Home / Rules / ISO/IEC 42005:2025

ISO/IEC 42005:2025: the rules AI agent steps rest on

Edition: ISO/IEC 42005:2025 (first edition, May 2025)
Checked current: 2026-10-07
Official source: https://www.iso.org/standard/42005
Licence: The standard is licensed: each rule is our statement of its clause, cited, never quoted.

Rule	Our statement	Steps that cite it
ISO/IEC 42005:2025 5.11 Approval process	The organization should record every approval, authorization and sign-off decision the assessment process calls for, including authorizing any exceedance of the thresholds set under 5.7 and the final approval of the assessment's results. Points to consider include the situations that call for a management decision (a threshold exceeded, an assessment completed), who takes the decision (which can differ by decision type), and whether approval from outside is needed, for example from regulators or external interested parties.	KC-06 S5 Who approves?
ISO/IEC 42005:2025 5.12 Monitoring and review	The organization should set and document how assessments are monitored and revised: how often (on a schedule, in response to trigger events, or both), who is responsible, and what the monitoring and review involve. Review outputs should include decisions on opportunities for continual improvement and needed changes in how assessments are planned and performed, changes to risk assessment and risk treatment processes, changes to the objectives for using AI, and changes to the thresholds used to rank impacts for sensitive or restricted uses.	KC-06 S6 The change request
ISO/IEC 42005:2025 5.4 Timing, reassessment triggers and triage	When setting up the process the organization should decide when assessments are carried out and in how much depth, and when an earlier assessment can be used again, adapted to a new purpose or updated, and to what extent. Timing can be shaped by applicable law; contractual and professional duties; the organization's own structures, policies, processes, procedures and resources, technology among them; the AI system's risk level (ISO/IEC 23894:2023, 6.3.4 offers guidance); what relevant interested parties, customers among them, expect; and the internal AI life cycle processes (Annex D covers timing alongside other assessments). Reassessment should be considered when something changes: the intended use, including who uses the system; customer expectations; the AI system itself (its data, its complexity or type, its performance); its operating environment; or the context around it (applicable law, contracts, internal policies, relevant interested parties, and the places and sectors the organization works in or plans to). Risk and impact assessments should be done before the triggering change is put into effect. Timing questions include how often assessments happen, at which life cycle stage, how often they are refreshed, what circumstances call for a new or updated assessment, and which other impact assessments are tied to it. The organization should also consider triage: if only high-risk AI systems are to be assessed, the process should state what makes a system high-risk and what triggers an assessment, and a shortened assessment can be used to decide whether a full one is needed.	KC-06 S1 Is an impact assessment required, and when?; KC-06 S6 The change request
ISO/IEC 42005:2025 5.7 Thresholds for sensitive and restricted uses	Documented thresholds, above all for how an AI system is used, are central to the process: the organization should set them for its own context, taking into account applicable law, the expectations of relevant interested parties, the current state of the art, the benefits the AI system brings, cultural, labour and societal norms, and the ethical frameworks for AI that apply. Depending on the kinds and number of thresholds it can add processes such as reviews and approvals; it may decide, for instance, that some uses, or some misuses that can reasonably be foreseen, count as sensitive or are forbidden by its policy, and law and other external requirements can also make a use sensitive. Where an assessment shows that a planned use is sensitive or restricted, the organization should record what the next steps in the overall process are. Automated credit decisions are given as an example of sensitive use because of their potential financial harm to individuals, and a system with a significant effect on human rights, particularly children's rights, whose harms cannot be removed by technical changes alone is given as a case to escalate to higher management, including on whether to redesign it. The organization should also decide how the overall scale of impact is set and how foreseeable impacts are measured, because a use that is not sensitive can still reach a very large population; it should weigh scale when shaping the assessment's components and set criteria for the types of impact an AI system can cause, for example ranking an impact as very severe when it can lead to significant adverse consequences for the organization, groups or societies.	KC-06 S2 How deep?
ISO/IEC 42005:2025 6.7.2 Directly affected interested parties	The organization should identify and document the individuals, groups and societies the system and its intended use may affect, both those who interact with it and those affected without interacting, for example AI users, vulnerable persons, workers and data subjects. Representatives of affected parties should be consulted during the assessment and the results recorded. The scale and subject of consultation should be proportionate to the initial risk level found in triage (more robust assessment and consultation where the risk is high); whatever the risk, consultation helps the organization understand interests and concerns and respond in good faith. Consultation should give timely and relevant information, include different viewpoints, take proper account of vulnerable groups' needs, protect affected parties from retaliation, keep confidentiality and anonymity where needed, and allow individuals to contest decisions and, where needed, seek redress.	KC-06 S3 Who is affected?