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 from a base of essentially zero insured claims.

What has replaced the missing experience is analogical transfer from adjacent lines plus catastrophe-style scenario loading, and it is showing up in filed rating plans rather than in academic papers.

Key Takeaways

• 1,082 claims is the limited fluctuation standard for full credibility on frequency, derived from (1.645/0.05) squared, and no AI liability book will approach it for several years.
• Seven of 13 lawyers' professional liability carriers reported AI-related claims increases on a panel insuring more than 80% of Am Law 200 firms, the closest thing to loss emergence data in any AI-adjacent line.
• Patent infringement at 11.9%, copyright at 11.2% and privacy-linked personal injury at 10.2% of claim categories across 700-plus US generative AI lawsuits argue for weighting technology E&O over bodily injury.
• Corgi's AI and Algorithmic Liability Endorsement launched in May 2026 as a modular add-on to technology errors and omissions rather than a standalone form.
• JCCP No. 5431 coordinated roughly a dozen wrongful-death and product liability cases against OpenAI in February 2026, none of them tried yet.

The Credibility Math Does Not Clear

The limited fluctuation standard for full credibility on claim frequency works out to roughly 1,082 claims, the volume needed for observed frequency to sit within 5% of the true mean at 90% confidence. Buhlmann-Straub softens that cutoff into a partial weight but still needs enough exposure years, and a stable enough variance structure, to estimate a credible mean and a between-risk variance.

AI liability, as a coverage distinct from the technology E&O and product liability forms it rides alongside, clears neither. Corgi's endorsement launched in May 2026 as an add-on to existing tech E&O policies. A book that young cannot produce 1,082 claims in an accident year and will not for some time even as premium grows.

The nearest proxy is a different line. EPIC's 16th annual survey found seven of 13 carriers, on a panel covering more than 80% of Am Law 200 firms, reporting an increase in AI-related claims. That is directional evidence AI is generating professional liability losses somewhere. It is not a frequency estimate for AI liability: the exposure base, coverage trigger and claimant population of a legal malpractice policy are not those of a policy written for an AI vendor or an enterprise deploying a third-party model.

Building a Prior From Adjacent Lines and Court Records

With classical credibility unavailable, the working method resembles Bayesian prior construction. Technology E&O supplies the financial-harm analogue. Product liability supplies the bodily-injury analogue for AI inside physical systems. Cyber supplies the accumulation analogue.

The claim mix suggests the weights. Of more than 700 cumulative US generative AI lawsuits filed between 2020 and 2025, patent infringement accounted for 11.9% of claim categories, copyright 11.2% and privacy-linked personal injury 10.2%. That argues for weighting tech E&O and the intellectual property analogues above pure bodily injury, at least until the insured mix shifts toward autonomous vehicles and connected medical devices.

The cyber analogue carries a warning the others do not. "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 of Gallagher's cyber liability practice. One foundation model behind dozens of insureds is the aggregation problem that reshaped cyber catastrophe modeling, and it argues for an explicit accumulation load on top of the per-risk blend.

Severity anchors are coming from case law because closed claims with awarded damages barely exist. A Florida federal judge ruled on May 21, 2025 that Character.AI's chatbot is subject to product liability law on the same footing as a defective vehicle. Character.AI and Google settled five related suits in January 2026, and a California Superior Court coordinated roughly a dozen cases against OpenAI into JCCP No. 5431 that February.

None yields a dollar figure. The ruling establishes exposure, the settlement terms are confidential, and the coordination signals volume and venue rather than size. What they support is a severity distribution with wide bands, anchored low against product and professional liability verdict ranges and stress-tested high against the accumulation scenario.

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 those factors has been validated against the class's own loss experience. Each is a hypothesis borrowed from adjacent underwriting logic, cyber's control questionnaires most directly, and treated as a rating variable until claims accrue to test it.

One Verdict Moves the Anchor

That thinness changes how an indication should update. In a line with 1,082 claims or more, a new verdict is one observation in a large sample and barely moves the rate. In AI liability, with counts orders of magnitude below that, a single high-severity verdict should move the severity anchor materially, because no larger sample absorbs it.

So the governance cadence has to be event-driven rather than annual. A ruling on the merits out of a coordinated proceeding like JCCP 5431 is a repricing event, and an annual review cycle designed for lines where a handful of cases is noise will miss it by months.

The filing conversation changes with it. Standard documentation walks a department through triangles, trend selections and an indication tied to the company's own experience. This one cannot, so it produces a methodology memorandum instead: the adjacent lines chosen, the weights and the reasoning, the scenario set and its probabilities, and a sensitivity analysis showing how the indication moves as those inputs shift. Reviewing that is closer to reviewing a catastrophe model's assumptions than a conventional indication.

That is separate from the NAIC track governing carriers' own use of AI, adopted in 23 states and Washington, DC. That governs how a carrier deploys AI internally; the question here is what supports the premium charged for insuring someone else's.

The constraint reaches reserving from the other side too. With no development pattern, initial reserves come from the same framework used to price the business and update on the same cadence, which is a harder case to make to an auditor than a stable triangle. Gallagher Re projects the standalone market could reach $4.8 billion by 2032. Credible volume sits years behind that, and the first rulings out of JCCP 5431 will recalibrate the anchor long before claim counts approach 1,082.