Build self-improving document processing using Amazon Bedrock AgentCore
Amazon Web Services Erik Cordsen ● Covered by 2 sources
AWS shows how to build document processors that teach themselves on new file layouts. That could cut the slog of template upkeep, which is where these systems usually go to die.
Based on reporting by Amazon Web Services, Erik Cordsen — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a different way to handle the paperwork swamp: let agents do the sorting, extracting, checking, and relearning instead of forcing humans to hand-tune templates every time a vendor changes a form.
The company frames the problem with some harsh numbers. IDC says document work costs information workers $19,732 a year each, cuts productivity by 21.3 percent, and can swallow up to 30 percent of staff time. Manual entry still slips 1 to 4 percent of the time. In other words, this is not a corner-case annoyance. It is a tax.
The proposed fix is an agentic IDP system built on Amazon Bedrock AgentCore. Eight specialized agents work across two graph-based workflows. One path handles new document types: an Analyzer pulls text with Amazon Textract, a Matcher looks for similar documents in Amazon S3 Vectors, and if nothing matches, other agents generate instructions, extract fields, validate against rules in Amazon Aurora DSQL, and save what they learned for next time. The second path skips the learning step when the format is already known.
AWS says the system can be deployed in under 20 minutes, though the actual deployment it describes takes about 15 to 20 minutes and creates six CloudFormation stacks. The sample workflow is built around purchase orders, but the company says the same pattern can be adapted for invoices, insurance claims, loan applications, contracts, and other business documents. The big idea is that the model does not just process a document once. It remembers the format.
The demo numbers are the part that will matter to operators. On a first run, a new purchase order from a retailer went through the training workflow and finished in 3 minutes 49 seconds. The next one from the same sender matched stored instructions with a confidence score of 3.946, skipped prompt generation, and finished in 1 minute 30 seconds. AWS also pegs monthly cost at about $58 for 10 training documents and 1,000 processed documents, with model inference making up roughly 85 percent of that bill.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring AI: less chatbot theater, more systems that stop breaking every time a supplier sneezes. The real tell is that AWS is selling memory, validation, and reuse, not just extraction. That is where enterprise AI usually earns its keep, and where the hand-built template crowd has been quietly losing weekends for years.
Read more about this at: Amazon Web Services