AI news ·
Amazon Bedrock AgentCore enables self-improving document processing that learns from each new format
Document tasks cost $19,732 per worker yearly and cut productivity 21.3%. This agentic system learns new formats automatically, processing repeat documents in 1 minute 30 seconds.
Document-related tasks cost organizations $19,732 per information worker each year and cut productivity by 21.3 percent, according to IDC research. For teams processing invoices from hundreds of suppliers, layouts that vary by issuer and change without notice break template-based systems, forcing staff to spend up to 30 percent of their time on document work with error rates between 1 and 4 percent.
Conventional intelligent document processing (IDP) tools require extensive upfront configuration for each format. An agentic approach built on Amazon Bedrock AgentCore changes this. Eight specialized AI agents collaborate across two workflows to classify, extract, validate, and learn from business documents automatically, handling new formats without manual template setup.
How the multi-agent system works
The solution runs on two graph-based workflows. When a document arrives, the Analyzer agent extracts text using Amazon Textract and identifies the document type and sender. The Matcher agent then searches for similar documents using vector similarity in Amazon S3 Vectors.
If no match is found, the system triggers a training workflow. The Instructions agent generates extraction prompts, the Extractor agent pulls structured data, and the Validator agent checks results against schemas and business rules in Amazon Aurora DSQL. When validation fails, the Instructions Fixer agent refines the prompts and retries. On success, the Save Instructions agent persists the instructions for reuse.
When a match is found, the automation workflow applies existing instructions directly. If validation fails, the Troubleshooter agent fixes the extracted data without modifying the stored instructions. The entire system deploys in approximately 15 to 20 minutes using AWS CDK.
Training on new document formats
A purchase order from a new retailer illustrates the training workflow. The Matcher agent finds no similar document in the vector store, so the system routes the job through training. The Instructions agent generates extraction prompts from the schema, and the Extractor agent produces structured JSON with line items, quantities, and prices.
The Validator agent then checks the result and fails because the vendor's SKU format does not match the internal format. The Instructions Fixer agent diagnoses the mismatch and adds an explicit SKU transformation rule to the extraction prompt. On the second pass, all line items validate successfully, and the refined prompt is saved to the vector store. The first run completes in 3 minutes and 49 seconds.
Reusing learned instructions
When a second purchase order arrives from the same retailer, the Matcher agent finds the stored instructions with a confidence score of 3.946, above the 3.7 threshold. The automation workflow applies the instructions directly, skipping prompt generation entirely. The SKU transformation rule learned during training applies automatically.
All line items validate on the first pass. End-to-end processing completes in 1 minute and 30 seconds, down from nearly 4 minutes on the initial run. The system improves through two feedback loops: the training loop refines extraction prompts after validation failures, and the processing loop fixes extraction errors while identifying patterns in format variations.
Customizing for your document types
The sample deployment includes a purchase order processing use case with a synthetic dataset and a pre-built JSON schema. To process your own documents, you need to replace the included schemas and validation logic. Define a JSON schema specifying field names, data types, and whether each field is explicitly stated or inferred. Build a validation Lambda function that checks extracted data against your business rules and returns pass/fail results with detailed error messages.
Update the agent prompts to reflect your document types and extraction requirements. Start with 5 to 10 representative samples per document type, including edge cases with format variations. Add document types incrementally by creating a JSON schema and corresponding validation function for each one. The agents handle the rest.
For teams building these systems, structured training on agent architectures can accelerate development. AI Agent Courses cover the patterns used in multi-agent orchestration, and AI Engineering Courses provide deeper technical grounding for customizing extraction and validation pipelines.
Why this matters for operations and development teams
An agentic IDP system removes the template maintenance burden that scales with every new vendor or document variant. Instead of configuring extraction rules for each format, operations teams upload documents and let agents generate, validate, and store instructions autonomously. Development teams can swap models for cost or accuracy tradeoffs through a centralized configuration file, using lower-cost models for classification and stronger models for extraction where needed.
For a workload of 10 training documents and 1,000 processed documents per month, the estimated cost runs approximately $58, with foundation model inference accounting for roughly 85 percent of the total. The serverless architecture means no idle costs. When changing models, AgentCore Evaluations confirms that extraction quality remains acceptable before production rollout.