How Actuaries Price AI Liability Coverage When the Loss Triangle Has No Rows

AI liability insurers are pricing a coverage class with almost no loss triangle: the classical full credibility standard calls for roughly 1,082 claims, while generative AI litigation grew 978% between 2021 and 2025 (Gallagher Re, March 2026) from a base of essentially zero insured claims. Actuaries are substituting analogical transfer and catastrophe-style scenario loading for the missing experience.

That substitution is now showing up in filed rating plans rather than academic papers. Seven of 13 lawyers’ professional liability carriers reported AI-related claims increases in EPIC Insurance Brokers’ 16th Annual LPL Claims Survey (EPIC, May 2026), the closest thing the market has to loss emergence data for any AI-adjacent liability line, and it covers legal malpractice, not the standalone AI liability products carriers such as Corgi have only started writing this year. When a regulator asks what supports the rate on a new AI liability form, the honest answer from most filing actuaries in 2026 is judgment layered on adjacent-line data, not a credible triangle drawn from the coverage itself.

The Credibility Math Doesn’t Clear

Classical credibility theory sets a numerical bar before an actuary can rely on a book’s own experience instead of a broader class average. The limited fluctuation standard for full credibility on claim frequency, derived from (1.645/0.05)², works out to approximately 1,082 claims, the volume needed for observed frequency to fall within 5% of the true mean at a 90% confidence level (Dean, Casualty Actuarial Society). Buhlmann-Straub credibility relaxes the all-or-nothing cutoff into a partial-credibility weight, but it still requires enough exposure years within the class, and a stable enough variance structure across risks, to estimate both a credible mean and a between-risk variance parameter. AI liability, as a coverage distinct from the tech E&O and product liability forms it typically rides alongside, does not clear that bar. Corgi’s AI and Algorithmic Liability Endorsement, a modular add-on to existing technology errors and omissions policies rather than a standalone form, only launched in May 2026 (PR Newswire, May 4, 2026). A book that young cannot generate 1,082 claims in any single accident year, and likely will not for several years even as premium volume grows.

The nearest available proxy, EPIC’s LPL survey, is itself a data point about an adjacent line, not the class being priced. Fifty-four percent of the 13 surveyed carriers, which collectively insure more than 80% of Am Law 200 firms, reported an increase in AI-related claims over the prior year (EPIC Insurance Brokers, May 2026). That is directional evidence that AI is generating professional liability losses somewhere in the system. It is not a frequency estimate an actuary can plug into a Buhlmann-Straub formula for AI liability rating, because the exposure base, coverage trigger, and claimant population of a legal malpractice policy differ from a standalone AI liability policy written for an AI vendor or an enterprise deploying a third-party model.

Constructing a Prior From Adjacent Lines

With classical credibility off the table, the working method looks closer to Bayesian prior construction than ratemaking in the traditional sense. An actuary starts from loss experience in adjacent lines, weights each by how closely its harm mechanism resembles what an AI system actually does, and blends the results into a starting frequency and severity assumption for the new class. Technology errors and omissions supplies the financial-harm analogue: professional service failures and software defects that cause a client’s business loss without physical injury. Product liability supplies the bodily-injury analogue for AI embedded in physical systems, autonomous vehicles, robotic surgical devices, and industrial control software, where the loss process resembles a defective-product claim more than a professional-negligence claim. Cyber liability supplies the data-processing and systemic-accumulation analogue: breach costs, business interruption, and the correlated-loss structure that shows up when many policyholders depend on the same underlying model, cloud provider, or training dataset.

The claim category data in Gallagher Re’s March 2026 report on the AI insurance gap, produced with MIT and Testudo Global and titled “Smart Systems, Blind Spots: Rethinking Insurance for the AI Era,” gives a rough guide to how those weights should split in practice. Of the more than 700 cumulative generative AI lawsuits filed in the U.S. between 2020 and 2025, patent infringement accounted for 11.9% of claim categories, copyright infringement 11.2%, and personal injury tied to privacy violations 10.2% (Gallagher Re, MIT, and Testudo Global, March 2026). That distribution argues for weighting the tech E&O and intellectual-property analogues more heavily than the pure bodily-injury analogue in a typical AI liability book today, at least until the mix of insureds shifts toward physical-system deployments such as autonomous vehicles or connected medical devices, where the product liability analogue should dominate the blend.

The cyber analogue carries a warning the other two do not: correlated failure across many insureds at once. “When you consider how much organizations are relying on AI platforms to provide critical services and products to their own clients, it creates the potential for the frequency and severity of claims to go up,” said John Farley, managing director of Gallagher’s Cyber Liability practice (Gallagher, “Not So Silent: Tackling the Complexities of AI Liability,” May 2026). A single foundation model, cloud provider, or training-data vendor sitting behind dozens of insureds is the same aggregation problem that reshaped cyber catastrophe modeling after prior large-scale outage and breach events. It argues for an explicit accumulation load layered on top of the per-risk analogical blend, not just a per-policy severity assumption borrowed independently from each adjacent line.

Catastrophe-Style Scenario Loading Instead of a Trend Line

Where a mature line extrapolates severity trend from several accident years of closed claims, AI liability actuaries are borrowing the scenario-based approach catastrophe modelers and pandemic-mortality actuaries use when there is no credible trend to extrapolate: define a small number of plausible loss scenarios, assign each a probability, and load the rate to the probability-weighted outcome rather than to a fitted curve. The scenario severity anchors increasingly come from early case law rather than closed claims, because closed claims with awarded damages barely exist yet in this class.

Three developments from the past 14 months illustrate the anchor points currently available. On May 21, 2025, a federal district judge in Florida ruled that Character.AI’s chatbot product is subject to product liability law on the same footing as a defective vehicle, allowing strict liability, negligence, and wrongful-death counts to proceed, the first ruling to establish that framework for a generative AI product. In January 2026, Character.AI and Google settled five related lawsuits over teen suicide and self-harm claims, among the first AI-chatbot-harm settlements of scale in the country, though the terms were confidential and included no admission of liability. And on February 3, 2026, a California Superior Court coordinated roughly a dozen wrongful-death and product-liability cases against OpenAI into a single proceeding, In re: ChatGPT Product Liability Cases, JCCP No. 5431, none of which had reached trial as of mid-2026.

None of those three data points gives an actuary a jury-awarded severity figure to anchor a scenario. The 2025 product-liability ruling establishes legal exposure, not a dollar amount. The Character.AI settlement establishes that claims of this type are worth resolving before trial, but confidential terms mean the number cannot be used directly in a rate indication. The JCCP coordination signals claim volume and venue concentration, both of which raise the probability of an eventual verdict but say nothing yet about its size. What actuaries can do with this record is build a severity distribution with wide bands, anchored at the low end to product liability and professional liability verdict ranges for comparable harm categories, and stress-tested at the high end against the accumulation scenario the cyber analogue implies. That is a materially wider confidence interval than a pricing actuary would tolerate in a mature line, and it is the honest output of the available information.

Rating Factors Showing Up in 2026 Filings

Even without a credible loss triangle, carriers writing AI liability in 2026 have converged on a similar set of rating factors, each intended to proxy for frequency or severity risk that cannot yet be measured directly from the class’s own experience.

Rating Factor	What It Proxies For
Model type (generative vs. deterministic)	Generative outputs carry open-ended hallucination and IP exposure; rules-based systems carry narrower, more predictable failure modes
Industry of deployment	Healthcare AI implicates bodily-injury and regulatory exposure; consumer recommendation engines skew toward financial and reputational harm; autonomous vehicle and industrial control skew toward product-liability severity
Human-in-the-loop presence	A human reviewer before consequential action reduces frequency and shifts liability allocation toward the deployer rather than the model provider
Training data provenance	Documented, licensed data sources reduce IP-infringement frequency; scraped or undocumented data raises both frequency and defense cost
Audit certification status	Third-party model audits and bias testing function as a proxy for governance quality, similar to how cyber underwriters use security control questionnaires

None of these factors has been validated against AI liability’s own loss experience, because that experience does not yet exist in credible volume. Each is instead a hypothesis borrowed from the underwriting logic of an adjacent line, cyber’s control questionnaires most directly, and treated as a rating variable until enough claims accrue to test whether it actually correlates with loss. The human-in-the-loop and audit-certification factors are the two most likely to move rates materially at renewal, since they are also the two a policyholder can change through its own governance practices rather than through the underlying model choice, giving carriers a lever for risk selection even before frequency data exists to price the factor precisely.

What a Judgment-Heavy Rate Filing Has to Document

Standard ratemaking documentation walks a state insurance department through loss development triangles, trend selections, and an indicated rate change tied to the company’s own experience or an approved rating bureau’s data. An AI liability filing built on analogical transfer and scenario loading cannot produce that package, because the triangles and trend lines it would rest on do not exist. What it can produce instead is a methodology memorandum: the adjacent lines selected, the weights assigned to each and the reasoning behind them, the scenario set used for severity, the probability weights assigned to each scenario, and a sensitivity analysis showing how the indicated rate moves if those weights or probabilities shift within a plausible range.

That is a materially different conversation with a regulator than a typical new-line filing. Reviewers accustomed to checking an actuary’s trend selection against several years of closed claims are instead being asked to evaluate whether the analogical weights and scenario probabilities are reasonable, a judgment call closer to reviewing a catastrophe model’s assumption set than reviewing a standard indication. This runs alongside, and is distinct from, the separate NAIC track governing how carriers use AI in their own underwriting and claims systems; 23 states and Washington, D.C. had adopted the NAIC Model Bulletin on insurer AI governance as of early 2026, with a 12-state AI Systems Evaluation Tool pilot running January through September 2026 (NAIC, March 2026). That workstream examines how carriers deploy AI internally. The rate-filing question addressed here is different: what actuarial support underlies the premium charged for insuring someone else’s AI risk. State DOIs reviewing AI liability filings in 2026 are increasingly asking pricing actuaries to show the sensitivity of the rate to each judgment input explicitly, rather than accept a single point estimate, because a point estimate built on this little data invites more scrutiny than the same estimate built on a decade of triangles.

Weighting the First Wave of Court Decisions

From evaluating pricing submissions for new specialty technology lines across several early-stage liability classes, the pattern of judicial outcome clustering in the first wave of case law has consistently been a more reliable signal for initial severity assumptions than the volume of premium written in years one through three of a new class. Premium volume in a nascent line mostly reflects how aggressively carriers are chasing growth, not how the underlying risk is developing. Case law clustering, where multiple similar claims land in the same venue or get coordinated into a single proceeding such as JCCP 5431, is a better early indicator of where severity is likely to land, because it shows plaintiffs’ counsel converging on a theory of liability that a court has allowed to survive a motion to dismiss.

The practical implication is a weighting scheme that looks different from how a mature line updates its indications. In a mature line with 1,082 or more claims, a single new verdict barely moves the indicated rate; it is one more observation in a large sample. In AI liability, with claim counts orders of magnitude below that threshold, a single high-severity verdict can and should move the severity anchor materially, because there is no larger sample to absorb it. Actuaries pricing this class in 2026 and 2027 need a governance process for updating the analogical weights and scenario probabilities every time a coordinated proceeding like JCCP 5431 produces a ruling on the merits, rather than waiting for an annual rate review cycle built for lines where quarterly movement in a handful of cases would be noise.

Why This Matters for Actuaries

The pricing problem does not stay contained to the rate filing. Reserving actuaries setting IBNR for AI liability books face the same thin-data constraint from the other direction: with no development pattern to select a loss development factor from, initial reserves have to be set from the same analogical and scenario framework used to price the business, then updated on the same event-driven cadence rather than a standard quarterly triangle review. Appointed actuaries certifying reserve adequacy for carriers writing this line will need to document why a judgment-based reserve is reasonable in the absence of development data, a harder case to make to an auditor or regulator than pointing to a stable triangle.

Capital allocation carries a parallel issue. A line priced substantially on scenario analysis rather than credible experience should carry a wider confidence interval around its indicated capital requirement, and risk officers modeling aggregate exposure across an AI liability book need to account for the correlated-failure scenario the cyber analogue implies, not just a sum of independent per-policy severities. Gallagher Re projects the standalone AI liability insurance market could reach $4.8 billion by 2032, up from a category that barely existed as a distinct line before 2025 (Gallagher Re, 2026). For consulting actuaries advising smaller carriers or MGAs entering this space without the data advantage of an early mover like Corgi, the analogical transfer framework described here is the realistic starting point, not a stopgap to be replaced once real data appears. Real data, at credible volume, is likely several years away. The first wave of court decisions from JCCP 5431 and its successors will recalibrate the severity anchor well before claim counts approach anything close to 1,082.