through the market, and your renewal is where the wording becomes your problem.

ISO, which writes the industry’s standard policy forms, issued three optional generative-AI exclusions for commercial general liability, dated January 2026. In July, The Insurer counted 41 property and casualty groups with at least one company filing to adopt an AI exclusion. Another 20 filed to delay adoption. Filing does not necessarily mean implementation.

The language is reaching past general liability into D&O (directors and officers) and E&O (errors and omissions) forms too. In financial lines it often arrives as an endorsement on the renewal quote, not a state filing. Don’t wait for the quote to read it. Some carriers are holding it in reserve: AIG reportedly has no current plans to apply its exclusion, but it has secured the option.

The market isn’t only moving one way. Armilla, backed by Lloyd’s, and Munich Re sell affirmative AI cover, and QBE has added a separate, capped limit for fines under the EU AI Act.

Affirmative AI coverage deserves consideration, but it still has to be matched to the failure you’re trying to insure.

Read the full essay: Who Pays When Your AI Gets It Wrong? (every source linked inside)

The LION Lens

Take an MGA, a managing general agent that writes business on a carrier’s paper under delegated authority. Its AI tool summarizes submissions and misses a change in an insured’s operations. The account renews, and the capacity provider later challenges the terms. That one file can draw five different allegations. A carrier and a board face their own versions:

The allegation	Where to test for a response	Ask before renewal
MGA: “You wrote outside your authority”	Your E&O, including any exclusion for claims by your own carrier	Does our E&O respond when our carrier says AI triage skipped a referral?
MGA: “You underwrote it badly”	Your E&O professional-services grant, its treatment of underwriting, and applicable exclusions	Does our E&O cover decisions made with AI help? Does the renewal add an AI exclusion?
MGA: “The information was wrong”	Your E&O, and what the agreement said each side had to check	What did our agreement say we had to verify, and did we?
MGA: “The tool missed it”	The vendor’s contractual obligations, liability cap, indemnity and remedies	Is the vendor’s liability cap anywhere near our exposure?
MGA: “You broke the agreement”	The delegated-authority agreement’s remedies, against your E&O’s contractual-liability exclusion	Does our E&O exclude liability we took on by contract?
Carrier: “Your AI lowballed my claim”	Your ICPL and any ECO/XPL (extra-contractual and excess-of-limits) provisions. Separate the underlying policy benefits from the alleged extra-contractual loss	Would an AI exclusion on our ICPL reach its ECO/XPL extension? Are punitive damages covered where the law allows?
Board: “You overstated your AI, or never oversaw it”	D&O, including any Side A protection for directors when the company can’t cover them, against the allegations and endorsements	Could an AI exclusion reach Side A, or claims about what we’ve said publicly about AI?
Anyone: “The AI vendor went down”	Cyber: dependent business-interruption terms, covered-provider definitions, system-failure triggers and limits	Does our dependent business-interruption cover respond to a vendor outage nobody hacked?

So what?

For MGAs this isn’t remote. In the UK, more than 80% of MGAs use AI, but only about half have a formal governance framework. Among Lloyd’s managing agents, the capacity side of that relationship, 72% already have one and another 21% are building one. The gap sits exactly where the delegated-authority agreement does.

Keep two maps: where your business uses AI, and where your insureds use it. An exclusion on the policies you issue does nothing for your own operational liability. Your own AI controls say nothing about your insureds’. Where the wording is silent, read further before anyone calls it covered.

The LION POV

Here’s how we’re advising clients: make every retained AI loss a decision, written down and sized to the benefit. For each material deployment, a one-page loss-allocation schedule answers four questions:

• Who owes the obligation if this fails?
• What remedy does the contract give us, and what caps or carve-outs limit it?
• Which policy section might respond, and what wording could stop it?
• What stays with us?

A contractual remedy is not an insurance recovery, and neither should be assumed because someone else supplied the technology.

The vendor agreement, the delegated-authority agreement, your own program and the policies you issue were written by different parties for different purposes. Nobody drafted them to be read together, which is exactly why they need to be.

This is the work we do with clients before renewal. We trace each AI scenario through the vendor contract, the delegated-authority agreement and the policy wording, while there’s still time to negotiate. Want that read on your own program? Reply to this email or book a time.

Ask your board: Which AI-related losses have we deliberately accepted, and which are we assuming someone else will pay?

Even a loss you’ve insured still costs you the operation while it’s down. The last essay puts a price on that.

---

Your AI business case is missing the cost of depending on it

An AI business case can look compelling right up until someone asks what it costs to keep the savings.

The Boston Consulting Group’s newest survey of 1,330 executives, published September 30, found that just under half said their companies now get meaningful value from AI. That leaves the other half still waiting. In its 2024 survey, BCG traced about 70% of the challenges to people and process, and only 20% to the technology.

The usual business case leaves most of the cost out. Take an illustrative mutual’s claims team. The figures are assumed, not client results: 10,000 files a year, an AI summary tool that saves 30 minutes a file, and staff at a fully loaded $75 an hour. Count what it costs to use the tool, and about 89% of the gross benefit disappears:

	Per year
Gross time saved (5,000 hours)	$375,000
Quality review, 10 minutes a file	($125,000)
Rework, 8% of files at one hour each	($60,000)
License and integration	($150,000)
Net, before anything breaks	$40,000

Review and rework are assumed to be new work on top of the prior process. License and integration is the annual cost assigned to this example.

Dependence already has a price on record. On July 19, 2024, one faulty CrowdStrike software update crashed Windows systems worldwide, with no AI involved. Parametrix, a cloud-outage insurer, estimated the direct loss to U.S. Fortune 500 companies, excluding Microsoft, at $5.4 billion, and put the insured share at no more than $540 million to $1.08 billion.

One vendor, one update, and most of the loss stayed with the companies that relied on it.

Read the full essay: Your AI Business Case Is Missing the Cost of Depending on It (every source linked inside)

So what?

Even the $40,000 is the generous version.

It assumes every saved hour turns into cash. If the released hours don’t reduce expense or avoid future spending, the labor benefit is capacity, not cash. The $150,000 vendor-and-integration cost still has to be funded. The margin is thin too: if rework rises to about 13% of files, with each reworked file still costing an hour, the net falls to zero.

A year into relying on AI summaries, could the mutual’s adjusters absorb the original work at current staffing if the service went down? For an MGA the dependence runs both ways. Below it sits the AI vendor. Above it sits a capacity provider whose agreement assumes you keep underwriting to standard, whatever happens to your tools. If that capacity provider asks how you’d keep writing through a week-long outage, the answer should already be on paper.

Ask management for a plain account of what stops, what slows and what still runs without the tool, and who would notice if a vendor’s model update made it quietly worse. Then test the fallback once and write down how long the manual work took. Exit terms (what data you can export, what has to be rebuilt) live in the vendor contract, where they’re a negotiating point before signing and a fixed cost after. Check the last row of the table in the second essay too. Does your cyber policy’s dependent business-interruption cover respond to an AI vendor outage, and at what sublimit and waiting period?

Losing a drafting assistant is an annoyance. Losing the system staff use to retrieve the evidence behind a claims decision stops the work.

A credible AI business case shows three numbers: the expected benefit, the full operating cost, and the cost of losing the capability.

Ask your board: After the full cost and the failure scenario, what operating advantage have we built, and can we keep it?

---

The Bottom Line

Governance tells you who’s accountable for AI. It doesn’t tell you whether the decisions got better, who pays when they’re wrong, or whether the savings survive the full cost.

Management can produce evidence on all three. Sample the files nobody challenged. Write a loss-allocation schedule for each material deployment. Rebuild the business case net of review, rework and the fallback.

Run the file review with counsel. If it finds errors, settle what they mean for notice to your insurers and for your application answers before the renewal goes out.

Then put that file in front of your underwriters. When an AI exclusion shows up in a renewal quote, it’s what you negotiate with. Read how the exclusion defines AI, whether it uses the broad “arising out of” wording, and which carve-backs survive, such as Side A and securities claims in D&O or professional services in E&O.

Three for your board on Monday

1. Find the errors nobody challenges, then settle notice. Where would AI errors surface in our claims and underwriting files, and do we have those numbers before the next spend? If the review finds errors, has counsel told us what they mean for notice to our insurers and our application answers before we sign?
2. Map every AI loss to a clause. For each material AI deployment, which policy section or contract clause responds, and which losses have we chosen to keep?
3. Rebuild the business case net. What’s the benefit after review, rework and integration costs, and what happens to the operation if the tool is down for a week?

If the honest answer to any of them is “nobody’s looked,” that’s where to start.

---

ICYMI: Last week we mapped Cybersecurity Awareness Month to your crime, cyber and D&O policies, including where the first AI exclusions are showing up. Read it here.

---

Thank you for reading today’s edition.

Want to share it? Copy the link below:

https://lionspecialty.kit.com/posts/90-of-appealed-ai-denials-were-reversed-plaintiffs-say-41-insurance-groups-filed-to-adopt-ai-exclusions-and-a-375k-ai-saving-that-nets-40k

And if this was forwarded to you, subscribe here.

Stay Covered Everybody,

-FLIP

P.S. If you’re briefing your own management team on AI risk, the three essays are written to be forwarded. And if you’re not sure which AI losses your policies would pick up, and which your contracts quietly leave with you, I can help you check. Reach out for a confidential conversation: book a time or reply to this email.

P.P.S. Nothing in this briefing constitutes legal advice. These are the opinions of the founder, and coverage is always subject to the terms of each policy. It’s market intelligence designed to help you ask better questions of your advisors and make sharper decisions at your next insurance renewal.

---

You're receiving this because you subscribed to the LION Specialty Boardroom Brief.
Unsubscribe  ·  Update your preferences