Architecting HIPAA-compliant AI agents to safeguard health data with AWS

This article presents a reference AWS architecture for building HIPAA-compliant AI agents in healthcare that protect patient data across new surfaces introduced by agentic AI workflows.

• AI agents multiply PHI exposure pathways through prompts, tool calls, memory, reasoning, and logs beyond traditional database access patterns
• Reference architecture organizes 11 technical and 5 operational controls across four trust zones: customer application, API boundary, agent platform, and observability
• Authorization controls enforce field-level and record-level access restrictions, preventing agents from returning unnecessary PHI like SSNs or claims history
• Input/Output sanitization uses AWS WAF, Amazon Comprehend, and Bedrock Guardrails to keep PHI out of model reasoning and responses
• Memory protection encrypts agent context with customer-managed KMS keys, applies per-session isolation, and auto-expires after 30 days
• Network isolation routes all PHI-bearing traffic through VPC endpoints and PrivateLink, keeping data off public internet
• Immutable audit trails use S3 Object Lock with 6-year retention; Amazon Macie backstop scanning catches missed PHI in logs
• Consequential actions like prior authorization decisions require explicit human approval before execution, closing the judgment gap
• Implementation follows three phases: foundational access control and audit, edge hardening before member launch, and ongoing detection and monitoring

The architecture enables healthcare organizations to deploy AI agents while maintaining HIPAA compliance by protecting PHI across all new agentic surfaces.