Blog

What Is Fine-Tuning in Large Language Models—and Why Does It Matter for Insurance AI?

Written by John Cottongim | May 28, 2025

Artificial intelligence has transformed our industry. In the past year, AI has moved beyond proving its potential to streamline operations, becoming a force that improves and enhances performance in areas from claims processing to underwriting to customer service.  

Many of these advances are powered by fine-tuned large language models (LLMs), which can fluently understand and generate natural language. However, to truly deliver value in insurance, these models need more than general language skills—they need to understand the insurance industry’s specialized vocabulary, rules, and workflows.  

And this is where fine-tuning comes in. 

 

What Is Fine-Tuning in Large Language Models? 

Fine-tuning involves “teaching” a pre-trained LLM that already understands general language patterns and training it further on a smaller, specialized dataset. Fine-tuned large language models have undergone this additional training round to become adept at interpreting claims notes, understanding policy terms, flagging compliance issues, and other insurance-related work.  

Think of it this way: a pre-trained LLM is like a new hire who’s an expert communicator, but unfamiliar with insurance. Fine-tuning is the onboarding process for teaching them your company’s terminology, policies, and best practices so that they can perform at their best in your environment.

Speaking the Language of Insurance 

Insurance has specialized terminology, acronyms/abbreviations, and highly nuanced regulatory language. A general-purpose LLM lacking contextual knowledge might misunderstand or misinterpret standard insurance terms (e.g., COI, “admitted/non-admitted”, “retention”), necessitating human rework of data and thus sapping operational efficiency. Insurance fine-tuned models learn to recognize and accurately process our industry’s unique lexicon. 

Accuracy and Efficiency Are Goals—Fine-Tuning Is How to Reach Them

Accuracy is critical whether it’s automating claims triage, supporting underwriters with risk analysis, or answering customer queries. Fine-tuned models deliver more precise results because the best models for insurance are trained on a vast corpus of real examples from insurance workflows, reducing the risk of mistakes and the need for manual corrections.

Adapting to Your Organization’s Needs

While many insurance practices are standardized, every insurer has its own products, processes, and documentation styles. Fine-tuning allows LLMs to be customized to reflect an organization’s unique way of working—specific policy language, claims procedures, or compliance requirements—without starting from scratch.

Maximizing the Value of Data

Insurance companies have access to vast amounts of proprietary data, from claims histories to customer communications. Fine-tuning enables you to use this data, teaching your AI to recognize patterns, flag anomalies, or generate insights that are directly relevant to your business.

 

 

Common Questions About Fine-Tuning and Insurance AI

  1. What is fine-tuning in large language models? 
    Fine-tuning is the process of further training a pre-trained LLM on a smaller, domain-specific dataset—such as insurance claims or underwriting notes—to perform specialized tasks more effectively. 
  2. What are fine-tuned models? 
    These are LLMs that have been adapted through fine-tuning to excel at tasks within a specific field, like insurance, rather than just general language tasks. 
  3. How many examples are needed to fine-tune an LLM? 
    The number can vary, but fine-tuning typically requires fewer examples than training a model from scratch. In many insurance applications, 100 well-chosen examples—such as annotated claims or policy documents—can have a significant effect. 20-30 samples are adequate in known use cases. Quality and relevance are key: the better your examples reflect real insurance tasks, the better your model will perform. 

 

Notes on Annotation

“Annotation” is a critical step in improving AI-powered document processing that provides models with human-reviewed examples, including claims forms, transcripts of customer interactions, geospatial data, and other structured and unstructured documents.  

Benefits of annotation include:

  • Continuous model improvement: Manual annotation provides the most accurate and confident tagging and labeling needed to extract data from a new set of documents. 
  • Data consistency: Annotation reduces output variance, leading to higher confidence for straight-through processing.
  • Faster model training: Annotated datasets accelerate the model training process by providing precise, structured examples.

 

 

Teachable Moments for AI: Getting Started with Fine-Tuning

For insurance leaders in IT, underwriting, and claims, the path to fine-tuning starts with identifying key workflows or pain points where AI can make a difference. The next step is gathering high-quality, representative data—such as claims, policies, and customer interactions—that can be used to teach the model. Finally, work with your organization’s AI experts to fine-tune and evaluate the model’s performance, ensuring it meets organizational needs before deployment. 

 

 

How Fine-Tuning Enables Agentic AI in Insurance

Agentic AI refers to systems that can autonomously pursue complex goals, make decisions, and adapt actions based on changing circumstances. Fine-tuning is foundational to agentic AI. To operate effectively and efficiently, an AI agent must deeply understand insurance-specific language, workflows, and regulatory requirements.  

Fine-tuning large language models with insurance data provides this critical knowledge, allowing the AI to reason, plan, and execute multi-step processes such as claims adjudication or underwriting reviews. By grounding the model in the realities of insurance, fine-tuning plays a critical role in transforming general LLMs into intelligent “agents” for insurance underwriting, claims management, and customer service.

Fine-tuning is the critical step that transforms large language models from generalists into insurance specialists. By teaching AI to understand our industry’s language and workflows, fine-tuning unlocks enhanced accuracy, efficiency, and value from your AI investments. As insurance’s tech transformation powers on, fine-tuned models provide an essential link to improving operational efficiency and delivering better outcomes for insurers and policyholders.

 

Curious how insurers are putting Roots to work? Check out our case studies to see insurance-specific AI in action.