ISO/IEC 5338 Human Oversight Across the AI Lifecycle

What ISO/IEC 5338 gives you, and what it leaves to you

ISO/IEC 5338:2023, titled "Information technology, Artificial intelligence, AI system life cycle processes", is a published standard from ISO/IEC JTC 1/SC 42. Its value for oversight is structural. It gives a regulated enterprise a common vocabulary for the stages an AI system passes through, from defining stakeholder needs through design, verification, validation, operation, monitoring and disposal, so that oversight can be located in a stage rather than described in the abstract.

A lifecycle process standard is not an oversight specification. It does not tell you which decisions need a human, how many reviewers are required, or what an approval record should contain. Those are implementation decisions. Treat any oversight checkpoint in your control framework as your own design, justified by your risk assessment and regulatory exposure, and use the lifecycle stages as the scaffolding that keeps those decisions complete and traceable.

This separation matters in audits. An assessor will ask what your organization decided, who owns the decision, and what shows it was carried out. Citing the standard answers none of those questions. Documented checkpoints, named roles and retained records do.

How it sits beside ISO/IEC 42001, the EU AI Act and ISO/IEC 42105

ISO/IEC 42001 is an AI management system standard, so it operates at the level of organizational policy, roles and continual improvement. ISO/IEC 5338 operates at the level of the system and its lifecycle processes. In practice the two meet in your control framework: management-system controls assign accountability and review cadence, and lifecycle processes tell you at which stage each control is exercised. Someone needs to own the reconciliation of the two, or oversight duties will be documented twice and enforced in neither place.

The EU AI Act, Regulation (EU) 2024/1689, includes human oversight provisions in Article 14. These are legal obligations, whereas the ISO/IEC standards are voluntary references. As of November 2026, the Commission's Digital Omnibus on AI, proposed on 19 November 2025, would set fixed deadlines of 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk AI systems embedded in products. It remains a proposal, so plan against current law and treat the proposed dates as a scenario.

A separate SC 42 project, ISO/IEC 42105, covers guidance for human oversight of AI systems. As of November 2026 it is listed as under development. It is reasonable to expect process-level lifecycle language and oversight-specific guidance to be addressed in different documents, so keep your internal mapping modular enough to absorb the guidance when it is published.

Operation stage: oversight at the level of individual interactions

Design reviews and governance committees cannot see what a system does at runtime. Once a model, agent or tool-calling assistant is in production, behavior is determined interaction by interaction: a prompt arrives, data is retrieved, a tool is invoked. Oversight at this stage means deciding, for each such event, whether it is permitted, whether sensitive content needs to be removed, and whether a person must approve before it proceeds.

This is where human-in-the-loop controls become concrete. A policy can allow an action, redact part of the content, block the action, or escalate it for human approval. Escalation is the explicit hand-off to a person, and it should be defined in advance: which actions qualify, who is eligible to approve, how long a request may wait, and what the system does if no one responds. A design that leaves the last question open usually fails open or stalls a business process.

Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalate for human approval. The scope is individual interactions and tool calls. Enforcement happens at runtime, not after the fact, which means the decision and the record of it come from the same event. This covers the operation-stage portion of an oversight program. It does not replace design-time review, governance roles or monitoring cadence, which remain organizational responsibilities.

Evidence that oversight happened in production

Reviewing activity after the fact can show what occurred, but it cannot show what the system was permitted to do at the time. Evidence is stronger when it is produced by the enforcement decision itself: the policy that applied, the action taken, and for escalations, the request, the approver and the time. That is the record an auditor needs to see that a human control operated and was not merely documented.

Trussed generates continuous audit-ready evidence from enforcement, mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST. It runs self-managed in the customer cloud, with configurable retention and data residency, which lets a regulated enterprise decide where evidence lives and how long it is kept. Trussed has run continuously in production at a regulated enterprise for more than 18 months, generating over 30 GB of audit evidence per day.

Be precise about what this evidence covers. It describes policy decisions on interactions and tool calls. It does not describe model lineage, dataset provenance or the full chain of a multi-step agentic workflow, and your oversight program should source those artifacts elsewhere if you need them.