What ISO/IEC 5338 Establishes for AI System Lifecycles
ISO/IEC 5338:2023, published by ISO/IEC JTC 1/SC 42, is an international standard covering AI system life cycle processes. Like other lifecycle standards in the systems and software engineering family, it organizes work into process groups spanning inception, development, deployment, operation, and retirement. Retirement and decommissioning represents the closing stage of that lifecycle: the point at which an AI system or agent is formally removed from active use. Because the standard's exact clause language for this stage was not available for direct citation in this guide, governance teams should confirm specific deliverables and terminology against the purchased standard text before using it in a formal compliance claim. What follows is guidance on implementing that closing stage in a way consistent with how lifecycle standards of this type are generally structured.
Why the Decommissioning Stage Is a Distinct Risk Point
AI agents accumulate a wider set of technical artifacts over their operating life than most conventional applications. These typically include service credentials, API keys, OAuth tokens, tool-call permissions, connector configurations to enterprise systems, and cached data. Each of these artifacts represents a standing access path. When an agent is retired informally, by simply disabling a schedule or removing it from a workflow, these artifacts frequently remain active. The result is orphaned access: credentials that still authenticate, permissions that still authorize tool calls, and no record confirming that access was ever reviewed or removed. This gap is what creates security exposure and, separately, an audit gap, since there is no documentation showing the system was retired in a controlled way.
The Role of Runtime Policy Enforcement and Agent Identity
Clean revocation depends heavily on how an AI agent's identity and permissions were architected in the first place. When credentials are embedded directly in agent code or configuration, decommissioning requires locating and removing them individually across every system they touch, which is slow and error-prone. Centralized agent identity management addresses this by giving each agent a managed identity that can be revoked in a single action rather than hunted down across systems. Runtime policy enforcement layers, which mediate tool-call and API access at execution time, extend this benefit further: permissions can be withdrawn centrally without modifying the underlying model or application code, and enforcement takes effect immediately rather than depending on every downstream system independently honoring a revocation request. Audit logging that is decoupled from the agent's runtime, so logs persist after the agent itself is retired, is what supports the retention requirement described above. Trussed AI's runtime governance capabilities, including agent identity management, runtime policy enforcement, and audit logging, are one implementation mechanism for satisfying these operational requirements, though the architectural pattern itself applies regardless of which enforcement layer an organization uses.