Insurance Document Digitization

Records Management Company — Multi-Agent Architecture on AWS

Click any component to expand details

Powered by AWS
Human Input

SRM Worker — Document Scanning

Staff scans medical documents and uploads to cloud folder

Process

  • • Worker receives physical medical documents
  • • Documents scanned using standard scanner
  • • PDFs uploaded to designated S3 folder
  • • Upload triggers automated processing

AWS Services

Amazon S3 S3 Event Notifications
Cloud Storage

Amazon S3 — Document Ingestion

Centralized intake bucket with event-driven trigger

S3 EventBridge

What Happens

  • • Documents land in S3 intake bucket
  • • S3 event notification triggers EventBridge rule
  • • EventBridge invokes Step Functions workflow
  • • Original document preserved for audit trail

AWS Services

Amazon S3 Amazon EventBridge AWS Step Functions

Durable storage with 99.999999999% durability

AGENT ORCHESTRATION
AWS Step Functions State Machine
Agent 1

Vision Agent

Autonomous field extraction via multimodal LLM

Lambda Bedrock
Capabilities
  • • Extracts hospital name, patient ID, insurer
  • • No predefined templates required
  • • Handles varying document layouts
  • • Returns confidence scores per field
AWS Stack
  • • AWS Lambda (Python runtime)
  • • Amazon Bedrock (Claude 3.5 Sonnet)
  • • Amazon Textract (fallback OCR)
Agent 2

Classification Agent

Intelligent routing by hospital & insurance carrier

Lambda SageMaker
Capabilities
  • • Categorizes by hospital (100+ facilities)
  • • Routes by insurance carrier
  • • Learns from correction feedback
  • • Handles new entities gracefully
AWS Stack
  • • AWS Lambda (inference)
  • • Amazon SageMaker (custom model)
  • • Amazon DynamoDB (entity lookup)
Agent 3

Validation Agent

Self-checks with confidence scoring & exception routing

Lambda SNS
Capabilities
  • • Validates extracted fields against rules
  • • Checks confidence thresholds (>85%)
  • • Routes low-confidence to human review
  • • Generates unique barcode identifier
AWS Stack
  • • AWS Lambda (validation logic)
  • • Amazon SNS (exception alerts)
  • • Amazon SQS (human review queue)
Confidence ≥ 85%?
No

Human-in-the-Loop

Exception queue for manual review

  • • Document queued for human review
  • • Reviewer corrects/confirms fields
  • • Corrections feed back to improve models
  • • Re-enters pipeline after approval
Yes

Auto-Approved

Proceeds to orchestration

  • • All fields above confidence threshold
  • • Barcode generated automatically
  • • Metadata package assembled
  • • Continues to ECM delivery
Agent 4

Orchestration Agent

Goal-driven delivery to ECM with retry logic

Lambda Step Functions
Capabilities
  • • Assembles final document package
  • • Generates SOA spreadsheet
  • • Uploads to Dokmee ECM via API
  • • Handles retries on failure
AWS Stack
  • • AWS Lambda (ECM integration)
  • • AWS Step Functions (retry logic)
  • • Amazon RDS (audit log)
Final Output

Dokmee ECM — Document Repository

Indexed, barcoded documents delivered to enterprise content management

Deliverables

  • • Processed PDF with barcode overlay
  • • Extracted metadata (hospital, carrier, patient)
  • • SOA.xlsx spreadsheet for batch
  • • Full audit trail in database

Integration

Documents automatically indexed and searchable in Dokmee ECM. Staff continues using familiar interface with dramatically improved data quality.

Real-Time Operations Dashboard

CloudWatch + QuickSight — Full pipeline visibility

Metrics
  • • Documents processed
  • • Extraction accuracy
  • • Human review rate
AWS Stack
  • • Amazon CloudWatch
  • • AWS X-Ray (tracing)
  • • Amazon QuickSight
Alerts
  • • Pipeline failures
  • • SLA breaches
  • • Anomaly detection
90%
Auto-Processed
0
Templates Needed
98%
Field Accuracy
<10s
Per Document

AWS Services Used

Amazon S3 AWS Lambda AWS Step Functions Amazon Bedrock Amazon SageMaker Amazon Textract Amazon DynamoDB Amazon RDS Amazon EventBridge Amazon SNS/SQS Amazon CloudWatch Amazon QuickSight

Dizzain — Autonomous AI Systems on AWS