On March 24, 2026, the Casualty Actuarial Society released The CAS AI Primer: Practical Guidance for Actuaries, a document produced by the CAS Artificial Intelligence Working Group. The stated goal is ambitious: provide actuaries with a starting point for responsible AI adoption, covering everything from model specialization to corporate governance. In a landscape where generative AI is reshaping actuarial workflows faster than most credentialing bodies can keep pace, any guidance from a major professional organization deserves careful attention.
After reading the Primer cover to cover, the verdict is mixed. The document gets several things right at a conceptual level, particularly its emphasis on continuous validation and the risks of overreliance on AI-generated outputs. But for a publication arriving in spring 2026, when actuaries are already using LLMs daily for code generation, reserve analysis, and regulatory filing preparation, the Primer reads more like an executive briefing for a chief actuary who has never touched an LLM than practical guidance for the working professionals it claims to target. The most consequential gaps are not in what it covers, but in what it leaves out entirely.
What the Primer Gets Right
Credit where it is due. The CAS AI Primer correctly frames three principles that should anchor any actuary's approach to AI adoption.
First, its emphasis on validation loops is sound. The Primer identifies three validation methods: human-in-the-loop review, cross-validation across different LLMs, and prompt sensitivity testing. This last point is especially valuable. From tracking LLM behavior across actuarial use cases over the past year, prompt sensitivity testing is one of the most underused techniques in practice. Rephrasing a loss development prompt slightly and watching the output shift dramatically is one of the fastest ways to gauge whether a model's response reflects genuine pattern recognition or surface-level text prediction. The Primer's inclusion of this method signals that the working group has practical experience with these tools.
Second, the three-tier framework for specializing AI models (prompt engineering, retrieval-augmented generation, and fine-tuning) provides a useful mental model for actuaries evaluating how deep they want to go. The progression from low-cost prompt adjustments to high-investment fine-tuning maps well to how most insurance companies are actually approaching the build-versus-buy decision.
Third, the corporate governance section correctly identifies the interdisciplinary nature of AI risk management, noting that risk management, legal, compliance, IT, data teams, and business units all need seats at the table. This mirrors what patterns we have seen across carrier AI governance programs: the organizations struggling most are those that have siloed AI adoption within a single department.
Gap 1: The Regulatory Landscape Is Nearly Invisible
The most glaring omission in the CAS AI Primer is the near-total absence of the regulatory framework that actuaries must navigate when deploying AI. The document mentions regulatory compliance in passing, noting that models "should always be monitored to ensure that they comply with all applicable laws and regulations." But it never specifies which laws and regulations those are, how they apply to actuarial work, or what compliance actually looks like in practice.
This is a significant miss, because the regulatory environment for AI in insurance has matured substantially since 2023. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, is the single most important regulatory document governing AI use in the U.S. insurance industry. As of early 2026, approximately 24 states have adopted the Model Bulletin in full or with minimal modifications, including Connecticut, Delaware, Kentucky, Maryland, Massachusetts, New Jersey, North Carolina, Oklahoma, Pennsylvania, Rhode Island, Vermont, and West Virginia.
The Model Bulletin requires insurers to develop, implement, and maintain a documented AI program (an "AIS Program") that governs the responsible use of AI in regulated insurance decisions. The program must include governance frameworks, risk management controls, internal audit functions, bias testing and mitigation procedures, and third-party vendor oversight. It also establishes that regulators may request documentation during examinations or investigations.
For practicing actuaries, this means that any AI tool used in pricing, reserving, underwriting, or claims handling is now subject to specific regulatory expectations in roughly half the states. The CAS Primer does not mention the NAIC Model Bulletin once.
Beyond the Model Bulletin, several states have enacted their own targeted requirements. New York's DFS Circular Letter 2024-7 requires insurers to demonstrate that AI systems do not proxy for protected classes. Colorado's C.R.S. 10-3-1104.9 requires quantitative disparate impact testing for AI models used in life insurance, with expansion to auto insurance and health benefit plans effective October 2025. California restricts health insurers from relying solely on automated tools for coverage determinations. A working actuary building or validating AI models needs to know these specifics. The Primer provides none of them.
Gap 2: ASOP No. 56 and the Professional Standards Framework
Perhaps the most puzzling omission is the absence of any substantive discussion of Actuarial Standard of Practice No. 56, Modeling. ASOP No. 56 has been effective since October 1, 2020, and it applies to actuaries in any practice area when performing actuarial services with respect to designing, developing, selecting, modifying, using, reviewing, or evaluating all types of models. The scope is explicitly comprehensive.
Here is the critical connection the Primer misses: AI is a model, and ASOP No. 56 applies to it directly. This is not a debatable interpretation. The American Academy of Actuaries' September 2024 discussion paper on Actuarial Professionalism Considerations for Generative AI states this plainly: "GenAI is a model; thus ASOP No. 56 applies." The Academy's March 2026 article in Contingencies on actuarial and algorithmic accountability reinforces the same point: "ASOP No. 56, Modeling, which affects all practice areas, provides guidance on any type of model, including related algorithmic approaches such as AI."
Under ASOP No. 56, an actuary using an AI model must evaluate the model's appropriateness for its intended use, assess the quality of input data, validate model output, disclose limitations, and document reliance on experts. Section 3.1 requires the actuary to consider whether the model is appropriate for the intended purpose and whether the structure, data, and assumptions are suitable. Section 3.2 addresses data quality. Section 3.6 covers model output validation. These are not abstract principles; they create enforceable professional obligations. An actuary who signs off on AI-generated reserve estimates without following ASOP No. 56 requirements could face a hearing before the Actuarial Board for Counseling and Discipline (ABCD).
The CAS Primer's validation section touches on similar concepts, but without connecting them to the binding professional standards that give them teeth. A practitioner reading only the Primer would come away thinking that validation is a best practice rather than a professional obligation.
Equally important are ASOP No. 23 (Data Quality) and ASOP No. 41 (Actuarial Communications), both of which apply to AI-assisted actuarial work. ASOP No. 23 requires actuaries to review data for reasonableness and consistency, which is directly relevant when using AI-processed or AI-augmented data. ASOP No. 41 requires clear communication of assumptions, methods, and limitations in actuarial work products, including disclosure of reliance on AI tools. The Academy's discussion paper explicitly addresses how these standards interact with generative AI use cases.
A document titled "Practical Guidance for Actuaries" that does not mention ASOP No. 56, ASOP No. 23, ASOP No. 41, or the Code of Professional Conduct leaves the most important part of practical guidance on the table.
Gap 3: The Model Comparison Table Has a Short Shelf Life
The Primer's appendix includes a comparison table of leading GenAI models: GPT-5, Grok 4, Gemini 2.5 Pro, Claude 4 Opus, and Llama 4 Maverick. The table lists context windows, pricing, benchmarks, insurance use cases, and access methods.
This is the kind of content that looks helpful on first read but creates more problems than it solves. Model capabilities, pricing, and benchmarks are changing on a quarterly (sometimes monthly) basis. The table's own footnote acknowledges this: "Metrics vary by provider and workload. Values shown are representative as of August 2025." By the time the Primer was published in March 2026, several of these metrics were already stale.
More fundamentally, a static model comparison table does not help actuaries make the decision they actually need to make: which model is appropriate for their specific use case within their specific regulatory and governance environment. The better approach would have been to publish decision criteria (latency requirements, data sensitivity thresholds, auditability needs, cost tolerances) that remain stable even as the models themselves evolve. The "matching model to task" quick guide on page 18 moves in this direction but stays too high-level to be actionable.
From working with actuarial teams evaluating LLMs, the most common mistake is not picking the "wrong" model. It is failing to establish evaluation criteria that reflect their regulatory environment and risk appetite before evaluating any model at all. The Primer reinforces this pattern by leading with product features rather than decision frameworks.
Gap 4: No Practical Workflow Examples
The Primer describes AI adoption in conceptual terms: specialize the model, validate the model, avoid pitfalls. But it never shows what responsible AI use looks like inside an actual actuarial workflow.
Consider a pricing actuary using an LLM to process unstructured loss descriptions into categorized segments for ratemaking analysis, which is one of the Primer's own examples. A genuinely practical guide would walk through the end-to-end workflow: how to structure prompts for consistent classification, how to validate LLM output against manually coded samples, what acceptance thresholds look like, how to document the methodology for regulatory filing support, what disclosures are required under ASOP No. 56, and how to handle cases where the LLM's classification confidence is low.
Or consider a reserving actuary using AI to assist with loss development factor selection. The Primer mentions "triangle development" as a potential AI task but does not address the professional judgment questions this raises. If an actuary uses AI-generated development factors in a Statement of Actuarial Opinion, how should that reliance be disclosed? What validation is sufficient? How does this interact with ASOP No. 43 (Unpaid Claim Estimates)?
The SOA's February 2024 research report, A Primer on Generative AI for Actuaries, provides a more technical overview of the mechanics of generative AI. The SOA also published Operationalizing LLMs in early 2025 with practical guidance on integrating LLMs. The CAS Primer could have built on these foundations with P&C-specific workflow examples. Instead, it stays at a level of abstraction that makes it difficult to translate concepts into daily practice.
Gap 5: The Bias and Fairness Treatment Is Thin
Algorithmic bias is arguably the highest-stakes AI issue in insurance today. State regulators are actively investigating whether AI-driven pricing and underwriting models produce unfairly discriminatory outcomes. Colorado requires quantitative disparate impact testing. New York requires documentation that AI systems do not proxy for protected classes. The NAIC Model Bulletin makes bias testing a core component of the required AIS Program.
The Primer's treatment of bias amounts to a single mention within the corporate governance section, noting that "mitigation of model bias" is an element of a robust governance framework. It does not define what bias means in an insurance context, distinguish between actuarially justified risk differentiation and unfair discrimination, describe methods for detecting proxy discrimination, or reference the substantial body of work that the Academy, the CAS itself, and state regulators have produced on this topic.
The Academy's Data Science and Analytics Committee has published multiple papers on algorithmic bias in insurance, including the 2021 issue paper "Big Data and Algorithms in Actuarial Modeling and Consumer Impacts" and a 2023 issue brief on "Discrimination: Considerations for Machine Learning, AI Models, and Underlying Data." These provide frameworks for bias detection that are directly applicable to actuarial AI implementations. The CAS Primer does not reference them.
For a CAS publication, this is a missed opportunity. Property and casualty actuaries are on the front lines of the bias debate because they work with the pricing and underwriting models most subject to regulatory scrutiny. A practical guide should equip them with the vocabulary, methods, and regulatory awareness to navigate this terrain confidently.
Gap 6: Environmental Impact Is Mentioned but Not Quantified
The Primer's closing paragraph on environmental sustainability is a welcome inclusion, but it reads as an afterthought. It correctly notes that large AI models require substantial computing resources with a significant carbon footprint. But for actuaries, who are trained to quantify risk and cost, the absence of any numbers is notable.
A more useful treatment would contextualize the cost-benefit tradeoff: what does a typical LLM API call cost in both dollars and energy? How does running a fine-tuned model on cloud infrastructure compare to using a general-purpose API? These are operational decisions that actuarial teams are making right now. The Primer could have framed environmental impact as an underwriting and enterprise risk management consideration, connecting it to the profession's existing expertise rather than treating it as an afterthought.
What Would Have Made This Better
The CAS AI Primer is not a bad document. It is an incomplete one, and its incompleteness is most damaging in the areas that matter most to practicing actuaries: professional standards compliance, regulatory navigation, and practical implementation.
A stronger version of this document would have included: a dedicated section on ASOP No. 56, No. 23, No. 41, and the Code of Professional Conduct as they apply to AI use; specific discussion of the NAIC Model Bulletin and its state-level adoption, with practical compliance guidance; at least two or three end-to-end workflow examples showing responsible AI use in pricing, reserving, or claims analysis; a substantive treatment of algorithmic bias detection and mitigation methods relevant to P&C actuarial work; a decision framework for model selection that is criteria-based rather than product-based; and cross-references to the Academy's professionalism discussion paper, the SOA's GenAI research, and the CAS's own prior publications on AI in insurance.
The Primer's own co-author, Shine Wang, FCAS, described it as "a starting point rather than a comprehensive rulebook." That framing is fair, and it is worth acknowledging that creating guidance in a rapidly moving field is genuinely difficult. The CAS AI Working Group also has an active Request for Proposals on Adapting Large Language Models for Specialized P&C Actuarial Reasoning, offering up to $40,000 in research funding, which signals that deeper work is coming.
But "starting point" documents carry a risk. If an actuary reads the CAS AI Primer and comes away believing they have a reasonable understanding of what responsible AI use looks like, they are missing critical pieces: the professional obligations under ASOP No. 56, the regulatory requirements under the NAIC Model Bulletin, and the state-specific mandates that are being enforced right now. The Primer would have been stronger with a prominent disclaimer directing readers to these resources.
Recommendations for Practicing Actuaries
Until the CAS produces more comprehensive guidance, actuaries using AI tools should supplement the Primer with the following resources:
Read the Academy's Actuarial Professionalism Considerations for Generative AI (September 2024). This is currently the most thorough treatment of how existing professional standards apply to AI use, with specific discussion of ASOPs No. 56, 23, and 41.
Review the NAIC Model Bulletin on the Use of AI Systems by Insurers and check whether your state has adopted it. The NAIC maintains a tracking document listing state-by-state adoption status.
Familiarize yourself with ASOP No. 56 if you have not reviewed it recently, with specific attention to Sections 3.1 (model appropriateness), 3.2 (data quality), and 3.6 (model output validation) as they apply to AI tools.
Document your AI use. If you are using LLMs to assist with any actuarial work product, maintain records of the models used, the prompts provided, the validation performed, and the professional judgment applied to the output. This documentation will be essential if your work is subject to regulatory examination or peer review.
Build validation into your workflow from the start. Do not bolt it on after the fact. The Primer's advice on cross-validation and prompt sensitivity testing is genuinely useful; the gap is in recognizing that these practices are not optional best practices but components of your professional obligations under ASOP No. 56.