A Practical AI Actuarial Solution Equipped with Four Technologies That Surpass the Limitations of LLMs

Authored by Justin Hwang, Head of the AI Project, Donghwan Lee, AI Team Lead and Scarlett Bae, AI Specialist

LLMs: The Most Advanced AI Models in Modern Society

Tracing the historical development of AI models reveals a remarkable evolution—from early rule-based systems to today’s highly sophisticated Large Language Models (LLMs) such as GPT and Gemini. The journey has progressed through stages of machine learning and deep learning, propelled by advances in both technology and data accessibility. As a result, AI has become indispensable across a wide range of industries.

Among the recent breakthroughs in AI, LLMs stand out as the most talked-about innovation. These deep neural network models are specifically designed to understand and generate human language by learning from massive volumes of text data—ranging from billions to trillions of words. LLMs excel at grasping complex relationships and contextual nuances within language.

However, despite their impressive capabilities, LLMs can still feel abstract or difficult to grasp for some users. Fundamentally, an LLM is a deep learning model—a complex composite made up of numerous nonlinear regression models. By incorporating transformer architecture, LLMs gain a strong ability to understand long-range dependencies in text, enabling them to engage in human-like natural language conversations. This combination of structure and functionality is what defines a Large Language Model.

[Figure 1. Neural Network Model and Nonlinear Regression Model]

 

Limitations of LLMs and How to Overcome Them

The strengths of Large Language Models (LLMs) are undeniable. With their exceptional natural language processing capabilities, they are widely used in document analysis, customer service, content creation, and many other areas. Their flexibility and scalability—rooted in extensive training on massive datasets—are unmatched.

However, LLMs also exhibit clear limitations. One of the most prominent is the phenomenon of hallucination, where the model generates information that sounds plausible but is factually incorrect. In addition, due to a lack of in-depth domain expertise and advanced reasoning capabilities, LLMs often struggle with complex Excel spreadsheets—where formulas and tables are interconnected across multiple sheets—perceiving them more like a maze than a structured dataset.

These limitations pose serious risks when applying LLMs to actuarial science, a field where accuracy and reliability are paramount. Outputs based on incorrect information can lead to financial instability, regulatory violations, and loss of trust. Therefore, before deploying LLMs in such high-stakes environments, it's essential to explore methodologies that reinforce factual grounding and logical structure.

Solution 1: Fine-Tuning - Enhance actuarial expertise through additional model training

Another practical limitation of LLMs is their lack of contextual adaptation to specific organizations or industries. Since LLMs are generally trained on broad public data sources such as Google or Wikipedia, they often fail to fully grasp the complex contract structures of insurance products or detailed regulatory frameworks like IFRS 17 and capital requirements. This lack of domain-specific knowledge is a significant barrier to practical application.

One effective way to overcome this challenge is through fine-tuning. Fine-tuning is the process of training a pre-trained LLM (base model) further using domain-specific data, terminology, and task patterns. In essence, it transforms a general-purpose AI into an "in-house expert," tailored to understand and perform tasks within a specialized professional context.

[Figure 2. Concept of Fine-tuning, Source: medium.com]

Fine-tuning is carried out in the following way:

  1. Collect domain-specific text data: Gather documents such as product brochures, foundational documents, actuarial and assumption documents, and internal manuals.

  2. Create instruction dataset: Design pairs of user questions and ideal answers that the model should provide.
    (e.g., Q: “What are the main benefits of this product?” → A: “It includes cancer diagnosis benefit, cancer hospitalization, surgery, and outpatient benefits. The diagnosis benefit is paid only once upon the first diagnosis…”)

  3. Retrain the LLM-based model: Partially update the parameters of the base model to build a language model tailored to the organization.

  4. Validation and performance evaluation: Test the model’s suitability and accuracy using real business-related questions.

This approach is especially effective in the following situations:
• When the domain is highly specialized, making it difficult to get consistent results with prompt-based methods alone
(e.g., insurance actuarial software code used to calculate premiums, reserves, or cash flows for a specific insurance product)
• When consistent standards must be maintained for repeated business questions
• When the model must be operated independently on-premise in a network-isolated environment without API access

RNA Analytics is currently conducting phased experiments using client-specific datasets to fine-tune the AI-driven actuarial automation system, R3S Modeling Agent, aiming to improve both task suitability and response consistency.

Fine-tuning is both a “technology for enhancing performance” and a “strategy for ensuring trust.” In the future, for AI to deliver consistent judgments and explainable results in the insurance industry, fine-tuning and organizational customization will be critical capabilities beyond simply choosing a pre-trained model.

Solution 2: RAG (Retrieval-Augmented Generation) - Improve accuracy by integrating external knowledge sources
One of the key strategies to overcome the limitations of traditional language models is RAG. RAG enables a Large Language Model (LLM) to retrieve information from external knowledge sources in real time (Retrieval), supplement its response with that information (Augmentation), and then generate the final answer (Generation). Unlike conventional LLMs that rely solely on pre-trained parameters, RAG significantly improves accuracy and reliability by referencing trusted external data.

Let’s take a closer look at the components and workflow of RAG:

  • Query: The user’s input question or request

  • Retrieval: Semantically relevant information is retrieved from an external knowledge base based on the input query

  • Augmentation: The retrieved information is combined with the original query and passed to the response generation model

  • Generation: The model generates the final response using the augmented input

In conclusion, RAG is a practical and powerful framework for structurally controlling hallucinations. It goes beyond merely compensating for the weaknesses of LLMs—it is emerging as a core technology for building trustworthy AI response systems. RAG is particularly well-suited to high-accuracy domains such as actuarial work in the insurance industry.

[Figure 3. RAG Flow and Retrieval Methodology]

Solution 3: Dataset & Document Formatting - Make documents easier for AI to read and understand

Above all, the decisive factor for the success of any AI project is the dataset. The quality of data directly impacts the performance of AI. Looking at the insurance industry as an example, some companies have already invested large sums in AI systems but have failed to achieve the expected results.

One key reason is the document format. Many documents containing insurance data are written in formats that AI struggles to read. This issue goes beyond simple typos or errors. Often, the documents are not structured in a way that AI can systematically understand. Since the quality of documents is the most critical factor in using AI to improve actuarial productivity, it is essential to address three key document preparation practices that must be improved.

First, it is time to move away from PDF-centric document creation.
PDFs are designed primarily for printing. While they look good to the human eye, their structure is ambiguous for machines to interpret. Recent attempts to analyze PDFs using AI technologies like OCR (Optical Character Recognition) or Vision Transformers have been made, but these methods have limitations in accuracy and require significant time and cost for preprocessing and postprocessing.

On the other hand, formats such as .docx, .tex, .html, and .md (Markdown) are text-based, globally recognized standards that AI can accurately parse. Notably, Microsoft’s open-source Markdown project has active participation from developers worldwide, making it a stable and reliable format for various insurance business documents.

If your company still relies on country-specific word processors or non-standard document formats, there is a high risk of falling behind in the wave of AI transformation. It is urgent to either develop dedicated tools to convert these documents into AI-readable formats or to accelerate a company-wide effort to transition to standardized document formats.

Second, Use LaTeX or KaTeX for Formulas Instead of Images

Complex formulas frequently appear in actuarial work. However, it is still common to see formulas inserted as images. Image-based formulas cannot be read by AI. Although OCR can somewhat recognize them, the accuracy is low and costs increase.

The correct approach is to input formulas using TeX syntax. Even if formulas appear well-rendered visually, if the underlying code is poorly structured, AI will not be able to interpret them. This is a classic case of “Garbage in, Garbage out.”

KaTeX is especially recommended. It renders quickly in web browsers and is easy for non-experts to learn, making it easy to adopt across an organization.

Third, Avoid Formatting Entire Documents as Tables

Arranging an entire document in table format to control layout may look neat to humans, but from an AI’s perspective, it’s almost like an encrypted document. Tables obscure the semantic structure of a document—such as headings, paragraphs, and sections—making it very difficult for AI to understand the context. Especially when the main text, subheadings, and explanations are all placed inside tables, AI struggles to separate and interpret the intended messages.

When preparing documents, it is essential to use semantic formatting features like ‘heading styles,’ ‘paragraphs,’ and ‘lists’ provided by word processors. This approach not only benefits AI readability but also greatly improves document searchability and maintenance.

Insurance companies possess vast amounts of data assets. However, if these data do not exist in a form that AI can read and write, their value cannot be realized.

AI transformation is not merely a matter of adopting new technology. It is a strategic shift to standardize information structures and create documents that both humans and machines can understand.

Starting now, review your company’s document formats. Remove PDFs and image-based formulas, and transition to AI-friendly document structures. The starting point for AI adoption is not the ‘algorithm,’ but the ‘document.’

Solution 4: Ontology-Based Database - Helping AI Understand Meaning

In the actuarial field, building an ontology-based database is critically important. Ontology clearly defines and structures concepts and terminology to help AI understand and process information more accurately. Using ontology in actuarial work greatly enhances interoperability between data and structural understanding, enabling more precise and faster decision-making.

Ontology is a system that explicitly defines the concepts and relationships within a specific domain, structuring knowledge such as insurance product structures, actuarial/statistical/financial techniques, legal and accounting regulations, and internal company policies and manuals. By storing the resulting knowledge graph in a database and using it as needed, large language models (LLMs) can respond to questions more accurately, understand context better, and infer relationships among information.

For example, when asked about the calculation method for reserves of a specific insurance product, an LLM can comprehensively search the ontology-based knowledge graph database for relevant regulations, actuarial methodologies, and similar existing products to generate a reliable answer. At the same time, it can visually present which data and concepts the answer is based on, enhancing transparency and user trust.

[Figure 4. Example of a Knowledge Graph for a Hypothetical Cancer Insurance Product]

To apply this technology in practice, collaboration among actuaries, data scientists, and AI engineers is essential, along with a phased strategy for building ontologies and knowledge graphs. Technologies for automated relationship extraction and updates, as well as the integration design between LLMs and knowledge graphs, are also emerging as critical components.

Donghwan Lee, Head of the AI Lab at RNA Analytics, stated, “LLMs have the potential to dramatically enhance the efficiency and accessibility of actuarial work, but ensuring reliability is essential for their safe application. Data quality, standardized document structures, and ontology-based databases are key to overcoming the limitations of AI and driving true innovation in actuarial practices.”

RNA Analytics’ AI Solution

RNA Analytics is developing next-generation AI-based solutions for the insurance industry. One such solution is R3S Thalexa™, designed to enhance actuarial modeling capabilities in conjunction with the company’s flagship actuarial software, R3S Modeler.

Thalexa—named after the thalamus of the brain, which serves as the hub for transmitting information—symbolizes its role as a central hub within the actuarial ecosystem, connecting and processing diverse information across the insurance industry.

 

R3S Thalexa™

Core Function 1: Knowledge Agent

  • Automatically extracts information from document-based input materials and populates the BB-Matrix (a table containing product structure and assumption-related data).

  • Automatically builds an ontology-based database from document-based input materials.

Core Function 2: Modeling Agent

  • AI-powered Chatbot: Provides real-time Q&A support for actuarial theory, insurance regulations, and R3S Modeler documentation.

  • R3S Modeler Code Assistance: Enhances modeling productivity by suggesting code in real time.

  • Autonomous AI Modeling System for R3S Modeler: Automatically generates actuarial models to minimize manual effort and maximize efficiency.

R3S Thalexa™ is an AI-driven insurance product pricing automation platform that extracts product structure and assumption data from documents, refines it into a model-ready format, and connects it through to premium and reserve calculations, as well as profitability analysis.

With Thalexa, insurers can dramatically reduce the time and human resources required during product revisions and new product development, enabling more sophisticated profitability assessments and faster responses to market demands.

[R3S Thalexa™ User Interface]

 

RNA Analytics is steadily expanding the core features of its AI solutions. In January 2025, the company developed its flagship actuarial language model, ActyLLM, and by July 2025, it will introduce prompt-based Q&A capabilities and model code suggestion features. By the end of the year, RNA plans to launch a fully autonomous model-building AI through the Modeling Agent, capable of generating complete model structures. The AI solution is designed to evolve into an increasingly powerful engine as user feedback accumulates after its release.

This groundbreaking technology is being developed and refined by RNA Analytics’ highly skilled AI professionals. The AI Lab at RNA is composed of experts with multinational experience across actuarial science, software development, data science, and artificial intelligence. Drawing on real-world expertise in markets such as Korea, the UK, and Australia, the team delivers tailored AI solutions to meet the unique needs of each client.

[Figure 5. R3S AI Solutions: System Architecture and Output Results]

Justin Hwang, Head of the AI Project at RNA Analytics, stated;

“Our AI solutions, led by ActyLLM and R3S Thalexa™, go beyond mere technology. They are designed to be strategic AI partners that drive business innovation and digital transformation across the insurance industry. We aim to deliver tangible value through automated product development, profitability analysis, and support for regulations like IFRS 17 and solvency requirements—ultimately empowering our clients to enhance their competitiveness.”

The adoption of AI in actuarial work has moved beyond the experimental phase. It is now evolving toward genuine operational automation and enhanced information accuracy, made possible by sophisticated knowledge structures and integrated system design.

RNA Analytics