How to Write a Clinical AI Decommissioning Plan

Why Decommissioning Requires a Formal Plan

No regulation names “clinical AI decommissioning” as a discrete requirement. Instead, the obligations that apply when retiring a clinical AI system are distributed across the HIPAA Security Rule, NIST account management and media sanitization guidance, and the NIST AI Risk Management Framework’s Manage function, which explicitly covers decisions to retire an AI system once risk exceeds acceptable tolerance.

Because no single framework addresses the full lifecycle, healthcare organizations must synthesize these controls into a decommissioning-specific workflow. Treating retirement as an informal IT ticket rather than a governed process is where most gaps appear: orphaned service accounts, unrevoked tool-call permissions, incomplete audit exports, and undocumented data disposition decisions.

Shutdown Sequencing

Shutdown should be staged rather than instantaneous. Separating read access, write access, and tool-call permissions lets operators stop the system from taking new actions while still supporting audit export, data disposition review, and clinical handoff. A single “kill switch” that removes all access at once often leaves the team unable to gather logs or confirm downstream effects before infrastructure is torn down.

Sequence typically moves from disabling outbound tool calls and write paths, to freezing new sessions, to read-only access for export and verification, and only then to full deactivation. Document each stage with a timestamp and an owner so the governance packet reflects what was offline when each control step completed.

Credential and Access Revocation

Revocation is broader than closing a vendor admin console. Clinical AI systems often hold service accounts, API keys, and integration permissions that span the identity provider, EHR layer, and third-party data feeds. Those must be deactivated in a coordinated pass so no path remains open after the system is retired.

• Deactivate service accounts and API keys tied specifically to the AI agent, not shared credentials used elsewhere.
• Coordinate revocation across the identity provider, EHR integration layer, and any third-party data feeds, not solely the AI vendor’s admin console.
• Confirm tool-call and integration permissions can be revoked individually without disrupting other connected clinical systems.
• Verify no downstream system retains a persistent trust relationship or cached credential after revocation.
• Log each revocation action with a timestamp for inclusion in the governance record.

Audit Trail Preservation and Data Retention

Audit log continuity must be handled separately from system shutdown. The HIPAA Security Rule requires audit controls to record and examine activity involving electronic protected health information, and documentation of security actions must be retained for six years from creation or last effective date. This means audit logs generated by the AI system need to be exported or archived before the underlying infrastructure is decommissioned, not assumed to persist automatically once the platform is shut down.

Data disposition decisions should follow NIST SP 800-88 sanitization categories (clear, purge, or destroy) depending on whether the underlying storage will be reused. Training data, inference logs, and PHI-containing records should be evaluated separately, since each may carry different retention or deletion obligations under organizational policy or state law that exceeds the HIPAA six-year minimum.

Governance Sign-off Before Final Deactivation

Final deactivation should not proceed until security, compliance, and clinical leadership have signed off on the completed workstreams. Sign-off confirms that credentials are revoked, audit exports are verified, data disposition decisions are recorded, and clinical teams have an agreed path forward without the system.

The governance record should include owners, timestamps for each major control step, the disposition outcome for each data class, and the approving parties. That packet becomes the durable evidence that retirement was intentional, controlled, and aligned with existing HIPAA, NIST, and AI risk management expectations rather than an ad hoc infrastructure change.