Implementing Financial Document Processing with Amazon Bedrock Data Automation

Amazon Bedrock Data Automation (BDA) is transforming the document processing workflows of financial institutions. A new approach has emerged that can automatically extract data with high accuracy from diverse financial documents, which were difficult to handle with traditional OCR software.

BDA goes beyond simple OCR, providing document understanding using foundation models. It features visual justification and confidence scores for explainability, as well as built-in hallucination mitigation, achieving industry-leading accuracy at low cost compared to foundation models like Anthropic Claude.

(Reference: Process financial documents using Amazon Bedrock Data Automation)

Customizing Extraction Patterns with Blueprint Settings

The core functionality of BDA lies in its blueprint feature, which allows output settings to be customized according to processing needs. A blueprint is a configuration template that defines how to extract data from documents, specifying:

  • The data fields to be extracted
  • The processing method for each field
  • The structure of the output format

You can choose from catalog blueprints or create custom ones, allowing for organization-specific extraction patterns. For different types of financial documents, such as bank statements, W-2 forms, 1099-B tax forms, and vendor contracts, custom blueprints can be created, and output can be generated and verified in the BDA console.

Typically, a single custom blueprint is sufficient for a specific document type, but in cases where workflow requirements differ or document formats change significantly, multiple custom blueprints may be necessary.

(Reference: Process financial documents using Amazon Bedrock Data Automation)

Building Multi-Agent Environments with Amazon Bedrock AgentCore

In a collaboration between Works Human Intelligence (WHI) and the AWS Generative AI Innovation Center (GenAIIC), using Amazon Bedrock AgentCore to build AI agents resulted in a 97% reduction in operational costs while improving business efficiency.

WHI developed and sold an integrated human resources system called “COMPANY” and built two AI agents for commute allowance approval and business support. Initially, they used LangGraph and Amazon ECS, but with the release of Amazon Bedrock AgentCore, they considered migration.

In the new architecture, sub-agents run individually on the AgentCore Runtime. To support multi-tenancy, Amazon DynamoDB and Amazon Cognito are used for tenant management, maintaining the flexibility that WHI built and managed. Slack is used as the entrance for the commute allowance agent, and authentication is performed at invocation before the appropriate sub-agent processes the request.

(Reference: Building AI agents for business support using Amazon Bedrock AgentCore)

Summary

  • The custom blueprint feature of Amazon Bedrock Data Automation enables the construction of automatic extraction systems for organization-specific financial document formats, such as bank statements and W-2 forms.
  • Utilizing the multi-agent environment of Amazon Bedrock AgentCore allows for the automation of routine tasks like commute allowance approval with a 97% cost reduction, as well as the implementation of authentication flows via Slack.
  • By combining the visual justification and confidence score features of BDA, high-accuracy data extraction and explainability from complex document layouts, which were previously difficult with traditional OCR, can be achieved.