ISO/IEC 42001 — AI management system

TL;DR: ISO/IEC 42001:2023 (published December 2023) is the first certifiable management system standard for AI. Same family as ISO 27001 (security) and ISO 9001 (quality) — Plan-Do-Check-Act, leadership accountability, risk-based — but scoped to organisations that develop, deploy, or use AI systems. Driver: regulatory pressure (EU AI Act, NIST AI RMF, US executive order 14110) plus enterprise need for assurance of supplier AI systems.

What it is

ISO 42001 establishes requirements for an AI Management System (AIMS). Like ISO 27001 builds an ISMS, ISO 42001 builds an AIMS — policies, procedures, roles, risk processes, controls, continual improvement — scoped to AI lifecycle.

Companion standards:

ISO/IEC 23894 — AI risk management (informative)
ISO/IEC 22989 — AI concepts and terminology
ISO/IEC 5338 — AI system life cycle processes
ISO/IEC 5259 series — data quality for analytics and ML
ISO/IEC 38507 — governance implications of AI for boards

42001 is the certifiable umbrella; the others fill in technical detail.

Preconditions / where it applies

Organisations developing AI models or systems (model labs, SaaS with ML features)
Organisations deploying AI in products
Organisations using AI procured from third parties (incl. ChatGPT-type integrations)
AI components subject to EU AI Act high-risk classification → strong commercial driver for 42001 certification

Structure (Annex A controls)

42001 Annex A has 39 controls in 9 categories:

A.2 Policies related to AI — AI policy, alignment with org strategy
A.3 Internal organisation — roles, responsibilities, board oversight
A.4 Resources for AI systems — data, tooling, computing, human resources
A.5 Assessing impacts of AI systems — AI impact assessment (FRIA-style)
A.6 AI system life cycle — design, development, verification, deployment, decommissioning
A.7 Data for AI systems — provenance, quality, bias mitigation
A.8 Information for interested parties of AI systems — transparency, explainability, user info
A.9 Use of AI systems — operational controls, monitoring
A.10 Third-party and customer relationships — supplier AI risk, customer agreements

AI Impact Assessment (A.5)

Closest analogue to the AI Act’s Fundamental Rights Impact Assessment (FRIA). Required for each AI system:

Intended purpose + reasonably foreseeable misuse
Affected individuals and groups
Risks: fairness, accuracy, reliability, robustness, security, privacy, transparency
Risk mitigation measures
Residual risk acceptance
Review trigger conditions

Document; update on material change; produce on auditor request.

Implementation tradecraft

Phase 1 — Scope and governance (Months 1-2)

Decide what’s in/out of scope (which AI systems, departments, lifecycle phases)
Appoint AIMS owner — usually CTO, CISO, or new “Head of AI Governance”
Board commitment statement (parallel to 27001 management commitment)
Inventory all AI systems — first-party + procured + embedded-in-SaaS (often the longest list)

Phase 2 — Risk and impact assessment (Months 2-4)

AI risk register: hallucination, bias, model drift, prompt injection, training-data leakage, supply chain (Hugging Face model integrity)
Per-system AI Impact Assessment using A.5 template
Map AI risks to existing 27001 risks (overlap with information security, privacy)

Phase 3 — Control implementation (Months 4-9)

Data governance: provenance, consent, retention, bias testing (A.7)
MLOps: versioned models, reproducible training, deployment gates (A.6)
Monitoring: drift detection, performance regression, fairness metrics in production (A.9)
Transparency artifacts: model cards, datasheets for datasets, system cards (A.8)
Vendor AI assessment: due diligence checklist for procured LLMs (A.10)

Phase 4 — Internal audit + management review (Month 10)

Mock audit by independent internal auditor or consultant
Top management reviews KPIs, audit findings, risks
Capture in records before certification

Phase 5 — Certification body audit (Month 11-12)

Stage 1 (document review)
Stage 2 (on-site / remote evidence sampling)
Three-year cycle with surveillance audits annually

Intersection with regulations

EU AI Act (in force 2024, full applicability 2026-2027) — 42001 conformance is strong evidence of compliance, particularly for high-risk AI providers; expected to be referenced in harmonised standards (hEN) under Annex IV
NIST AI Risk Management Framework (US) — not certifiable; 42001 is the closest certifiable framework aligned to it
US Executive Order 14110 (Biden, partly rescinded under Trump 2025) — referenced AI safety practices; many requirements survive in agency policies
UK AI Regulation — principle-based, sector-led; 42001 cited as voluntary baseline
China — Generative AI Measures — Chinese national standards diverge; 42001 useful for multinationals
GDPR + EU AI Act overlap — AIMS must cover personal data dimensions; 27701 + 42001 reinforce each other

Common implementation pitfalls

Treating AI as a security control alone — AIMS covers ethics, fairness, transparency, beyond CIA triad
Vendor AI invisibility — internal teams use ChatGPT, Copilot, Gemini, embedded LLMs; inventory is harder than first-party AI
Bias testing without representative test sets — A.7 requires demonstrable mitigation, not just a checkbox
No model decommissioning process — A.6 requires lifecycle including retirement
Transparency theater — generic disclaimers don’t meet A.8; need system-specific explanations

Certification market signal

42001 certifications are early; first wave 2024-2025. Expect strong demand from:

Enterprise AI vendors selling to regulated buyers (banks, insurers, governments)
Public sector AI providers
Multinationals preparing for EU AI Act high-risk obligations
Healthcare and HR-tech AI providers (sectoral scrutiny)

Many orgs will pair 42001 + 27001 + 27701 as the “trust trifecta” for AI products.

OPSEC for compliance team

AIMS overlaps deeply with ISMS — share auditors, processes, document control
Keep AI Impact Assessments TLP:AMBER — they describe weakness/limitations of your AI
Model cards published externally are sanitised versions of the assessment, NOT the assessment itself
Track regulatory updates; the AI compliance landscape changes faster than other frameworks