ISO/IEC 42001 expects human oversight to be designed into how an AI system is used, assigned to competent people, and evidenced through records. For AI agents, that means deciding which actions a person must approve, naming who approves them, giving them the information and time to judge, and keeping records that show oversight changed outcomes when it needed to.
What does the standard expect of human oversight?
ISO/IEC 42001:2023 is an AI management system standard. It does not contain a single oversight clause; oversight runs through it. In substance:
- The organisation defines roles and responsibilities for AI, including who oversees systems in use.
- Its AI risk assessment and AI system impact assessment identify where human oversight is needed to treat risks and limit impacts on people.
- Controls on the responsible use of AI systems expect the organisation to define processes for use, including human oversight, and to use systems according to their intended use and documentation.
- Competence requirements mean those exercising oversight have the knowledge and training to do it.
- Monitoring and documented information provide evidence that oversight operates.
Our explainer on what human oversight is covers the concept, and our piece on ISO/IEC 42001 controls mapped to agents shows where oversight sits among the annex controls.
Who does it apply to?
Any organisation that adopts ISO/IEC 42001, whether developing, providing or using AI. For a vendor, oversight design is part of what it provides: the features that let customers hold actions and review them. For a deployer, oversight is how it uses the system: who approves, when and with what information.
What does oversight mean when the system is an agent that acts?
Oversight of a model usually means a person reviews its outputs. Oversight of an agent is harder, because agents take many small actions, quickly, and the risk lies in which ones pass without review.
The useful design question is not "is there a human in the loop?" but "which actions does a person approve, and can they really judge them?". We answer it in three layers.
By capability and depth. Decide for each capability whether the agent reads (depth 1), drafts for a person to send (depth 2), executes routine work within bounds (depth 3), or takes actions that move money or are hard to undo (depth 4). Oversight is heaviest at depth 2, where every output waits for a person, and at the edges of depth 3 and 4, where actions outside bounds are held.
By exception. For depth 3 and 4, the person does not approve each action; they approve the bounds, and the system holds anything that falls outside them. Oversight then means reviewing held actions and sampling allowed ones.
By named person. Every held action is routed to a named person, not a shared queue with no owner. Our template for a human oversight procedure for agents gives the fields to define.
Oversight also fails in predictable ways: approvers clearing items in seconds, approving everything, or lacking the context to disagree. A standard-compliant system will detect that, not just record the approvals.
How do you meet the expectation with controls and evidence?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Oversight identified through risk and impact assessment | Each agent capability assessed for where a person must approve | Risk and impact assessments naming oversight points |
| Oversight assigned | Named overseers and deputies for each agent and capability | Role assignments per agent |
| Overseers competent | Training on the agent's purpose, limits and failure modes | Training records |
| Overseers informed | Held actions arrive with the agent's reasoning, sources and what it did not check | Hold records with context attached |
| Oversight operates | Held actions approved or rejected by the named person | Approval logs with identity and time |
| Oversight is real | Review times and rejection rates monitored per overseer | Oversight quality reports |
| Allowed actions sampled | Periodic sample of actions taken without approval | Sampling results |
| Oversight changes when risk changes | Oversight points reviewed on agent change or incident | Review records |
Roles are the foundation; our piece on ISO/IEC 42001 roles and responsibilities for agents covers how to assign them.
What will an auditor ask?
- Where does your risk assessment require human oversight for this agent?
- Who are the named overseers, and how were they trained?
- Show us held actions from last month. Who decided, how long did they take, how many did they reject?
- What information does an overseer see when deciding?
- How do you detect oversight that has become a formality?
- What actions does the agent take without approval, and how are they sampled?
What does the standard not cover?
It does not set legal oversight requirements, such as Article 14 of the EU AI Act for high-risk systems or Article 22 of the GDPR for solely automated decisions, though a good oversight design helps meet both. It does not tell you which actions must be approved; your risk assessment does. And it does not prove the agent performs well; testing and certification do. For background, see what ISO/IEC 42001 is and our guide agent governance is a bound, not a model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme, each action needs a depth and every held verdict names the person who must decide, with the agent's reasoning and sources attached. Every verdict is filed in an append-only record, so review times, rejections and allowed actions can be sampled without reconstruction. A certification run that clears no depth drops the agent to depth 1 the same day.
Questions people ask
Does every agent action need human approval to satisfy ISO/IEC 42001?
No. The standard asks for oversight proportionate to risk. Approving bounds and reviewing exceptions can satisfy it for routine actions, while high-impact actions are held for a person.
Who should be the overseer for an agent?
Someone who understands the business process well enough to disagree with the agent, with the authority to stop it. Usually that is a team lead in the process the agent works in.
How do you show an auditor that oversight is not a rubber stamp?
Show review times, rejection rates and cases where overseers changed the outcome. A record of approvals alone does not prove judgement was exercised.