What ISO/IEC 5338 Covers
ISO/IEC 5338:2023, published by ISO/IEC JTC 1/SC 42, defines lifecycle process requirements for AI systems. It extends the established system and software lifecycle standards ISO/IEC/IEEE 15288 and ISO/IEC/IEEE 12207, adapting their process groups to account for characteristics specific to AI, including data-dependent behavior, model training, and continuous learning. The standard applies to organizations of any size or sector that develop, provide, acquire, or use AI systems.
Unlike a management system standard, ISO/IEC 5338 is not certifiable on its own. It describes process outcomes an organization should achieve at each lifecycle stage rather than specific technical mechanisms for achieving them. This distinction matters for governance leaders: conformance to ISO/IEC 5338 demonstrates that a structured process exists, but the standard leaves implementation detail, such as monitoring thresholds or drift detection methods, to the adopting organization or to complementary standards.
Lifecycle Stages and Process Groups
ISO/IEC 5338 organizes AI lifecycle activity into stages including inception, development, deployment, operation, monitoring, and retirement. These map to process groups analogous to ISO/IEC/IEEE 15288's agreement, organizational project-enabling, technical management, and technical processes, adapted for AI-specific tasks.
The standard treats data engineering as a distinct lifecycle process rather than a subset of development, reflecting the centrality of data acquisition, labeling, and quality assurance to AI system behavior. It also recognizes continuous learning and model updating as lifecycle processes separate from traditional software maintenance, since AI systems can change behavior after deployment without a discrete new release. Retirement includes planning for secure data disposal and removal of dependencies from other production systems, which becomes more complex when AI systems are integrated into broader enterprise workflows.
Where the Standard Stops Short
ISO/IEC 5338 does not specify technical thresholds or metrics for acceptable model performance, drift, or safety. It defines that monitoring and risk management processes must exist and produce certain outcomes, but it does not dictate how organizations should log runtime activity, detect anomalous behavior, or enforce action-level controls.
This gap is most pronounced for autonomous or agentic AI components. The standard, as published, does not separately address AI agents that make runtime decisions or invoke external tools. Organizations deploying these systems must map agentic behavior into the existing operation and monitoring processes ISO/IEC 5338 defines, then build the supplementary technical controls, such as tool permission boundaries, action logging, and approval workflows, needed to actually govern that behavior. The standard establishes that this governance must occur; it does not specify the mechanism.
Key Distinction
ISO/IEC 5338 confirms that a governance process must exist at each lifecycle stage. It does not define the runtime mechanisms, such as permission enforcement or action logging, required to prove that process is working for autonomous systems.
Related Standards and How They Fit Together
ISO/IEC 5338 sits within a broader SC 42 portfolio of AI standards. ISO/IEC 42001:2023 establishes requirements for an AI management system (AIMS) and can be used alongside ISO/IEC 5338 to operationalize governance across the lifecycle at an organizational level. ISO/IEC 23894:2023 provides guidance on AI risk management and is referenced as complementary to lifecycle process standards.
Because ISO/IEC 5338 leaves risk management detail to standards like ISO/IEC 23894, governance leaders should determine explicitly which framework governs which decisions. Attempting to treat ISO/IEC 5338 as a complete risk management or certification framework on its own will leave gaps. Enterprises evaluating third-party certification of AI governance should assess ISO/IEC 42001 in parallel, since ISO/IEC 5338 is process guidance rather than a certifiable management system.
Operationalizing Conformance: Runtime Controls for Agentic AI Systems
For enterprises deploying AI agents that call tools, access enterprise systems, or make autonomous decisions, ISO/IEC 5338's operation and monitoring processes provide the process framework but not the enforcement layer. Satisfying the standard's intent for these systems requires runtime governance: establishing agent identity, applying least-privilege permissions, enforcing tool approval workflows, and maintaining audit logs of agent actions.
This is where Trussed AI's runtime governance and security capabilities are relevant. Trussed AI provides runtime policy enforcement, agent permission controls, and audit logging designed to give enterprises the operational evidence and control layer that process standards like ISO/IEC 5338 assume exist but do not define. This does not substitute for the governance program required by the standard; it supports the technical control gap that lifecycle process requirements leave open, particularly for agent-to-agent interactions and MCP-based tool access.