AI Governance Standards

ISO/IEC 5338 Verification and Validation Process for AI Systems

ISO/IEC 5338 defines Verification and Validation as a distinct technical process within the AI system life cycle. It requires documented evidence that AI systems, including autonomous agents, meet functional, safety, and performance requirements from initial design through deployment and ongoing operation, not a single pre-release test gate.

Explore Runtime Governance Talk to an Expert

ISO/IEC 5338 V&V at a Glance

Verification

Confirms AI system implementation and model artifacts meet defined requirements before and during deployment.

Validation

Confirms the deployed AI system performs as intended in its actual operational context.

Lifecycle Integration

V&V spans requirements, design, deployment, and operation stages, not a single pre-release checkpoint.

Runtime Validation

Ongoing monitoring and revalidation address model updates, drift, and dynamic agent behavior after deployment.

What ISO/IEC 5338 Verification and Validation Covers

ISO/IEC 5338:2023, published by ISO/IEC JTC 1/SC 42, defines AI system life cycle processes covering organizational, project, technical, and technical management activities. Within this framework, Verification and Validation is identified as a distinct technical process, separate from the general testing activities many engineering teams already perform. The standard adapts life cycle process concepts from ISO/IEC/IEEE 15288, which addresses system life cycle processes, and ISO/IEC/IEEE 12207, which addresses software life cycle processes, extending them to account for characteristics specific to AI systems such as data dependency and probabilistic outputs.

For AI governance leaders, this means V&V cannot be treated as a single quality assurance checkpoint before release. It is positioned as a life cycle process that spans multiple stages, from early requirements definition through deployment and operation, and it requires documented evidence linking requirements, model or version artifacts, and validation results. The full clause-level detail of ISO/IEC 5338 is available only in the published standard text, so organizations implementing V&V programs should confirm specific activity requirements against the official document rather than relying solely on secondary summaries.

How AI System V&V Differs From Traditional Software V&V

Traditional software V&V, as commonly practiced under ISO/IEC/IEEE 12207, generally assumes deterministic, specification-based behavior that can be verified against a fixed set of requirements. A given input produces a predictable output, and passing a test suite at release time is typically sufficient evidence of conformance.

AI systems, including autonomous agents, do not fit this model cleanly. Their behavior depends on training data, model versions, configuration parameters, and, for agents, external tool calls and retrieved inputs that can change independently of the core application code. AI system verification therefore needs to cover both the software implementation and the data and model artifacts that determine runtime behavior. Because model updates, retrieval-augmented inputs, and environmental drift can alter an AI system's behavior after deployment, validation evidence gathered once before release does not remain sufficient on its own. Continuous or periodic revalidation becomes necessary to maintain confidence that a deployed system still meets its original requirements.

Frequently Asked Questions

Is ISO/IEC 5338 a certification standard?

ISO/IEC 5338 defines AI system life cycle processes, including verification and validation, but available evidence does not indicate it functions as a certification scheme comparable to ISO/IEC 42001, which establishes management system requirements. Confirm certification status and applicability directly against the official standard before making compliance claims.

How does ISO/IEC 5338 apply specifically to autonomous AI agents?

Agents that invoke external tools or take independent actions introduce additional verification surface area beyond model output accuracy, including tool-call correctness and action outcomes. This expands what verification and validation activities need to cover compared with a single-model AI system.

How often should AI systems be revalidated after deployment?

Because model updates, retrieval inputs, and environmental drift can change agent behavior after release, revalidation should occur when models, prompts, or retrieval sources change, supported by continuous monitoring rather than a single pre-deployment test gate.

Technical Controls That Support Ongoing Validation

  1. 1

    Behavioral logging and decision traces

    Systems should expose logs and decision traces sufficient to support runtime validation, not only pre-deployment test evidence.

  2. 2

    Model and version separation

    Keeping model or version artifacts separate from application logic allows validation evidence to be tied to a specific model release.

  3. 3

    Tool-call and permission boundaries

    For agents that invoke external tools or take independent actions, defined action-permission boundaries allow individual tool calls and outcomes to be verified and audited, not just model outputs.

  4. 4

    Versioned validation baselines

    Maintaining baselines that can be re-tested when models, prompts, or retrieval sources change supports revalidation without rebuilding evidence from scratch.

  5. 5

    Drift-aware monitoring

    Monitoring should distinguish expected variability from out-of-tolerance drift, since not every behavioral change indicates a validation failure.

Governance Considerations for Enterprise AI V&V Programs

  • Align V&V activities with a broader AI management system framework, such as ISO/IEC 42001, so technical testing is supported by organizational oversight rather than operating in isolation.
  • Confirm the applicability of ISO/IEC 5338 in your jurisdiction and sector before citing it as evidence of regulatory compliance.
  • Validate internal interpretations of the standard against its official published text, since secondary summaries may not reflect exact clause requirements.
  • Assign clear accountability for reviewing runtime drift alerts and approving corrective action, rather than leaving ongoing validation decisions solely to engineering teams.
  • Document how agent identity, least-privilege permissions, and tool-call logging feed into the validation evidence set used for internal and external review.

Support ISO/IEC 5338 Validation With Runtime Governance

Trussed AI provides runtime governance and security controls, including agent identity, least-privilege permissions, tool approval workflows, and audit logging, that generate the evidence enterprises need for ongoing AI system validation.

Explore Runtime Governance