When LLMs Draft the Reserve Opinion: The Actuarial Liability Gap

An actuary who signs a reserve opinion substantially drafted by an LLM remains fully accountable for it. ASOP No. 41 requires an individually named actuary as author, not an organization or a tool, and the American Academy of Actuaries has stated an actuary cannot rely on a generative AI result without independent validation.

No ASOP or ABCD guidance yet addresses AI-drafted documentation specifically, and every framework that does exist was built for a different artifact.

Key Takeaways

• A 1.08 development factor selected over 1.15 is the kind of judgment an LLM will explain fluently regardless of whether the explanation matches what the actuary actually did.
• 24-plus states have adopted the Model Bulletin since December 2023 and the Evaluation Tool has been piloting across 12 states since early 2026. Documentation sits outside all of it.
• Chubb and Travelers have state approval to add explicit AI exclusions to general liability, D&O and errors and omissions forms, targeting losses tied to generative AI outputs.
• $2 million to $50 million are the reported limits on standalone AI liability products now appearing because general E&O carriers are retreating from the exposure.

Documentation Was Never in Scope

Every existing AI governance framework in insurance was written around a specific mental model of the AI system: a scoring engine, a pricing algorithm, a claims triage classifier. Each asks what the model predicts, what data trained it, how it was validated and what governance surrounds deployment. None was written assuming the artifact in question might be a paragraph of prose explaining why the actuary selected a 1.08 loss development factor over a 1.15 alternative, or a section of a Statement of Actuarial Opinion summarizing the basis for a risk margin.

Documentation is frequently the work product a regulator, auditor or opposing expert actually reviews. A P&C Statement of Actuarial Opinion and its supporting summary, filed under the NAIC's Actuarial Opinion and Memorandum Regulation, exist so a state department can assess reserve adequacy without rebuilding the analysis. If the prose explaining methodology and judgment calls was substantially machine-generated and lightly reviewed, the artifact carries a provenance question the AOMR was never designed to surface.

The asymmetry in attention is stark. The Model Bulletin on insurer AI use dates to December 2023 and has been adopted in 24-plus states, and the Evaluation Tool has been piloting governance examinations across 12 states since early 2026. The opinion an appointed actuary signs sits outside all of it.

Tier Three Is Where It Bites

Not all AI-assisted drafting carries the same exposure, and conflating the levels is where informal discussion of this goes wrong.

Tier	What the AI does	Professional risk
AI-assisted editing	Grammar, spell-check, tone, formatting consistency across a report already drafted by the actuary	Minimal; functionally equivalent to a word processor's editing tools
AI-generated structure	Outlining sections, drafting boilerplate disclosure language, standardizing table formats, summarizing prior-year memoranda for continuity	Moderate; requires verification that boilerplate reflects the current engagement's actual facts and assumptions
AI-generated substance	Drafting the rationale for a specific assumption, explaining why a method was selected over an alternative, characterizing a risk margin's adequacy	High; the actuary's judgment, not the AI's, is what ASOP No. 41 and the signed opinion certify

The third tier is the one the frameworks miss. A language model asked to explain why a 1.08 development factor was selected produces fluent, plausible actuarial reasoning whether or not that reasoning matches what the actuary did, because it optimizes for coherent text rather than fidelity to a particular analyst's judgment process. The SOA Research Institute's "Operationalizing LLMs" frames exactly this as a core risk category: outputs fluent but not grounded in the underlying data or the analytical path taken. When that text becomes the rationale paragraph in a signed opinion, the actuary has certified a chain of reasoning that may not be the one followed.

ASOP No. 41 is the standard that speaks to this without anticipating it. Its requirement that reports name individual actuaries and credentials rather than a firm exists, as Pinnacle Actuarial Resources puts it, to ensure "another actuary qualified in the same practice area could make an objective appraisal" of the work. Peer review under that standard presumes the artifact reflects one named actuary's traceable judgment, not a probabilistic text-completion process sitting underneath a human signature.

The Actuarial Standards Board is mid-revision on the standard, with a second exposure draft adding a positive disclosure requirement when an actuary uses an assumption or method it did not itself select. That was not drafted with LLM use in mind, but the direction extends naturally to a rationale an actuary did not derive.

What is underspecified is the standard of care for verification. When a junior analyst drafts a section for senior review, practice defines adequate review: cross-check figures against source triangles, confirm the narrative matches selected assumptions, test whether the described method was applied. No comparable practice exists for LLM-drafted prose, and the failure mode differs in kind. A human who misunderstands an assumption produces text that is internally inconsistent or flags its own uncertainty. A model produces text that is fluent regardless of accuracy.

The profession's own disciplinary body is already fielding the question. The ABCD's 2025 Annual Summary reports Requests for Guidance received during the year concerning the use of Copilot and other AI tools in actuarial calculations. Its jurisdiction is tool-agnostic by design, and it has never distinguished between an error originating in a spreadsheet macro, a vendor model, a junior analyst's draft or a model's output. The signature owns the result.

The Coverage and the Examination Both Assume a Number

Two mechanisms that would normally absorb this exposure are structured around numerical output, and neither reaches prose.

Professional liability is contracting rather than expanding. Carriers including Chubb and Travelers have received state approval to add explicit AI exclusions to general liability, D&O and errors and omissions forms, targeting losses tied to generative AI outputs. One legal analysis put it directly: "the era of 'Silent AI,'" where policyholders assumed existing E&O implicitly covered AI risk absent an explicit exclusion, "is definitively over".

Standard actuarial E&O was underwritten against human analytical error, missed data, a flawed but explainable judgment call. A claim alleging a signed opinion relied on an LLM-drafted rationale misstating the actuary's own methodology fits neither: the reserve number may be correct, and the narrative was generated rather than poorly written. Standalone AI liability products have appeared for that reason, with reported limits between $2 million and $50 million and a claims-made product launched in January 2026. Whether any of them, or an existing actuarial policy, responds is untested, because no such claim has been publicly litigated to a coverage determination.

The examination side has the same shape. The NAIC's July 22, 2026 actuarial panel pairs AI governance trends with a pilot update, and the March 2026 Issue Brief states that existing state insurance laws apply "regardless of whether decisions are made by humans, algorithms, or third-party vendors", a framing built for underwriting and pricing decisions.

The Evaluation Tool's four exhibits follow it: usage inventories, risk assessments, high-risk model detail and data lineage, all structured around systems producing decisions or scores. None was built to ask a signing actuary whether the narrative justification in a reserve opinion was independently authored or substantially machine-drafted. A panel can name the gap and put it on the working group's agenda. Retrofitting an instrument built around decisioning models into one that audits documentation provenance is a separate work stream that does not currently exist.