Architecting HIPAA-Compliant AI: A Practical Checklist for Engineering and Procurement Teams

Introduction

The fastest way to derail a healthcare AI project is to discover a compliance problem after the demo. A team wires a language model into a workflow, someone pastes real patient text into a prompt, and only later does anyone ask where that text went, who can see it and what the contract says.

HIPAA-compliant AI development is mostly ordinary secure engineering applied carefully to new places where protected health information can leak: prompts, logs, embeddings, fine-tuning datasets and third-party APIs. This checklist is written for engineering leads and buyers who need to make concrete decisions. It is educational and not legal advice, so involve your privacy officer and counsel.

What "HIPAA-Compliant" Actually Means

No government body certifies a piece of software as "HIPAA compliant." Compliance describes how covered entities and their business associates handle protected health information (PHI) under the Privacy, Security and Breach Notification Rules. A tool can support compliance, but the obligations sit with the organizations using it and the vendors handling PHI on their behalf.

So the first questions are: is the data PHI; are we a covered entity or business associate, or building for one; and could we avoid PHI by using properly de-identified data for a given task.

Business Associate Agreements Come First

If a vendor creates, receives, maintains or transmits PHI on your behalf, a business associate agreement (BAA) is typically required. HHS guidance on cloud computing addresses a common misconception: a cloud provider offering only "no-view" services, meaning it cannot see the data because you hold the encryption keys, still qualifies as a business associate and must comply with applicable breach notification requirements.

For AI work, apply this to every component that can touch PHI: cloud hosting, model APIs, transcription services, monitoring tools, error-tracking platforms and support vendors. Confirm in writing whether data is retained, for how long, and whether it is used to train or improve models.

Where PHI Hides in an AI System

Prompts and completions; application and error logs; vector databases and embeddings; fine-tuning and evaluation datasets; caches, queues and analytics tools; and developer laptops and test environments.

The Engineering Checklist

• Data governance: Classify data and map every data flow involving protected health information (PHI). Apply the minimum necessary principle and define retention and deletion policies for prompts, outputs, logs and datasets.
• Access and identity: Enforce role-based access control (RBAC) and the principle of least privilege. Require multi-factor authentication (MFA) for systems that access electronic protected health information (ePHI). Maintain individual user accounts and avoid shared credentials.
• Encryption and network controls: Encrypt ePHI at rest and in transit. Isolate PHI workloads in private networks, restrict unnecessary outbound traffic and manage encryption keys and secrets through a dedicated service.
• Audit and monitoring: Log access to PHI and relevant AI system activity while preventing logs from unnecessarily exposing PHI. Review logs regularly, alert on unusual access patterns and track model and prompt versions to maintain traceability.
• AI-specific safeguards: Filter or redact identifiers before sending text to external services wherever feasible. Test for prompt injection and data leakage, separate tenant data in multi-customer systems, and maintain human review wherever AI outputs could affect patient care or billing decisions.
• Vendor and lifecycle management: Maintain an inventory of vendors and subprocessors that handle PHI. Document a risk analysis, assess third-party security practices and prepare and rehearse an incident response plan.

De-identification: An Option, with Caveats

HIPAA recognizes two de-identification methods, Expert Determination and Safe Harbor, described in HHS guidance. Safe Harbor requires removing 18 categories of identifiers and having no actual knowledge that remaining information could identify a person. Expert Determination relies on a qualified statistical expert documenting that re-identification risk is very small. HHS acknowledges that neither method eliminates all risk of re-identification.

For AI teams, this matters because free text is hard to de-identify reliably. Treat automated de-identification as a control to validate, not a switch that removes obligations, and have your privacy officer sign off on the method.

Regulatory Outlook

In January 2025, HHS proposed a major update to the HIPAA Security Rule that would make several currently "addressable" safeguards, such as encryption and multi factor authentication, mandatory, and would add requirements like asset inventories and regular testing. As of late September 2026, secondary sources report the rule is still not final, with a target date that has slipped to around mid-2027. That timeline can change, so verify the current status with HHS or the Federal Register directly.

Questions to Ask a Development Partner

• Will you sign a Business Associate Agreement (BAA), and which of your subprocessors would handle PHI?
• Where will data be hosted and processed?
• How do you keep PHI out of logs, prompts and test environments?
• How do you handle model providers' data retention and training policies?
• What access controls, audit logging and incident response processes do you follow?
• Who owns the source code, the data and the trained artifacts?
• How will compliance evidence be documented for our audits?

Bottom Line

Building AI for healthcare does not require exotic security. It requires disciplined attention to where PHI flows, contracts that match those flows and controls you can demonstrate. Start with a data map and a vendor inventory, involve compliance early and design so that models see only what they need.

This article is educational and is not legal advice. Consult qualified counsel and your privacy officer for your situation.

Suggested Internal Links

• Cloud Solutions
• DevOps and Secure Delivery Pipelines
• Healthcare AI and Software Development

Sources / References

• HHS, "Guidance on HIPAA and Cloud Computing."
• HHS OCR, "Guidance Regarding Methods for De-identification of Protected Health Information."
• HIPAA Journal, "De-identification of Protected Health Information: 2026 Update." (Secondary source)
• Security Rule NPRM status (secondary sources; verify against HHS/Federal Register):
• Compliancy Group, Proposed HIPAA Security Rule Update and Risk Analysis Requirements.

CTA

If you're scoping an AI or software project that will handle patient data, an early architecture review can save rework later. Healthcare is one of the industries Techverse Solutions serves, and we're happy to walk through your data flows, hosting options and vendor questions on a scoping call. Book a consultation to get started.