ISO 42001 Lead Implementer cheat sheet

An AI management system (AIMS) is the set of interrelated policies, objectives, processes and responsibilities an organization uses to direct and control its development and use of AI; ISO/IEC 42001 specifies requirements for such a system.

ISO/IEC 42001 follows the Harmonized Structure (Annex SL): scope, normative references, terms, context, leadership, planning, support, operation, performance evaluation, improvement - enabling integration with ISO 9001, ISO/IEC 27001 and ISO 22301.

Core AI-specific concepts the standard introduces: AI system, AI system impact, AI system lifecycle, responsible AI, risk and impact assessment, and the AI system owner and AI system impact levels.

An AI system is an engineered system that generates outputs such as predictions, content, recommendations or decisions for a given set of human-defined objectives; it operates with varying levels of autonomy and adaptivity.

Risk in ISO/IEC 42001 is uncertainty affecting objectives (likelihood x consequence) and is assessed against the organization's own criteria; impact is the effect on individuals, groups, society and the environment, assessed separately.

The standard is designed to be technology-neutral and applicable to any organization that provides or uses AI-based products or services, regardless of size, type or sector.

AIMS scope must be documented and can be bounded by organizational units, products, services or geographic locations; excluded activities still must not undermine the system's integrity.

Accountability is a first-order principle: every AI system must have a named owner with defined authority, and roles for AI risk, compliance and data governance must be assigned and communicated.

Certification against ISO/IEC 42001 is a third-party conformity assessment performed by an accredited certification body; it attests to the management system, not to the safety of any individual AI model.

Clause 4 requires determining external and internal issues, identifying interested parties and their relevant requirements, and defining the AIMS scope and processes; these are inputs, not documentation exercises.

Clause 5 requires top management to demonstrate leadership by establishing the AI policy, ensuring resources, integrating AIMS requirements into business processes, and assigning AIMS responsibility and authority.

The AI policy must be appropriate to the organization's purpose, include commitments to applicable requirements and continual improvement, provide a framework for AI objectives, and be communicated and available.

Clause 6 requires addressing risks and opportunities, AI risk assessment, AI risk treatment, an AI system impact assessment, and documented AI objectives with plans to achieve them.

Clause 7 covers resources, competence, awareness, communication and documented information; competence records are required for persons doing work that affects AI performance or conformity.

Clause 8 requires operational planning and control, AI risk assessment and treatment performed and retained as documented information, and control over externally provided processes, products or services.

Clause 9 requires monitoring, measurement, analysis and evaluation, internal audit at planned intervals, and management review with defined inputs and outputs.

Clause 10 requires handling nonconformities and corrective actions and the continual improvement of the AIMS's suitability, adequacy and effectiveness.

Annex A controls are grouped by theme (policies, internal organization, resources, impact assessment, AI system lifecycle, data, information for interested parties, use of AI systems, third-party relationships); Annex B provides implementation guidance and Annex C potential objectives and risk sources.

A defensible sequence is: secure sponsorship and scope, gap analysis against Clauses 4-10 and Annex A, risk and impact assessment, control selection and treatment plan, implementation, internal audit, management review, certification audit.

The gap analysis output is a register of missing controls and missing documented information, each with an owner and a target date; it becomes the project plan baseline.

AI risk criteria must be defined before assessment: acceptable risk thresholds, impact levels, assessment frequency, and criteria for escalating to top management.

An AI system impact assessment considers effects on individuals, identifiable groups, society and the environment, and is required for AI systems the organization develops, provides or uses.

Control selection follows a risk-based approach: Annex A controls are the reference set, but applicability is decided by the organization and justifications for exclusions must be retained as documented information.

Scope decisions must be consistent with the risk assessment: an AI system excluded from scope but still used by the organization creates a boundary risk that the certification body will probe.

Implementation planning should account for realistic timelines: gap closure, evidence accumulation over at least one full internal audit and management review cycle, and certification body scheduling lead times.

A Statement of Applicability (or equivalent documented justification) records each Annex A control, whether it is applicable, and the reason; it is a primary audit entry point.

Resource planning must cover AI competence (data science, model validation, legal, ethics), tooling for logging and model inventory, and budget for external audit and surveillance cycles.

Maintain an AI system inventory as a controlled record: each system's purpose, owner, lifecycle stage, data sources, model type, users, and risk and impact rating.

Implement controls across the AI lifecycle: requirements and design, data collection and preparation, model building and validation, deployment, operation and monitoring, and decommissioning.

Data governance controls include provenance and lineage records, quality criteria, lawful basis and consent where applicable, bias examination, and retention and disposal rules.

AI risk treatment options mirror the risk discipline: avoid, mitigate through controls, transfer (for example contractual or insurance), or accept with documented authority.

Operational controls to evidence: change management for models and prompts, version control of datasets and models, environment separation, access control, logging and traceability of AI outputs.

Human oversight must be designed, not declared: define which decisions require human review, what information reviewers receive, and how override and escalation work.

Third-party and supplier controls cover due diligence on foundation models and AI services, contractual allocation of responsibilities, transparency on model changes, and monitoring of supplier performance.

Incident and nonconformity handling must include AI-specific events: harmful or biased output, model drift, data leakage, prompt injection, and misuse by users.

Documented information required includes the AI policy, scope, risk and impact assessments, treatment plans, SoA, competence records, audit results and management review outputs.

Monitoring and measurement must determine what is measured, the methods, when measurement occurs, who analyzes results, and when results are evaluated - all defined in advance.

Measure both management-system performance (audit findings closed on time, training completion, review cycle adherence) and AI system performance (accuracy, drift, bias indicators, incident rates).

Model drift monitoring compares production behavior against a defined baseline; triggers for retraining or decommissioning must be pre-agreed, not decided ad hoc.

Internal audit must be planned, objective and impartial; auditors must not audit their own work, and the programme considers the importance of processes and prior results.

Management review inputs include audit results, nonconformities and corrective actions, monitoring and measurement results, risk and impact assessment status, and interested-party feedback; outputs include improvement decisions and resource needs.

Evidence for audit must be retrievable and dated: logs, tickets, meeting minutes, training records and assessment reports, not retrospective reconstructions.

Key performance indicators should be traceable to AI objectives; an objective without a metric, target date and accountable owner is not auditable.

Nonconformities are graded by the certification body as major (systemic or unaddressed requirement) or minor (isolated lapse); both require root-cause analysis and corrective action, and majors typically require a follow-up audit.

Continual improvement uses the Clause 10 loop: identify nonconformity or opportunity, correct, determine root cause, act on cause, review effectiveness, and update risk and impact assessments as needed.

Corrective action must address cause, not symptom; evidence includes the root-cause method used, the action taken, and verification that the nonconformity did not recur.

Certification follows a two-stage audit: Stage 1 reviews readiness and documentation, Stage 2 evaluates implementation and effectiveness; certification is valid for three years with surveillance audits.

Stage 1 findings are typically documentation gaps and scope questions; Stage 2 findings are evidence and effectiveness failures, so the organization must be operating the AIMS, not just documenting it.

The certification body samples; therefore every Annex A control claimed in the SoA must have retrievable evidence available at short notice, including for AI systems not sampled in the previous cycle.

Recertification before the three-year expiry repeats a full audit; surveillance audits at least annually check specified areas and continued conformity.

Nonconformity closure timelines are set by the certification body; unresolved major nonconformities can suspend or withdraw certification.

Improvement opportunities should feed back into risk treatment and objectives, closing the loop between performance evaluation and planning.

Preparation activities that reduce audit risk: a full internal audit covering all clauses and applicable Annex A controls, a management review with complete inputs, and a dry-run evidence walkthrough per AI system.