{
  "id": 7809711,
  "title": "Build a serverless PII redaction pipeline with Amazon Bedrock Data Automation",
  "url": "https://urgent.news/2026/09/16/build-a-serverless-pii-redaction-pipeline-with-amazon-bedrock-data",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-16T15:17:37.000Z",
  "source": {
    "name": "AWS Machine Learning",
    "slug": "aws-machine-learning",
    "url": "https://aws.amazon.com/blogs/machine-learning/build-a-serverless-pii-redaction-pipeline-with-amazon-bedrock-data-automation/"
  },
  "original_language": "en",
  "account": "Organizations processing large volumes of scanned documents daily, such as medical forms, insurance claims, and financial records, often face the challenge of personally identifiable information (PII) redaction before sharing or further processing. Manual redaction is both time-consuming and prone to errors, posing compliance risks. Moreover, redaction demands a precise approach, considering the detection and removal of various sensitive data points on a single page. While traditional methods combine OCR with pattern matching or custom ML models, they encounter challenges with degraded text, lack field-level business logic expressiveness, and require ML expertise for development and model retraining as document formats evolve.\n\nThis article presents a method for automating end-to-end PII detection and redaction at scale on AWS. The approach utilizes Amazon Bedrock Data Automation (BDA), AWS Step Functions, and AWS Lambda to create a serverless batch architecture. This workflow is illustrated in Figure 1, which showcases the integration of generative AI document understanding capabilities of foundation models. These models can interpret documents holistically, including layout, field labels, and context, and can identify the specific entity each piece of information belongs to using clear, plain-language instructions. Amazon Bedrock Data Automation offers a custom blueprint feature, enabling precise extraction of structured information from unstructured documents, images, audio, and video through plain-language instructions. The service delivers extracted field content, confidence scores, and bounding box coordinates for downstream processing. By customizing a blueprint for PII redaction, it becomes an effective detection engine tailored to batch processing needs.\n\nTo illustrate, the focus of this post is on PII redaction for Attending Physician Statements before downstream claim processing. The use case establishes the criteria for redaction in Table 1. The sensitive data includes patient names, dates of birth, home addresses, and contact information, while non-sensitive data comprises physician names, examination dates, office addresses, and medical notes. The fields are located in structured form fields, handwritten text, and multiple instances across documents. The solution for removal involves identifying bounding box coordinates for eligible fields and applying black box redaction in post-processing.\n\nCreating the blueprint involves defining the target schema through the AWS Management Console, AWS Command Line Interface (CLI), or developer SDKs. The console provides a guided walkthrough to generate a blueprint schema from a sample document. This schema should include the fields eligible for redaction, their data type, a brief natural language description, and transformation specifics like date formats. For instance, for the patient's date of birth field which is a date type, the blueprint instructs BDA to extract only specific date instances, excluding dates such as appointments and signatures, even when they appear in different formats on the same page. This ensures that the attending physician's printed name and signature are not included in the redaction scope.\n\nThe example provided depicts a handwritten Attending Physician Statement before and after redaction, demonstrating the effectiveness of the pipeline (Figure 2). Additionally, Figure 3 displays the blueprint within the console, displaying the extracted information for field groups such as Emergency Contact, Family Members, Government IDs, and Insurance Identifiers, which are organized into 9 field groups with a total of 37 fields. This blueprint schema for the specific PII redaction of Attending Physician Statements encompasses a range of details across fields, tailored to the unique requirements of the use case.",
  "summary": "Learn how to automate end-to-end PII detection and redaction from scanned documents at scale using Amazon Bedrock Data Automation with a custom blueprint, AWS Step Functions, and AWS Lambda. A custom blueprint redacts sensitive fields with field-level precision, and a token matching quality check raises recall across degraded and handwritten documents.",
  "key_points": [
    "Amazon Bedrock Data Automation automates PII redaction pipeline on AWS",
    "Utilizes generative AI document understanding for holistic interpretation",
    "Custom blueprint feature enables precise extraction of structured information"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "AWS Machine Learning",
        "title": "Optimizing agent system prompts with Amazon Bedrock AgentCore",
        "url": "https://urgent.news/2026/09/16/optimizing-agent-system-prompts-with-amazon-bedrock-agentcore",
        "published": "2026-09-16T15:47:39.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}