Amazon Web Services (AWS) published a deployment guide for an agentic intelligent document processing (IDP) system built with Amazon Bedrock AgentCore. The system uses eight specialized agents across two graph-based workflows to classify, extract and validate business documents, then retain instructions learned from corrections. AWS says organizations can deploy the multi-agent system in under 20 minutes and process new document types without manually configuring templates.

The guide cites IDC research estimating that document-related challenges cost organizations $19,732 per information worker annually and reduce productivity by 21.3%. Staff may spend up to 30% of their time on document tasks, while manual data entry error rates range from 1–4%. The guide says conventional IDP tools need upfront configuration for each format and can fail when layouts change across suppliers.

In the identification workflow, an Analyzer agent extracts text with Amazon Textract and identifies the sender and document type. A Matcher agent searches Amazon S3 Vectors for similar documents using vector similarity. When it finds no match, the training workflow has an Instructions agent generate extraction prompts, an Extractor agent produce structured data, and a Validator agent check it against schemas and business rules in Amazon Aurora DSQL. If validation fails, an Instructions Fixer revises the prompts; after a successful check, a Save Instructions agent stores them for reuse.

For a known document type, the automation workflow reuses stored instructions. An Extractor processes the document, and a Troubleshooter can correct extracted data if validation fails without changing those instructions. The guide’s example uses a purchase order from Two Wheel Store: the first run found that vendor SKU SB-004 did not match Acme Bikes’ internal format AB-SB-004. The Instructions Fixer added a rule to prefix vendor SKUs with AB-, after which seven line items passed validation. The first run took 3 minutes 49 seconds.

For a second order from the same retailer, the Matcher found the stored instructions with a confidence score of 3.946 against a 3.7 threshold. The learned rule converted EL-002 to AB-EL-002, HB-071 to AB-HB-071 and PD-002 to AB-PD-002; all three line items passed on the first attempt, with an order total of $5,979.19. Processing took 1 minute 30 seconds. AWS describes two feedback loops: saving prompt refinements after validation failures and using Troubleshooter analyses to identify patterns in document variations.

The example deployment uses six AWS CloudFormation stacks and takes approximately 15–20 minutes. It requires an AWS account, enabled Amazon Nova and Anthropic Claude models in Amazon Bedrock, access to AgentCore, S3 Vectors and Aurora DSQL in a supported region, AWS CDK v2, Docker, Python 3.11+, Node.js 18+ and permissions to create the listed AWS resources. Documents can arrive through Amazon Simple Email Service or direct S3 upload; EventBridge triggers Lambda, which creates a DynamoDB tracking job and invokes AgentCore runtime. The guide says agents access AWS resources through AgentCore Identity and Amazon Cognito, and users can monitor jobs through a chat-enabled dashboard and AgentCore Observability with CloudWatch telemetry.

For 10 training documents and 1,000 processed documents per month in US East (N. Virginia), AWS estimates a cost of approximately $58 per month. Foundation model inference accounts for approximately 85% of that total, or roughly $0.055 per document; actual cost depends largely on the models assigned to each agent. The included use case is a synthetic bike company purchase order workflow. Organizations adapting it must replace the sample schemas, validation logic and prompts with versions for their document types, and AWS recommends starting with 5–10 representative documents per type, including edge cases. The guide also recommends business-rule validation, performance monitoring, incremental addition of document types and guardrails for production.

Amazon Web Services (AWS) published a deployment guide for an agentic intelligent document processing (IDP) system built with Amazon Bedrock AgentCore. The system uses eight specialized agents across two graph-based workflows to classify, extract and validate business documents, then retain instructions learned from corrections. AWS says organizations can deploy the multi-agent system in under 20 minutes and process new document types without manually configuring templates.

The guide cites IDC research estimating that document-related challenges cost organizations $19,732 per information worker annually and reduce productivity by 21.3%. Staff may spend up to 30% of their time on document tasks, while manual data entry error rates range from 1–4%. The guide says conventional IDP tools need upfront configuration for each format and can fail when layouts change across suppliers.

In the identification workflow, an Analyzer agent extracts text with Amazon Textract and identifies the sender and document type. A Matcher agent searches Amazon S3 Vectors for similar documents using vector similarity. When it finds no match, the training workflow has an Instructions agent generate extraction prompts, an Extractor agent produce structured data, and a Validator agent check it against schemas and business rules in Amazon Aurora DSQL. If validation fails, an Instructions Fixer revises the prompts; after a successful check, a Save Instructions agent stores them for reuse.

For a known document type, the automation workflow reuses stored instructions. An Extractor processes the document, and a Troubleshooter can correct extracted data if validation fails without changing those instructions. The guide’s example uses a purchase order from Two Wheel Store: the first run found that vendor SKU SB-004 did not match Acme Bikes’ internal format AB-SB-004. The Instructions Fixer added a rule to prefix vendor SKUs with AB-, after which seven line items passed validation. The first run took 3 minutes 49 seconds.

For a second order from the same retailer, the Matcher found the stored instructions with a confidence score of 3.946 against a 3.7 threshold. The learned rule converted EL-002 to AB-EL-002, HB-071 to AB-HB-071 and PD-002 to AB-PD-002; all three line items passed on the first attempt, with an order total of $5,979.19. Processing took 1 minute 30 seconds. AWS describes two feedback loops: saving prompt refinements after validation failures and using Troubleshooter analyses to identify patterns in document variations.

The example deployment uses six AWS CloudFormation stacks and takes approximately 15–20 minutes. It requires an AWS account, enabled Amazon Nova and Anthropic Claude models in Amazon Bedrock, access to AgentCore, S3 Vectors and Aurora DSQL in a supported region, AWS CDK v2, Docker, Python 3.11+, Node.js 18+ and permissions to create the listed AWS resources. Documents can arrive through Amazon Simple Email Service or direct S3 upload; EventBridge triggers Lambda, which creates a DynamoDB tracking job and invokes AgentCore runtime. The guide says agents access AWS resources through AgentCore Identity and Amazon Cognito, and users can monitor jobs through a chat-enabled dashboard and AgentCore Observability with CloudWatch telemetry.

For 10 training documents and 1,000 processed documents per month in US East (N. Virginia), AWS estimates a cost of approximately $58 per month. Foundation model inference accounts for approximately 85% of that total, or roughly $0.055 per document; actual cost depends largely on the models assigned to each agent. The included use case is a synthetic bike company purchase order workflow. Organizations adapting it must replace the sample schemas, validation logic and prompts with versions for their document types, and AWS recommends starting with 5–10 representative documents per type, including edge cases. The guide also recommends business-rule validation, performance monitoring, incremental addition of document types and guardrails for production.