Intelligent Document Processing for a Government Health Services Provider

A national health services organization was manually processing thousands of member-submitted documents daily — IDs, medical records, income verification, and enrollment forms. A team of 40+ reviewers handled classification, data extraction, and validation, creating a multi-day backlog that delayed member onboarding. We built an AI-powered document processing pipeline that achieves 92% straight-through processing, reduced manual review by 80%, and cut member onboarding time from days to minutes.

The Challenge

Members enrolling in state health programs submit an average of 4–6 documents each: government IDs, proof of income, medical records, and program-specific forms. The organization received 8,000+ documents daily across all state programs. Each document required manual classification (what type is it?), data extraction (pulling key fields), validation (does the data match the application?), and routing (which team handles exceptions?). A team of 40+ reviewers processed documents with a 3–5 day backlog. Quality was inconsistent — error rates varied from 4% to 12% depending on the reviewer and document complexity. Members calling to check on application status created additional support load. The backlog grew during enrollment periods, causing compliance risk for processing deadlines.

Our Approach

Document Analysis & Model Design — Weeks 1–3

Analyzed 28 document types across all state programs, mapped extraction requirements, defined accuracy targets, and designed the AI pipeline architecture.

• Document taxonomy: 28 types across enrollment, verification, medical, and identity categories
• Field extraction mapping: 120+ fields across all document types
• Accuracy and confidence threshold design per document type
• Human-in-the-loop workflow design for low-confidence predictions
• Training data preparation from 50K+ historical documents

AI Pipeline Development — Weeks 4–8

Built the end-to-end processing pipeline: Textract for OCR and structure extraction, Bedrock for intelligent classification and context understanding, and Step Functions for orchestration.

• Amazon Textract for OCR, table extraction, and form field detection
• AWS Bedrock (Claude) for document classification and semantic validation
• Custom extraction models for state-specific form layouts
• Step Functions orchestrating the full pipeline: ingest → classify → extract → validate → route
• Confidence scoring with automatic routing to human review below threshold
• DynamoDB for processing state and audit trail

Validation, Rules & Integration — Weeks 9–10

Implemented business rule validation, cross-document consistency checks, and integration with the member enrollment system for automated application updates.

• Business rule engine: income thresholds, date validity, ID format checks
• Cross-document validation: does the name on the ID match the application?
• Integration with member enrollment system via secure APIs
• Exception routing: ambiguous documents sent to specialized review queues
• PII handling: encryption at rest and in transit, access logging, auto-redaction in logs

Testing, Tuning & Production — Weeks 11–12

Ran the pipeline against historical documents to validate accuracy, tuned confidence thresholds to balance automation with accuracy, and deployed to production with monitoring.

• Validation against 10K historical documents with known correct outcomes
• Confidence threshold tuning: optimized for 92% automation at 96.5% accuracy
• A/B comparison: AI extraction vs. historical human extraction for quality baseline
• Production deployment with real-time accuracy monitoring and drift detection
• Reviewer dashboard for human-in-the-loop queue management

Outcomes

Straight-Through Processing

92% of documents are now processed end-to-end without human intervention — classified, extracted, validated, and routed automatically within minutes of submission.

Processing Speed

Member document processing dropped from a 3–5 day backlog to minutes. Members enrolling online see real-time document status instead of waiting for manual review.

Quality Improvement

AI extraction achieves 96.5% accuracy — higher than the 88–96% range from manual reviewers. Consistency is uniform regardless of volume, time of day, or document complexity.

Team Redeployment

The review team was reduced from 40+ to 8 specialized reviewers handling complex exceptions. Remaining staff were redeployed to member support and program operations roles.