Automating Medical Claims for Allianz Bulgaria with Claude Models on AWS

The Challenge

Allianz Bulgaria faced a significant operational bottleneck, manually processing thousands of scanned insurance claims daily. These submissions included highly variable Bulgarian medical and financial documents - such as ambulatory sheets, invoices, receipts, prescriptions, epicrises, and bank statements. The challenge was compounded by poor scan qualities, files containing multiple merged document types, and heavily handwritten doctor notes. Traditional Optical Character Recognition (OCR) software failed to accurately parse these unstructured, variable layouts. Furthermore, any automated solution had to strictly adhere to GDPR and Allianz's rigorous internal security standards, meaning data could not leave the European Union, and the per-document cost needed to remain viable at a high daily scale.

The Solution

Cloud Office engineered a fully automated, serverless pipeline on AWS that processes claims from arrival to structured data with zero manual intervention. Incoming files are automatically unpacked, split into individual pages, and sorted by document type. Leveraging Claude Opus model on AWS Bedrock hosted within EU data centers, the model reads the document images directly - bypassing traditional OCR - and extracts the required fields into structured JSON format. We tailored specific instructions for each document type, accounting for both standardized state medical forms and free-form private clinic templates. Every extracted data point is assigned a confidence score (from 0 to 1), ensuring that doubtful values are flagged for human review. Finally, all data passes through Allianz's internal security screening before entering the core claims system. Access is governed securely by AWS IAM permissions, completely eliminating the need for stored API keys.

The Results

The deployed system successfully processes approximately 3,000 documents per day, achieving an extraction accuracy of over 95%. Less than 1% of the processed documents return an unusable result, and the system's confidence scoring ensures that these outliers are safely flagged for human verification rather than passing incorrect data downstream. This transformation has revolutionized the claims department: staff no longer manually type data and instead focus solely on reviewing flagged anomalies. The platform manages the entire daily volume securely, automatically, and highly cost-effectively.