AWS has published a serverless workflow that uses Bedrock Data Automation to locate personally identifiable information in scanned documents and then applies redactions at the returned page coordinates. Unlike simple pattern matching, the system uses document layout, labels and context to distinguish fields that should be hidden from similar fields that should remain visible.
In the example for medical forms, a custom blueprint targets patient names, birth dates, addresses and contact details while preserving physician names, appointment dates and clinical notes. Bedrock returns extracted values, confidence scores and bounding boxes. The pipeline converts PDF pages to images, draws black boxes over selected coordinates and orchestrates batch processing with AWS Step Functions and Lambda. Each page is sent separately to keep the model’s scope focused.
The sample blueprint contains 37 fields across nine groups and relies on plain-language instructions. AWS tested 12 documents spanning 47 pages, from clean forms to handwriting and 100-DPI scans. Initial tests found every target at least once but sometimes missed repeated information in narrative text or handwritten notes, so human-labeled ground truth and precision-recall testing remain necessary. The design reduces custom model training, but it is not a compliance guarantee: organizations still need use-case-specific blueprints, quality evaluation and secure handling of both source and redacted files.