AI-Powered Insurance Claims Automation Platform: A Case Study of Hobbiton Investments, Zambia

The Zambian insurance industry processes vehicle accident claims through manual inspection procedures that are time-consuming, inconsistent, and susceptible to fraudulent manipulation. This dissertation presents the design, implementation, and evaluation of an AI-powered insurance claims automation platform developed in collaboration with Hobbiton Investments, a Zambian insurance technology company. The system employs a YOLOv8m convolutional neural network trained on 2,280 annotated vehicle damage photographs to automatically detect the presence, type, location, and severity of vehicle damage from client-submitted photographs and video frames. The trained model was further fine-tuned on an augmented dataset of 51,051 images to improve performance on visually variable damage classes. The platform integrates a REST API backend, a multi-stakeholder web portal providing role-specific interfaces for policyholders, insurers, and repair shop operators, an automated claims workflow engine, and a seven-check AI-assisted fraud detection module. The final system achieves an overall mean Average Precision at IoU threshold 0.50 (mAP@50) of 0.635 with a precision of 0.824 and recall of 0.577, with the structurally significant damage classes of Broken Part, Cracked, and Dent achieving individual mAP@50 values of 0.876, 0.940, and 0.716 respectively. The complete system is deployed online and accessible from any internet-connected device, reducing damage assessment time from a typical fourteen to forty-five day manual process to under two seconds. The work demonstrates the feasibility of deploying computer vision-based claims automation within the resource constraints of the Zambian insurance market.