Incident Cause Classification in Insurance claims using Generative AI

Thumbnail Image

Files

master_Uzair_Mohammad_2025.pdf (4.83 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-05-20

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

81

Series

Abstract

Automation plays a critical role in modern insurance operations, enabling companies like OP-Pohjola to process claims more rapidly, reduce manual workloads, and improve customer satisfaction. However, claim automation rates are often reduced when essential categorical data are left unspecified during claim notification. Customers frequently struggle to choose from predefined incident categories, submitting instead free-text descriptions of their incidents. These unstructured inputs disrupt downstream automation workflows, forcing claims into manual handling, even when they would otherwise qualify for automated processing. This thesis explores the potential of large-scale generative language models, particularly GPT-4o, to address this challenge. The study introduces a solution that classifies missing incident cause categories by analyzing free-text descriptions and other structured claim data. The system is planned to be integrated as an API within OP-Pohjola’s internal claims handling platform, where it will be automatically triggered when claims lack the necessary categorization. By mapping user-inputted text to predefined category labels, the solution restores automation eligibility to such claims. The experimental results presented here are promising. GPT-4o demonstrates the ability to reliably and accurately infer incident categories from a wide range of diverse and informal claim descriptions without the need for domain-specific training or large annotated datasets. By fine-tuning prompts and restructuring the classification taxonomy, the system achieved remarkable improvements in classification accuracy, while maintaining impressively low false positive rates. The results also show a direct and measurable increase in simulated automation rates, streamlining claim processing and reducing manual intervention. This approach offers a timely solution to the growing demand for efficient and scalable automation in insurance. It bridges the gap between unstructured user input and the structured data requirements of automated workflows, providing a lightweight alternative to traditional machine learning methods. The findings not only demonstrate the viability of generative models in practical insurance applications but also underscore their broader potential to improve automation rates in various industries.

Description

Supervisor

Kannala, Juho

Thesis advisor

Valli, Jaakko

Keywords

insurance, claims, automation, generative AI, text classification, GPT-4o

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202506174914

Collections

[dipl] Perustieteiden korkeakoulu / SCI