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 under existing standards: 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 (Academy of Actuaries, October 2024). No ASOP or ABCD guidance yet addresses AI-drafted documentation specifically.

From tracking AI adoption across carrier actuarial functions over the past 18 months, a consistent sequence has emerged: AI enters the analytics layer first, pricing models, reserve triangles, underwriting scores, then documentation follows within a year or two, and governance frameworks arrive last, usually only after a complaint, an audit finding, or a regulatory question forces the issue. Reserving and pricing AI has now had three-plus years of sustained regulatory attention: the NAIC's Model Bulletin on insurer AI use dates to December 2023 and has been adopted in 24-plus states (Quarles Law Firm, 2026), and the NAIC's AI Systems Evaluation Tool has been piloting governance examinations across 12 states since early 2026. Documentation has had almost none. The written opinion an appointed actuary signs, certifying that reserves make good and sufficient provision for unpaid claim liabilities, sits entirely outside that apparatus, even as practitioners increasingly use large language models to draft the memorandum that supports it.

Why the Documentation Layer Was Never Built to Be Governed

Every existing AI governance framework in insurance, the NAIC Model Bulletin, ASOP No. 56 on modeling, the Colorado AI Act, the EU AI Act's high-risk classification, was written with a specific mental model of what "the AI system" is: a scoring engine, a pricing algorithm, a claims triage classifier. Each of those frameworks asks the same set of questions, roughly: what does the model predict, what data trained it, how was it validated, and what governance sits around its deployment. None of them was written with the assumption that 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.

That gap matters because documentation is not incidental to actuarial work; it is frequently the work product a regulator, auditor, or opposing expert actually reviews. A Property & Casualty Statement of Actuarial Opinion and its supporting Actuarial Opinion Summary, filed under the NAIC's Actuarial Opinion and Memorandum Regulation, exist specifically so a state insurance department can assess reserve adequacy without rebuilding the analysis from scratch. If the prose explaining methodology, assumptions, and judgment calls was substantially generated by an LLM and only lightly reviewed, the artifact a regulator relies on carries a provenance question that the AOMR's disclosure framework was never designed to surface.

Three Tiers of AI Involvement, Three Different Risk Profiles

Not all AI-assisted drafting carries the same exposure, and conflating the tiers is where most informal discussion of this issue goes wrong. It is useful to separate three distinct levels of AI involvement in an actuarial document, because professional risk scales sharply between them.

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 where the governance blind spot actually bites. A large language model asked to "explain why a 1.08 development factor was selected" will produce fluent, plausible-sounding actuarial reasoning regardless of whether that reasoning matches what the actuary actually did, because the model is optimizing for coherent text, not for fidelity to a specific analyst's judgment process. The SOA Research Institute's January 2025 guide, "Operationalizing LLMs," by Caesar Balona, frames this as a core risk category for actuarial GenAI use: outputs that are fluent but not necessarily grounded in the underlying data or the actuary's own analytical path. When that fluent-but-ungrounded text becomes the rationale paragraph in a signed opinion, the actuary has certified a chain of reasoning that may not be the one that was actually followed.

What ASOP No. 41 Already Requires, and Where It Runs Out

The one standard that speaks most directly to this problem, without ever anticipating it, is ASOP No. 41, Actuarial Communications. The standard requires that actuarial reports identify individual actuaries by name and credential as authors, not merely the name of a firm or organization, a requirement Pinnacle Actuarial Resources has described as ensuring "another actuary qualified in the same practice area could make an objective appraisal" of the work (Pinnacle Actuarial Resources, December 2023). That single sentence does most of the load-bearing work here: peer review under ASOP No. 41 presumes the artifact under review reflects one named actuary's judgment, traceable and defensible, not a probabilistic text-completion process layered underneath a human signature.

The Actuarial Standards Board is mid-revision on ASOP No. 41 itself, with a second exposure draft that sharpens the distinctions among "actuarial communications," "actuarial reports," and "actuarial documentation" and adds a positive disclosure requirement when an actuary uses an assumption or method it did not itself select. That revision was not drafted with LLM drafting in mind, but its direction, tightening what must be disclosed about the origin of an assumption or method, points toward the natural extension: a method or rationale an actuary did not itself derive, whether from a colleague, a vendor model, or a language model, arguably falls under the same disclosure logic already being written into the standard.

The Academy's October 2024 professionalism paper on generative AI is the closest thing to authoritative guidance that exists today, and it states plainly that ASOP Nos. 56 (Modeling), 23 (Data Quality), and 41 (Communications) apply directly to GenAI use even though all three predate widely available generative AI. The paper's central instruction, that an actuary cannot use a GenAI result without independent validation and simply attribute an error to "what the model produced," extends logically from model outputs to drafted text. But logical extension is not the same as explicit guidance, and the ABCD confirmed in its own 2025 activity that actuaries are actively asking: the board's 2025 Annual Summary reports receiving Requests for Guidance during the year specifically concerning the use of Copilot and other AI tools in actuarial calculations, the first documented signal that the profession's own disciplinary body is fielding real-world questions this specific.

Accountability Under ABCD: The Signature, Not the Tool, Is What Gets Disciplined

The ABCD's jurisdiction is instructive precisely because it is tool-agnostic by design. The board evaluates whether a member's conduct met the Code of Professional Conduct and applicable ASOPs; it has never distinguished, and has no structural reason to start distinguishing, between an error that originated in a spreadsheet macro, a vendor's pricing model, a junior analyst's draft, or an LLM's output. The actuary who signs the opinion owns the result regardless of its drafting mechanism. That principle is not new and does not require new guidance to apply.

What is genuinely underspecified is the standard of care for verification. When a junior analyst drafts a section for a senior actuary's review, decades of firm practice and training define what "adequate review" looks like, cross-checking figures against source triangles, confirming the narrative matches selected assumptions, testing whether a described method was actually applied. No comparable body of practice yet exists for reviewing LLM-drafted actuarial prose, where the failure mode is different in kind: a human junior analyst who misunderstands an assumption typically produces text that is internally inconsistent or flags its own uncertainty, while an LLM produces text that is fluent and internally consistent regardless of whether it accurately reflects the underlying analysis. A reviewing actuary trained to catch the first failure mode is not automatically equipped to catch the second.

The E&O Question Nobody Has Tested Against a Real Claim

Professional liability coverage compounds the exposure rather than resolving it. The insurance market's own posture toward AI-related claims has hardened sharply through 2026: 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 (Fenwick, 2026; Risk & Insurance, 2026). One legal analysis of the shift put the change bluntly: "the era of 'Silent AI,' where founders hoped their existing E&O policy implicitly covered AI risks due to lack of explicit exclusion, is definitively over" (Insuriam, 2026).

Standard actuarial professional liability policies were underwritten against a specific loss model: human analytical error, missed data, a flawed but explainable judgment call. That model does not map cleanly onto a claim alleging that an actuary's signed opinion relied on an LLM-drafted rationale that misstated the actuary's own methodology, a scenario that is neither a pure modeling error (the reserve number itself may have been correctly calculated) nor a pure communications failure in the traditional sense (the narrative was generated, not merely poorly written). Standalone AI liability products have begun to appear specifically because general E&O carriers are retreating from this exposure: Munich Re and other specialty carriers now offer AI liability coverage with limits reported between $2 million and $50 million, and a claims-made product from the carrier Testudo launched in January 2026 targeting enterprises deploying generative AI (Risk & Insurance, 2026). Whether any of these newer products, or a firm's existing actuarial E&O policy, would respond to a claim alleging AI-assisted opinion drafting is untested, because no such claim has yet been publicly litigated to a coverage determination. Firms that have not asked their broker this specific question are carrying an unpriced gap.

What the July 22 NAIC Panel Can Realistically Address

The NAIC's Big Data and Artificial Intelligence (H) Working Group holds a public actuarial panel on AI governance trends on July 22, 2026, paired with a progress update on the AI Systems Evaluation Tool pilot running across 12 states through September 2026. The working group's own March 2026 Issue Brief on artificial intelligence and state insurance regulation is explicit about scope: existing state insurance laws apply "regardless of whether decisions are made by humans, algorithms, or third-party vendors" (NAIC, March 2026), a framing built for underwriting and pricing decisions, not for the provenance of the prose in a filed opinion.

That mismatch defines the ceiling on what the July 22 session can accomplish. The Evaluation Tool's four exhibits are structured around AI systems that produce decisions or scores, usage inventories, risk assessments, and data lineage for models that classify or predict. None of the four exhibits was built to ask a signing actuary whether the narrative justification in a reserve opinion was independently authored or substantially machine-drafted. A panel discussion can surface the question and put it on the working group's radar; it cannot retrofit an examination instrument built around decisioning models into one that also audits documentation provenance without a separate work stream. The most realistic outcome of the July 22 panel is that it becomes the moment the gap gets named publicly by regulators, which is a necessary precursor to guidance but is not itself guidance.

Where Professional Guidance Is Actually Needed

Three distinct gaps stand out as the most urgent for the SOA, CAS, and Academy to close, in rough order of tractability. First, an explicit disclosure convention: whether an actuarial report should state, at whatever level of specificity the ASB decides is workable, that AI tools were used in drafting substantive sections, mirroring the assumption-and-method disclosure logic already being written into the ASOP No. 41 revision. Second, a peer review standard specific to AI-drafted text, since the existing "objective appraisal" bar in ASOP No. 41 assumes a reviewer is checking a colleague's judgment, not auditing whether generated prose accurately represents judgment that happened somewhere else. A reviewer verifying AI-assisted work needs to understand the specific failure mode of the tool used, not just the actuarial conclusion being reviewed, which is a competency current CPD requirements do not yet name. Third, explicit ABCD guidance on the standard of care for verifying AI-generated substance versus AI-generated structure, since the current guidance treats all AI use as a single undifferentiated category subject to "independent validation" without specifying what validation looks like when the artifact being validated is prose rather than a number.

None of these gaps requires a new ASOP. Each is closable through interpretive guidance layered onto ASOP Nos. 41, 56, and 23, the same three standards the Academy's 2024 paper already identified as applicable, plus targeted ABCD guidance responding to the Copilot-related requests the board has already logged. The CAS's 2026 request for proposals on adapting large language models for specialized P&C actuarial reasoning suggests the profession is investing in making LLMs more reliable for actuarial tasks; it has not yet produced parallel investment in the professional standards for reviewing what those tools produce once they are reliable enough to be trusted with a first draft.

Why This Matters for Signing Actuaries Now

An actuary does not need to wait for ASOP guidance or ABCD clarification to reduce exposure today. Documenting which sections of a reserve memorandum or opinion involved AI drafting, at what tier, and what independent verification was performed creates exactly the kind of contemporaneous record that protects an actuary if a certified number is later contested, mirroring the workpaper discipline already expected under ASOP No. 41's disclosure requirements. Confirming with a broker, in writing, whether current E&O coverage responds to a claim involving AI-assisted drafting closes an unpriced gap before a claim forces the question. And treating AI-drafted rationale paragraphs with the same skepticism applied to a vendor model's black-box output, verifying the described method against what was actually done, not just checking that the prose reads plausibly, addresses the specific failure mode LLMs introduce that human junior-analyst drafts do not.

The profession moved deliberately on AI governance for pricing and reserving models because those models produce numbers that are relatively easy to audit against ground truth. Documentation is harder to audit precisely because its output is prose, and fluent prose is much better than a wrong number at hiding the fact that no one independently verified it.