AWS introduced a serverless pipeline for automated PII redaction using Amazon Bedrock Data Automation

Amazon Web Services published a guide to a serverless pipeline for automated detection and redaction of personally identifiable information (PII) in documents and images, built on Amazon Bedrock Data Automation (BDA) in production in combination with AWS Step Functions and AWS Lambda. The solution is intended for batch processing of large volumes of documents, such as medical forms, insurance claims or financial records, where sensitive information must be removed before sharing with third parties.

According to AWS, traditional approaches combining optical character recognition (OCR) with pattern matching or custom ML models run into limitations: they fail on degraded text, struggle to express business logic at the individual field level and require ML expertise when document formats change. The proposed solution instead uses foundation models that can interpret a document page holistically, including its layout, field labels and context, and uses a custom blueprint to define sensitive fields through natural-language instructions without training custom models. The output consists of the field content, a confidence score and bounding box coordinates for further processing.

As an example, AWS describes PII redaction in Attending Physician Statement forms before further processing of insurance claims, where the blueprint defines 37 fields in 9 field groups — sensitive fields include the patient name, date of birth, address and contact details, while the physician name, examination dates and medical notes are not sensitive. AWS recommends tailoring the blueprint to each specific use case, and the schema can be created through the console, CLI or SDK. The rest of the article containing further details is not available.