Organizations are losing an estimated $19,732 annually per information worker due to document-related challenges, impacting productivity by over 21 percent, according to IDC research. A significant contributor to this inefficiency is the constant maintenance required for template-based document processing systems, especially when dealing with the diverse and often-changing layouts of invoices from hundreds of suppliers. Manual data entry further compounds the problem, with error rates ranging from 1 to 4 percent, leading staff to spend up to 30 percent of their time on document-related tasks. Conventional intelligent document processing (IDP) solutions often fall short, demanding extensive upfront configuration for each document format and struggling with layout variations.
The Chenab Times has learned that Amazon Web Services (AWS) has introduced a new capability within Amazon Bedrock AgentCore designed to build agentic IDP systems. This system aims to classify, extract, validate, and learn from business documents with significantly reduced human intervention. Amazon Bedrock AgentCore provides a platform for building and optimizing agents at scale, supporting various frameworks and models. These agents are engineered to automatically handle new document formats such as supplier invoices, purchase orders, insurance claims, and loan applications, promising to deploy a multi-agent IDP system in under 20 minutes and process new document types without manual template configuration. Crucially, the agents can train themselves on new document formats and reuse learned information, streamlining the extraction process for subsequent documents.
Challenges with Traditional Rule-Based Document Processing
Many existing IDP pipelines necessitate manual definition of templates, extraction rules, and validation logic for each distinct document format. This approach proves fragile, as even minor layout changes from a new vendor can cause the pipeline to fail. This creates several core challenges: template rigidity requiring constant manual updates, a lack of inherent learning capabilities leading to repeated extraction errors, and a substantial scaling burden where each new vendor or document variant multiplies maintenance effort and operational costs.
An agentic approach offers a solution by moving away from static templates. AI agents analyze each document, automatically generate extraction instructions, validate the results, and learn from any corrections. Over time, this fosters a knowledge base of document formats, enabling the system to process similar documents with minimal additional configuration.
Solution Architecture: A Collaborative Agent Network
The proposed solution employs a multi-agent architecture comprising eight specialized agents that collaborate across two graph-based workflows running on Amazon Bedrock AgentCore. These workflows are orchestrated using Strands Agents, an open-source SDK for building AI agents.
The Identification workflow begins with an Analyzer agent that uses Amazon Textract to extract text and identify the document type and sender. A Matcher agent then searches for similar documents by leveraging vector similarity within Amazon S3 Vectors. For new document types, a Training workflow is initiated. Here, an Instructions agent generates extraction prompts, an Extractor agent pulls structured data, and a Validator agent verifies the results against predefined schemas and business rules stored in Amazon Aurora DSQL. If validation fails, an Instructions Fixer agent refines the prompts and retries the process. Upon successful validation, a Save Instructions agent persists these instructions for future use.
When a match for a known document type is found, an Automation workflow is triggered. In this scenario, the Extractor agent directly applies existing instructions. If validation fails during this workflow, a Troubleshooter agent attempts to correct the extracted data without altering the original instructions.
Infrastructure and Deployment on AWS
The underlying infrastructure architecture integrates several AWS services. Business documents can be received via Amazon Simple Email Service (Amazon SES) or uploaded directly to Amazon S3. Amazon EventBridge triggers an AWS Lambda function, which in turn creates a tracking job in Amazon DynamoDB and invokes the AgentCore runtime. AgentCore orchestrates document processing through Strands Agents, facilitating secure access to AWS resources via machine-to-machine credentials managed by AgentCore Identity and Amazon Cognito.
Lambda functions are made agent-compatible tools through the Model Context Protocol (MCP) interface of AgentCore Gateway. Agents utilize Amazon Textract for text extraction, S3 Vectors for document classification based on embeddings, and Aurora DSQL for data validation against business rules. Enterprise data can be accessed through MCP servers for validation. A chat-enabled dashboard hosted on Amazon Elastic Container Service (Amazon ECS) allows users to monitor status and troubleshoot issues. Agent performance is monitored through AgentCore Observability and Amazon CloudWatch telemetry.
Deployment and Usage Walkthrough
Deploying the solution infrastructure and web UI to an AWS account requires prerequisites including access to Amazon Bedrock with specific models enabled, AgentCore capabilities, AWS CDK v2, Docker, and appropriate IAM permissions. After cloning a specified GitHub repository and setting up the environment, deployment is initiated via a Python script. This process deploys multiple AWS CloudFormation stacks, with an approximate deployment time of 15–20 minutes.
Upon successful deployment, users receive login credentials and a CloudFront URL. The system allows users to upload documents, initiating the agent workflow. For new document types, the training workflow generates and saves extraction instructions. For known types, the automation workflow uses these stored instructions. Users can review extracted data and validation results in the ‘Review Jobs and Instructions’ tab. The system demonstrates how accuracy improves over time through feedback loops in both the training and processing stages, with the potential for reduced processing times for subsequently encountered document formats.
Customization and Best Practices
To tailor the solution for specific use cases beyond the provided purchase order example, users must define their own document schemas using JSON, build custom validation processes with Lambda functions integrated into Amazon Aurora DSQL, and update agent prompts to reflect their specific document types and extraction requirements. Best practices include starting with representative document samples for training, using business rule validation connected to reference data, monitoring agent performance, incrementally adding document types, and applying Amazon Bedrock Guardrails for sensitive document processing. Reviewing the OWASP Top 10 for LLMs is also recommended for security considerations.
The solution emphasizes a serverless and on-demand operational model, with estimated monthly costs for a moderate workload being approximately $58, largely driven by foundation model inference. Users are instructed to run a cleanup script to remove deployed resources and avoid ongoing charges after they have finished using the system.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
