GenAI Beyond the Basics: Advanced Concepts for Actuaries

Generative AI (GenAI) tools such as ChatGPT are rapidly reshaping how actuaries approach problem-solving, analysis, and communication. Beyond their familiar chat interfaces, these tools offer far more powerful capabilities through programmatic access, allowing users to interact directly with the underlying Large Language Models (LLMs) via APIs (Application Programming Interfaces). Compared with manual, web-based use, API access lets actuaries embed GenAI seamlessly into their existing workflows, process larger volumes of data, and automate repetitive tasks at scale. In this three-hour web session, participants will follow live demonstrations of advanced GenAI concepts in a Jupyter notebook that introduces each concept and applies it to a concrete actuarial use case. The notebook will be shared with all attendees to encourage experimentation and support adoption in their own work.

After a foundations-and-best-practices block that covers how to access LLMs both through cloud APIs and locally via Ollama, together with practical prompt-engineering techniques, we will explore the following advanced GenAI concepts:

• Structured Outputs: Generating responses in structured formats like JSON to support easier and more reliable downstream processing.
• Function Calling: Enabling LLMs to execute predefined functions, such as calculations or database queries, to perform specific operations.
• Fine-Tuning: Customizing pretrained LLMs with domain-specific data to improve accuracy and relevance in generating responses.
• Retrieval-Augmented Generation (RAG): Combining LLMs with external data sources to produce contextually enriched outputs.

The session then closes with a dedicated introduction to Agentic AI, showing how a single LLM can be turned into an autonomous, tool-using agent and how several such agents can be orchestrated as a multi-agent system that collaborates to automate complex, multi-step actuarial workflows.
For each concept, the session will explain its purpose and underlying principles, demonstrate it through a dedicated actuarial use case, and point to further applications and resources. It will conclude with a forward-looking outlook on emerging developments, followed by an open Q&A and discussion.

The aim of this web session is to equip participants with the knowledge and practical skills needed to apply advanced GenAI concepts in actuarial contexts. By the end of the session, participants will be able to:

• Understand how to interact with LLMs programmatically through APIs.
• Run open-weights LLMs locally via Ollama and weigh the trade-offs between local and cloud deployment.
• Apply Python programming skills to develop and integrate GenAI-powered solutions into actuarial workflows.
• Apply prompt-engineering best practices, such as few-shot prompting, chain-of-thought reasoning, and clear role and
  output-format instructions, to obtain reliable and reproducible results.
• Generate structured outputs in JSON format to ensure compatibility of the responses with actuarial processes and systems.
• Leverage Function Calling capabilities to trigger the execution of predefined functions in the underlying programming language, such as calculations or database queries.
• Customize LLMs using fine-tuning techniques to address domain-specific requirements.
• Combine external data with LLMs through Retrieval-Augmented Generation to produce contextually relevant outputs.
• Build agentic AI solutions, including multi-agent systems in which several specialized agents collaborate to automate complex, multi-step actuarial tasks.

By balancing conceptual understanding with hands-on programming, the session gives participants practical tools to enhance the accuracy, efficiency, and creativity of their actuarial work.