2026 CAS AI in Action: Practical Innovations Transforming Actuarial Practice Seminar

The CAS AI Primer in Action

Description:

Generative AI has moved from novelty to necessity in actuarial work, but most teams face the same question: where do we actually start? This session walks through the practical adoption framework developed by the CAS Artificial Intelligence Working Group in The CAS AI Primer, covering the spectrum from exploratory chatbot use to fully agentic workflows. Attendees will learn how to identify high-value AI use cases, choose the right level of model specialization (prompt engineering, RAG, fine-tuning, or platform-native customization), and understand the unique risks agentic AI introduces — cascading errors, prompt injection, and automation bias — along with the governance needed to manage them. 

Learning Objectives:

1. Apply the four-level AI adoption framework (exploratory chatbots, AI specialization, infrastructure agents, custom builds) to assess where their own team's use cases fit and what a realistic next step looks like.
2. Distinguish between the main model specialization methods — prompt engineering, RAG, fine-tuning, and platform-native customization (e.g., Claude Projects/Skills, Custom GPTs) — and identify which is appropriate for a given actuarial task and risk tolerance.
3. Identify the key risks specific to agentic AI systems and outline best practices for responsible AI use.
4. Understand the key corporate and regulatory considerations that shape AI implementation in actuarial contexts.

Speaker Bios:

Nick Easley, ACAS, has over ten years of property and casualty insurance experience. He currently serves as Senior Actuarial Associate at MS Transverse Insurance Group, where his work encompasses loss reserving, program pricing, and actuarial technology development.

Puxuan (Shine) Wang, FCAS, FCIA, is Manager of Corporate Actuarial Forecasting at Wawanesa Mutual Insurance, where the focus is capital projection, financial condition testing, and financial planning. Shine has a background spanning advanced analytics, data science, and P&C reserving — including IFRS 17 and reserving optimization.

Custom AI Development for Reinsurance Pricing Workflow

Description:

Reinsurance pricing depends on broker submissions that are heterogeneous, internally linked, and often difficult to validate before actuarial judgment can be applied. We develop two complementary Custom GPT tools for this workflow. The first compares prior- and current-year broker submissions, using an actuarial rule catalog and few-shot examples to identify and localize pricing-relevant inconsistencies. The second analyzes a validated submission, resolves missing or ambiguous actuarial assumptions through a structured clarification step, and executes experience-rating calculations through a deterministic Python function library with complete input and output provenance. In out-of-sample validation examples, the custom validation tool materially outperformed generic ChatGPT. Experience-rating examples further illustrate triangle-based Chain Ladder (CL) and Bornhuetter-Ferguson (BF) calculations and claim-level excess-layer analysis using various methods. The tools produce standardized human-readable reports together with machine-readable JSON registers that preserve workbook locations, assumptions, function arguments, calculation outputs, and limitations. The results show how customizing the GPTs can make LLM assistance substantially more reliable and operationally useful in reinsurance pricing.

Learning objectives:

1. Identify and explain pricing-relevant inconsistencies across consecutive-year broker submissions, including claim-level, policy-level, and cross-tab issues.
2. Apply a controlled custom-AI workflow to calculate experience-rating quantities such as on-level premium, trended and developed losses, layer losses, and loss ratios.
3. Understand how customizing the AI assistance tools can make it substantially more reliable and operationally useful in reinsurance pricing.

Speaker Bios:

Phoebe Choi, ACAS, is a reinsurance pricing actuary who helps teams make AI reliable enough to trust in real pricing work. She has priced US Casualty treaties at Hamilton Reinsurance and spent nearly four years in reserving at Arch Reinsurance optimizing quarterly IBNR meeting workflow, and performed studies across property catastrophe, D&O, and medical malpractice. A research collaborator on a CAS-funded AI project detecting errors in broker submissions, she blends hands-on reinsurance experience with a data science toolkit in machine learning, Python, and SQL to make custom AI practical at the desk.

Tsz Chai (Samson) Fung, Ph.D., FSA is a tenure-track Assistant Professor of Actuarial Science in the Maurice R. Greenberg School of Risk Science at Georgia State University. His research program is centered on actuarial P&C reserving and claims analytics, with a particular emphasis on stochastic and micro-level reserving, dependence modeling, and statistical inference for insurance data. His work is multi-disciplinary, drawing from actuarial science, statistics, econometrics, and machine learning to design models that are both theoretically grounded and practically implementable. His publication record spans leading outlets in actuarial science (e.g., NAAJ, IME, SAJ, ASTIN), risk management and insurance (Journal of Risk and Insurance), econometrics (Journal of Econometrics), operations research (European Journal of Operational Research), statistics (Journal of the Royal Statistical Society, Annals of Applied Statistics), and machine learning (Neurocomputing). He received multiple competitive research grants from the Society of Actuaries (SOA) and the Casualty Actuarial Society (CAS) as Principal Investigator and Co-Principal Investigator. These projects span a range of topics, including longitudinal actuarial modeling, trade credit insurance networks, reinsurance pricing, social inflation, healthcare reserving, and the development of AI-driven methods for insurance analytics.

Practical AI for Actuaries

Description: 
This session focuses on practical applications of AI for actuarial work. We will demonstrate tangible examples of how frontier AI tools can be used for actuarial workflows, share our research findings building a custom AI assistant, and discuss the current capabilities, limitations, and future opportunities for actuaries adopting AI.

Learning Objectives:

1. Identify applications of frontier AI tools to real-world actuarial tasks.
2. Describe the design of an actuarial AI coding assistant and interpret benchmark-testing results for reserving and ratemaking workflows.
3. Explain the limitations of AI for actuarial calculations and identify opportunities for its use within the actuarial profession.

Speaker Bio:

Sabrina Tan, FCAS, has over five years of actuarial experience and is currently an Actuary at Uber. She has experience in pricing, predictive modeling, and machine learning applications in insurance. She is currently part of Uber’s US Pricing team, where she conducts pricing analyses, negotiates insurance rates and terms, and provides analytics support. She has a consulting background and worked on various pricing and reserving automation projects, leveraging AI/ML and developing actuarial software tools.

Catastrophe Duration and Loss Prediction via Natural Language Processing

Description:

Textual information from online news is more timely than insurance claim data during catastrophes, and there is value in using this information to achieve earlier damage estimates. In this paper, we use text-based information to predict the duration and severity of catastrophes. We construct text vectors through Word2Vec and BERT models, using Random Forest, LightGBM, and XGBoost as different learners, all of which show more satisfactory prediction results. This new approach is informative in providing timely warnings of the severity of a catastrophe, which can aid decision-making and support appropriate responses.

Learning objectives:

1. Explain how information from online news articles can be used to generate timely estimates of catastrophe duration and insurable economic losses.
2. Assess the potential of online news articles to provide early warnings of extreme catastrophe events.
3.  Identify the words and topics that are the most significant indicators of catastrophe severity.

Speaker Bio:

Dr. Han Li is an Associate Professor at the Centre for Actuarial Studies, Department of Economics. She is an Associate of the Institute of Actuaries of Australia, and has a broad range of research interests around longevity and mortality risks, ageing and retirement, and the impact of climate change on insurance industry. Specifically, much of her research expertise centers on actuarial modelling and forecasting using advanced econometric and statistical techniques. She has attracted research funds from the Australian Research Council, the Society of Actuaries, the Casualty Actuarial Society, the SCOR Corporate Foundation for Science, and the Australia-Germany Joint Research Cooperation Scheme (DAAD). Han's research has been published in top-tier journals including Insurance: Mathematics and Economics, ASTIN Bulletin, North American Actuarial Journal, Scandinavian Actuarial Journal, Annals of Actuarial Science, Journal of the Royal Statistical Society: Series A, Journal of the Royal Statistical Society: Series C, Journal of Forecasting, International Journal of Forecasting, Energy Economics, Annals of Operations Research, BMC Medicine, and Population Health Metrics.

LLMs for Unstructured Claims Data

Description:

Most of what an insurer collects in their claims records is stored in adjuster notes, medical records, transcripts, and claimant statements. These unstructured data sources often contain rich information that is not converted into structured forms and used for reserving, ratemaking, and claims management. This session presents a CAS-funded prototype that uses large language models to extract actuarially meaningful variables from unstructured claims documents in a repeatable, auditable way. We show the two-stage extraction and compound scoring pipeline created in the project, a variable taxonomy of 36 candidate features (14 implemented in the prototype), and a validation design built on synthetic Synthea-based records scored independently by clinician reviewers. The talk closes with how extracted features integrate with methods such as chain ladder reserving, the limitations of the current approach, and what a production-scale deployment would require.

Learning objectives:

1. Identify the actuarial value locked in unstructured claims data 
2. Explain how LLM-based structured extraction differs from summarization
3. Evaluate how LLM-extracted features can be integrated into existing actuarial methods 

Speaker Bio:

Rob Lieberthal, PhD, is the Founder and Senior Principal of Lieberthal & Associates LLC, where his work spans health economics, actuarial science, and applied machine learning. He previously developed a machine learning method for fraud detection in healthcare claims that matched physician-expert accuracy and currently serves as Senior Advisor on Behavioral Economics to the CMS Center for Medicare and Medicaid Innovation (CMMI). He is the co-lead author of "Leveraging LLMs for Unstructured Claims Data Analysis,” along with Dr. Richard Tran of MDSight. This was a Casualty Actuarial Society-funded study demonstrating how large language models can extract structured actuarial variables from medical records and claims documents to improve reserving accuracy. He is also the author of "What Is Health Insurance (Good) For?” a book published by Springer that examines the structure, funding, and future of the U.S. health insurance system. Rob holds a PhD in Health Economics from the University of Pennsylvania Wharton School of Business and a Bachelor’s in Mathematics from Boston University.