HIPAA AI Compliance Ownership: Dividing Responsibility Between Compliance and Platform Teams

What HTI-5 changes, and for whom

As of November 2026, HHS has finalized HTI-5. The final rule was announced on October 26, 2026 and supersedes the proposal published in June 2026. It is a deregulatory rule from ONC that streamlines the Health IT Certification Program by removing and revising certification criteria, in alignment with Executive Order 14192.

The changes apply to certified health IT developers. For example, clinical decision support §170.315(a)(9) is removed and decision support interventions §170.315(b)(11) is revised, both effective on the effective date of the final rule. Several other criteria are removed effective January 1, 2027, and a revision to transitions of care §170.315(b)(1) takes effect January 1, 2030.

This matters for ownership. HTI-5 does not describe runtime controls for AI that a provider organization must operate. The first question for compliance is whether the organization develops certified health IT or only uses it. If it only uses it, the HTI-5 work is to review the changed criteria with the vendor. The obligations attached to the organization's own handling of PHI in AI workflows come from HIPAA.

The HIPAA anchors for AI ownership

The Security Rule requires covered entities and business associates to assess risks to the confidentiality, integrity and availability of ePHI through risk analysis and risk management. That duty includes the risk of third-party access to systems containing ePHI. A reasonable reading is that models, agents, tools and MCP servers that touch PHI belong inside that analysis. Developers of certified health IT often act as business associates, so vendor obligations and the covered entity's own risk analysis overlap and should be documented.

The Security Rule also assumes a designated Security Official. A January 2025 proposal to modify the Security Rule would require compensating controls to be reviewed, documented and signed by that official at least every 12 months, or sooner when the environment or operations change. As of November 2026 it is a proposal, not final law, but it indicates where attestation sits: with a named compliance role.

Technical safeguards sit in 45 CFR 164.312, which includes the audit controls standard at 164.312(b). The gap appears when these anchors are owned separately. Compliance writes a policy that no system enforces. The platform team builds controls that nobody has mapped to a HIPAA requirement. When an auditor asks what the AI was allowed to do and what it did, evidence is assembled by hand after the fact.

A responsibility map by obligation

The table below is an editorial model, not a division prescribed by HHS. It assigns each decision to the function that is best placed to own it and names who supports. Adjust it to your organization's structure, but keep one accountable owner per row.

• AI use policy (permitted PHI, tools and actions): Compliance accountable and defines. Platform team consulted on feasibility.
• Regulatory interpretation, including HTI-5 applicability: Compliance accountable. Platform team informed of changes that affect systems.
• Runtime policy enforcement: Platform team responsible for implementation. Compliance approves the policy being enforced.
• Access, retention and deployment: Platform team responsible. Compliance sets requirements, such as retention periods and data residency.
• Audit controls evidence (164.312(b)): Platform team produces. Compliance verifies the mapping against current regulation text.
• Risk analysis and attestation: Compliance accountable and signs. Platform team supplies the inventory of AI systems and the evidence.

Runtime enforcement and evidence as the shared layer

Policy that is reviewed after the fact only tells you what went wrong. Policy enforced at runtime changes what happens. For AI, that means a decision point in the execution path, where each interaction and tool call can be allowed, redacted, blocked or escalated for human approval before PHI moves.

The same decision point produces the evidence. Each enforced decision can be recorded with the policy that applied, the action taken and the outcome. Compliance gets a record of what the AI was permitted to do and what it did. The platform team gets operational proof that the controls run. Neither team has to reconstruct events from application logs.

Trussed is one example of this pattern. It enforces policy in the execution path across models, agents, tools and MCP servers, and generates continuous audit-ready evidence mapped to HIPAA. It deploys self-managed in the customer cloud with configurable retention and data residency, and it runs in production with healthcare organizations. The ownership model above does not depend on any one product.