According to the AWS Machine Learning blog post, cowritten by Ricoh engineers, Ricoh's framework built on the AWS GenAI Intelligent Document Processing (IDP) Accelerator reduced customer onboarding time from 4-6 weeks to 2-3 days and decreased engineering hours per deployment from about 80 to under 5, a reduction of over 90%. The solution now processes over 10,000 healthcare documents monthly, with the infrastructure in place to scale to 70,000 documents as client needs grow, and the capacity is projected to grow sevenfold. The post also reports potential annual savings exceeding 1,900 person-hours.
Ricoh USA's healthcare practice serves major health insurance payers, managed care organizations and healthcare providers, processing insurance claims, grievances, appeals and clinical records. Each new customer implementation had required unique development and tuning by specialized engineers, and the workflows demanded 40-60 hours of developer time per customer to set up. Documents ranged from 15 to 50 pages and varied by provider, so template-based extraction proved ineffective. The solution also had to meet HITRUST, HIPAA and SOC II requirements. Jeremy Jacobson, AI Architect, Portfolio Solution Development at Ricoh, says: “We needed to move from bespoke builds to a platform.”
Ricoh selected Processing Pattern 2 of the accelerator, which combines Amazon Textract for OCR with Amazon Bedrock foundation models for classification and extraction, so that documents are classified first and then extracted based on that classification. Documents land in Amazon S3, which triggers AWS Lambda functions that call Amazon Bedrock to determine document types such as claims, appeals, faxes, grievances and prior authorization requests. Results and confidence scores are written to Amazon DynamoDB. Confidence scores come from comparing Amazon Textract and Amazon Bedrock extractions, and fields below customer-defined thresholds, typically 70-85%, are routed to human review queues. Dynamically inserting context into prompts improved extraction accuracy by 15-20% compared with static prompts.
Each customer has its own Amazon EventBridge rule and Amazon SQS queue for multi-tenant isolation, with the queue controlling request rates to Textract and Bedrock, and each deployment is a configurable AWS SAM application deployed as an AWS CloudFormation stack. The IDP-Common library served as the integration layer with Ricoh's existing workflows, and audit logging uses Amazon CloudWatch and AWS CloudTrail. The classification system distinguishes between grievances and appeals, which carry different regulatory timelines, and the architecture is designed to handle traffic spikes of up to 1,000 documents in minutes.
Details