Lloyd’s-backed AI capacity more than doubled to $25M as twenty-three states made governance documentation an underwriting precondition.
The Thesis
The U.S. liability insurance market repriced AI as a distinct risk class in early 2026, and access to coverage now turns on documented governance evidence the underwriter can read, score, and price.
The Signal
Three moves repriced the AI insurance market this quarter.
What happened. Verisk’s Insurance Services Office released three optional endorsements effective January 1, 2026: CG 40 47, CG 40 48, and CG 35 08. CG 40 47 excludes any claim arising out of generative AI across both bodily injury and personal/advertising injury coverage. ISO forms underpin roughly 82% of U.S. P&C policies, so these are the standard language any carrier can attach at any 2026 renewal. W.R. Berkley filed Form PC 51380, an “absolute AI exclusion” for D&O, E&O, and fiduciary liability barring coverage “based upon, arising out of, or attributable to” AI. AIG and Great American sought clearance for similar language. Philadelphia Insurance and Hamilton Select have already excluded AI-related claims from E&O coverage, with Hamilton’s endorsement specifically naming ChatGPT, Bard, Midjourney, and DALL-E.
Why it matters. Until January, AI exposure was covered by default. The policy never mentioned AI, and when something went wrong the carrier paid. That silent-coverage era ended in a similar way to how silent cyber ended at Lloyd’s in 2019. A single field on a renewal form now eliminates AI-related coverage entirely, with trigger language broad enough that “arising out of” applies even when AI is one input among many. The April 2026 securities class action Dunn v. Upstart Holdings is a preview of what the coverage gaps look like in real litigation: investors allege Upstart’s executives misled them about calibration failures of the company’s “Model 22” AI underwriting system. Any company whose D&O carrier has attached the absolute AI exclusion is now exposed to that defense cost without coverage.
Second-order effect. An undocumented AI deployment in 2026 produces a denied claim at the moment of loss. Companies relying on traditional policy language are running uninsured the moment their carrier attaches the endorsement.
What happened. A February 2026 joint report from Lockton Re, Lockton International, and Armilla AI argued for AI as its own insurable class, separate from cyber, E&O, and D&O, on the grounds that AI failure modes (hallucination, model drift, agentic chain-of-action errors) do not map onto existing wordings. The same month, Armilla expanded its standalone AI Liability Policy to $25 million per organization, up from prior caps under $10 million, underwritten by Lloyd’s syndicates. Armilla remains the only Lloyd’s coverholder dedicated specifically to AI liability at this scale. Testudo, also Lloyd’s-backed, launched a parallel product covering hallucination, IP infringement, and regulatory penalties. No major admitted-market carrier has stepped in to fill the gap created by the ISO exclusions.
Why it matters. Specialty capacity is scaling at exactly the moment the primary market is excluding the risk, which is the structural fingerprint of a new insurable class forming. Affirmative AI coverage is becoming a standalone procurement decision with its own diligence cycle, the same path cyber took between 2014 and 2018. The absence of admitted-market competition tells you the pricing is not yet attractive enough for incumbents to compete, which is exactly when buyers can negotiate the cleanest terms.
Second-order effect. Brokers face a fiduciary question they did not face six months ago: when a carrier attaches CG 40 47 and the insured’s AI footprint is uncovered, the broker’s failure to recommend a Lloyd’s-backed standalone product becomes a documentable negligence exposure.
What happened. The NAIC Model Bulletin on AI Systems has been adopted by 23 states plus DC as of April 1, 2026. It requires insurers to maintain a written AI Systems Program covering four pillars: governance accountability with named owners, risk management proportional to consumer harm, third-party vendor oversight that does not transfer responsibility, and lifecycle documentation retrievable on demand. In January 2026, the NAIC launched its AI Systems Evaluation Tool pilot across twelve states, giving market conduct examiners a structured framework for reviewing AI governance programs in real exams running through September 2026.
Why it matters. The same governance posture insurers must now document for their own regulators is the posture they are demanding from commercial customers seeking AI coverage. A buyer’s underwriting submission needs the four pillars in writing — model inventory mapped to risk tier, vendor diligence file for each foundation model provider, monitoring telemetry, and an incident response runbook with named escalation paths. Companies that produce this material on ten business days’ notice access affirmative coverage; companies that cannot are pushed into the exclusion column.
Second-order effect. Documented AI governance moves from a compliance artifact to a precondition for procurement, financing, and M&A diligence — a roughly twelve-month compression of the shift cybersecurity governance underwent between 2018 and 2022.
The Playbook
Six questions to ask before the next renewal package or board meeting.
A single line in W.R. Berkley’s absolute AI exclusion now appears in 2026 D&O, E&O, and fiduciary renewals: coverage prohibits any claim “based upon, arising out of, or attributable to” any use, deployment, or development of AI. Read aloud in a risk meeting and the implication lands fast: every AI deployment under that policy line is now uncovered.
Before the renewal call, have the broker search every renewing CGL, D&O, E&O, EPLI, and cyber policy for any carrier-specific AI exclusion language, including ISO forms CG 40 47, CG 40 48, CG 35 08 and the W.R. Berkley absolute AI exclusion. Request the verbatim policy text in writing. The broker summary will likely use softer language than the policy.
For every AI use case in production — including agents inside Microsoft Copilot, Google Gemini Enterprise, and Salesforce Agentforce that arrived as features of an existing subscription — identify which policy line responds today, which would after the new endorsement, and which is uncovered.
Carriers writing affirmative AI coverage now request a documented governance pack: model inventory with risk tier, vendor diligence files, monitoring telemetry, incident response runbook, and board AI risk reporting cadence. Companies with this material access coverage; companies without it pay the exclusion.
Get terms from at least two Lloyd’s coverholders for affirmative AI liability coverage. Armilla is the dominant scaled provider; Testudo and Relm offer narrower products. No admitted-market carrier yet competes directly. Surplus lines placement means longer underwriting timelines and higher governance disclosure than admitted-market products.
Foundation model providers cap their liability at the contract value; enterprise exposure has no ceiling. The vendor contract clause and the insurance policy together determine the residual exposure landing on the balance sheet, and 2026 is the year that exposure becomes a board reporting line.
The Verification Test
Claim. My company’s AI deployments are not insured under our current policy.
Test. Have the broker pull the endorsement schedule on every primary and excess liability policy and search for “CG 40 47,” “CG 40 48,” “CG 35 08,” “artificial intelligence,” and “generative.” For any match, request the full endorsement language and a written confirmation of which AI use cases are excluded.
Pass criteria. AI exclusion attached to a policy line where AI is in production, with no affirmative coverage placed elsewhere; or the policy was renewed in Q1 2026 with no AI-specific question on the application. Either signals material uninsured AI exposure today.
Fail smell. The broker confirms no AI exclusions are attached and no AI questions appear on the application, yet cannot produce a written carrier confirmation of which AI use cases are covered. The absence of an exclusion does not equal the presence of coverage. If the carrier has not affirmatively spoken to AI, the silence will be litigated against the policyholder.
The Metric
What it measures. The maximum a single company can insure against AI-related claims through Armilla, the leading Lloyd’s-backed AI-only insurer — more than doubled from prior caps under $10 million before January 2026.
Why it matters now. When specialty insurers raise their limits, it means they’re confident enough in their ability to price the risk to put more money behind it. That’s happening at the exact moment traditional carriers are pulling back. Two sides of the market moving in opposite directions is how a new insurance category is born. Cyber went through the same split and it took five years to mature. AI is moving through it in roughly twelve months.
Source. Armilla AI / FF News, January 22, 2026.
Your liability policy used to be the seatbelt. The carrier just unbolted it from the floor, and most riders won’t notice until the crash, the moment a claim is filed.
The Lens — Horizon Search Institute
NIST AI RMF implementation is being cited in carrier diligence as evidence of organizational control, the first time a voluntary framework has flowed directly into liability underwriting. Lexology
Underwriters reviewing AI risk now ask whether human reviewers retain decision authority over AI-generated outputs, treating that authority as a measurable risk control. Lockton Re
Twenty-three states plus DC have adopted the NAIC AI Model Bulletin, and the NAIC’s January 2026 Evaluation Tool pilot is producing examiner-ready governance scorecards in twelve of those states. NAIC
Links Worth Your Time
-
The Evolving Contours of Artificial Intelligence as a D&O Exposure (Hunton)
The Dunn v. Upstart filing is the first major D&O case where an AI underwriting model’s calibration failure is the alleged misrepresentation. Read this before any audit committee briefing on AI.
-
NAIC Intensifies AI Regulatory Focus (Crowell & Moring)
Walks through the March 2026 NAIC Issue Brief, the twelve-state AI Systems Evaluation Tool pilot, and the federal preemption question raised by Executive Order 14365. The clearest map of how the regulatory architecture is being built right now.
-
Business Insurance — Insurers, brokers adjust as AI exclusions emerge
Original reporting on which carriers are actually attaching CG 40 47, with named broker sources at Gallagher, Axa XL, Coalition, Zurich, and Westfield Specialty. The piece to send to the broker before the renewal call.
-
NAIC AI Model Bulletin (PDF)
The actual regulatory text. The state-by-state adoption pattern is the structural fact behind the underwriting shift.
-
Carrier Management — AI Needs Its Own Risk Class: Lockton Re
The Lockton Re / Armilla joint report condensed; reads as the industry’s first formal argument for treating AI as a separate insurable line.
- Geoffrey B. Fehling, Michael S. Levine, and Yaniel Abreu, “The Evolving Contours of Artificial Intelligence as a D&O Exposure,” Hunton Andrews Kurth Insurance Recovery Blog, May 5, 2026. hunton.com
- Vanessa A. Perumal et al., “NAIC Intensifies AI Regulatory Focus: What Health Insurance Payors Need to Know,” Crowell & Moring LLP, March 25, 2026. crowell.com
- Dunn v. Upstart Holdings, Inc., complaint filed April 7, 2026, U.S. District Court for the Northern District of California (securities class action alleging misrepresentations about AI underwriting Model 22).
- Matthew Lerner, “Insurers, brokers adjust as AI exclusions emerge,” Business Insurance, April 14, 2026. businessinsurance.com
- Insurance Services Office, Form CG 40 47 01 26 — Exclusion: Generative Artificial Intelligence (Commercial General Liability Endorsement, effective January 1, 2026), referenced through state regulatory filings and trade press summaries.
- W.R. Berkley Corporation, Form PC 51380 — Artificial Intelligence Absolute Exclusion (D&O, E&O, Fiduciary Liability, 2025–2026 filings), quoted in Harvard Law School Forum on Corporate Governance, “The Hidden C-Suite Risk of AI Failures,” September 22, 2025.
- Carrier Management, “AI Needs Its Own Risk Class: Lockton Re,” February 17, 2026. carriermanagement.com
- FF News, “Armilla AI Raises Lloyd’s-Backed Coverage to $25M as Traditional Insurers Retreat from AI Risk,” January 22, 2026. ffnews.com
- National Association of Insurance Commissioners, Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, December 2023 (state adoption tracked through April 2026). content.naic.org
- Risk Specialty Group, “AI Liability Insurance for Architects: 2026 Coverage Gaps and Risk Management” (citing Verisk, Hunton Andrews Kurth, and National Law Review reporting on Philadelphia Insurance and Hamilton Select E&O exclusions), January 2026. riskspecialtygroup.com