Last updated: September 29, 2026 · Status of every framework and standard checked on nist.gov, iso.org, cencenelec.eu and EUR-Lex on this date
The three AI risk management frameworks that matter in 2026 do different jobs. The NIST AI RMF is a free, voluntary playbook (Govern, Map, Measure, Manage) that you can’t be certified against. ISO/IEC 42001 is a certifiable AI management system standard with 38 Annex A controls. EU AI Act Article 9 is a legal duty for high-risk AI providers, applying from December 2, 2027. ISO 42001 certification does not by itself prove AI Act compliance: it isn’t a harmonised standard, and the standard written for Article 9 (prEN 18228) is still a draft.
Most “NIST vs ISO 42001” articles are written by compliance vendors, and they stop at the comparison table. Before writing, we compared eight of them question by question. None explained what the EU AI Act’s Article 9 actually requires. None said whether ISO 42001 counts as AI Act compliance, gave the status of the EU harmonised standards, showed a filled-in risk register, or covered LLM-specific threat catalogues like OWASP and MITRE ATLAS. This guide adds those, with dates checked for 2026. If you’re looking for how AI is used in risk management (fraud, credit), see AI for risk management. This page is about managing the risks of AI.
How do NIST AI RMF, ISO 42001 and the EU AI Act compare?
| NIST AI RMF 1.0 | ISO/IEC 42001:2023 | EU AI Act Art. 9 (+ Art. 17) | |
|---|---|---|---|
| What it is | Voluntary US framework plus a free Playbook | International management-system standard (AIMS) | Binding EU law for providers of high-risk AI systems |
| Binding? | No | No, but customers increasingly require it in contracts | Yes. Fines up to €15M or 3% of turnover (see AI Act fines) |
| Certifiable? | No. You can claim alignment only | Yes, by accredited bodies (ISO/IEC 42006:2025) | Conformity assessment, not certification of the organisation |
| Structure | 4 functions, 19 categories, 72 subcategories | Clauses 4–10 plus Annex A: 38 controls in 9 objectives | Risk management system (Art. 9) plus quality management system (Art. 17), technical documentation, logging, oversight |
| Scope | Any AI system, any organisation | The organisation’s AI management system | Each high-risk AI system, across its lifecycle |
| Cost | Free | Audit fees plus internal effort | Compliance cost plus legal exposure |
| Status in 2026 | 1.0 (Jan 2023) “being revised”; no new version yet | Current. Guidance on it: ISO/IEC 23894 (risk), 42005 (impact assessment) | Applies from Dec 2, 2027 (Annex III) and Aug 2, 2028 (Annex I) after the Digital Omnibus |
The short version: NIST tells you what to think about, ISO 42001 tells you how to run it as a managed, auditable system, and the AI Act tells you what the law requires for specific high-risk systems. The three layer on top of each other; they don’t replace each other.
What is the NIST AI RMF and what’s new in 2026?
The NIST AI Risk Management Framework 1.0 (NIST AI 100-1, January 2023) organises AI risk work into four functions:
- GOVERN: policies, accountability, culture. It applies across the other three (6 categories).
- MAP: understand the context, intended use and possible impacts of each system (5 categories).
- MEASURE: test, evaluate and track the identified risks (4 categories).
- MANAGE: prioritise, treat, monitor and respond (4 categories).
For generative AI, NIST published the Generative AI Profile (NIST AI 600-1, July 2024). It lists 12 risks: CBRN information; confabulation; dangerous, violent or hateful content; data privacy; environmental impacts; harmful bias and homogenisation; human-AI configuration; information integrity; information security; intellectual property; obscene or degrading content; and value-chain and component integration.
2026 status. The July 2025 US AI Action Plan told NIST to revise the framework. As of September 29, 2026, NIST says only that “the AI RMF 1.0 is being revised”. No 1.1 or 2.0 has been published, so 1.0 is still the version to use. NIST has also been drafting SP 800-53 control overlays for AI systems (COSAiS), but no final overlays have been released. Don’t confuse either with NIST AI 300-1, a July 2026 draft on public-facing AI documentation.
What does ISO/IEC 42001 require?
ISO/IEC 42001:2023 is the first certifiable standard for an AI management system. It follows the same clause structure as ISO 27001:
- clauses 4–10: context, leadership, planning, support, operation, performance evaluation, improvement;
- Annex A: 38 controls grouped under 9 objectives (A.2–A.10), from AI policy and roles to data, the AI system lifecycle and third-party relationships.
Certification bodies are audited against ISO/IEC 42006:2025, so check that yours is accredited to it. Big AI providers certified to 42001 include Anthropic (January 2025), AWS (November 2024, covering Bedrock and other services), Microsoft (Copilot and Copilot Chat) and Google (Google Cloud, Workspace and the Gemini app). When you buy AI from them, their certificate is useful evidence for your supplier assessment. It doesn’t replace your own risk management.
How much does ISO 42001 certification cost and how long does it take?
We found no authoritative public price list. Certification bodies quote per engagement, and none of the eight articles we compared gave a sourced figure. What drives the cost is predictable: the number of AI systems and sites in scope, how many audit days the body calculates, whether you already run ISO 27001 (a big head start, since the clause structure is shared), and the internal hours to write policies and run the risk and impact assessments. Vendor guides typically quote 6–12 months from start to certificate. Treat any single “ISO 42001 costs $X” number as marketing.
ISO/IEC 23894 and 42005: the two guidance standards
- ISO/IEC 23894:2023 is guidance on AI risk management, built on ISO 31000. You can’t certify against it. Use it to design the risk process that 42001 and Article 9 expect.
- ISO/IEC 42005:2025 is guidance on AI system impact assessment: effects on individuals, groups and society. It maps to 42001 and is the closest ISO equivalent to the AI Act’s fundamental-rights focus.
What does EU AI Act Article 9 require?
Article 9 requires every provider of a high-risk AI system to run a risk management system: a continuous, iterative process across the system’s whole lifecycle. It must:
- identify and analyse the known and reasonably foreseeable risks to health, safety and fundamental rights;
- estimate and evaluate those risks under intended use and reasonably foreseeable misuse;
- evaluate new risks found through post-market monitoring;
- adopt targeted measures, so that the remaining residual risk is judged “acceptable”;
- test the system against metrics and probabilistic thresholds defined in advance;
- consider the impact on people under 18 and other vulnerable groups.
Article 17 adds a quality management system. Both apply from December 2, 2027 for Annex III systems (employment, credit, education, essential services and others) and from August 2, 2028 for AI embedded in regulated products. The Digital Omnibus (Regulation (EU) 2026/1744) pushed both back from August 2026. See our EU AI Act guide for the full timeline.
Does ISO 42001 certification prove EU AI Act compliance?
No. Under the AI Act, following a harmonised standard cited in the Official Journal gives a “presumption of conformity”. ISO/IEC 42001 is not one. The European Commission’s Joint Research Centre found it “not aligned in objectives and approach with the AI Act”. It manages an organisation’s AI governance, while the Act regulates individual products and their risks to fundamental rights.
The European standards written for the Act are arriving late:
| Standard (CEN-CENELEC JTC 21) | Supports | Status, Sep 2026 |
|---|---|---|
| EN 18286:2026 — quality management system | Art. 17 | Published (approved July 2026). Presumption of conformity only once cited in the Official Journal |
| prEN 18228 — AI risk management | Art. 9 | Draft. Enquiry closed end of July 2026; target Q4 2026 |
| Other JTC 21 drafts (bias, data, logging, accuracy) | Arts. 10–15 | In development |
What this means in practice: 42001 is still worth doing if you sell to enterprises or want an auditable backbone. Much of its evidence (policies, risk assessments, lifecycle controls) will be reused for Article 9. But plan a gap analysis against prEN 18228 once it’s final, rather than assuming your certificate covers you.
Which framework should you start with?
| You are… | Start with | Why |
|---|---|---|
| A startup using LLM APIs, no high-risk use | NIST AI RMF + GenAI Profile, plus OWASP LLM Top 10 | Free, fast to adopt, and matches what enterprise customers ask about. Skip certification until a customer requires it |
| A B2B SaaS vendor selling AI features to enterprises | ISO 42001 (especially if you already hold ISO 27001) | Procurement teams increasingly accept a certificate instead of lengthy questionnaires |
| A provider of a high-risk system in the EU (hiring, credit, education…) | AI Act Art. 9 + 17 as the requirement, with 42001 as the operating system, and prEN 18228 when final | Legal duty from Dec 2027; building on 42001 avoids two parallel systems |
| A bank or insurer | Your existing model risk management (SR 11-7 in the US) extended with the NIST GenAI Profile | Supervisors already expect model validation. Add GenAI-specific risks rather than a new framework |
| A deployer of someone else’s AI | A lightweight register + vendor assessment; AI Act Art. 26 duties if high-risk | You can’t audit the model, but you own its use: oversight, logging, instructions |
How do the frameworks map to each other? A simplified crosswalk
NIST publishes an official AI RMF ↔ ISO/IEC 42001 crosswalk. The table below is our simplified, three-way version, including the AI Act articles. It’s for orientation, not a substitute for the official crosswalk.
| NIST AI RMF | ISO/IEC 42001 | EU AI Act | Evidence you’d show |
|---|---|---|---|
| GOVERN | Cl. 5 Leadership, 6 Planning, 7 Support; A.2 policies, A.3 roles | Art. 17 QMS; Art. 4 AI literacy | AI policy, named owners, training records |
| MAP | Cl. 6.1 risk & impact assessment; A.5 impact assessment (with ISO 42005) | Art. 9(2)(a)-(b) risk identification; Art. 27 fundamental rights impact assessment (certain deployers) | System inventory, intended-use statement, impact assessment |
| MEASURE | Cl. 9 performance evaluation; A.6 lifecycle verification | Art. 9(6)-(8) testing; Art. 15 accuracy, robustness, cybersecurity | Test plans, eval results, red-team reports |
| MANAGE | Cl. 8 operation, 10 improvement | Art. 9(2)(d) risk measures; Art. 72 post-market monitoring; Art. 73 serious incidents | Risk treatment decisions, monitoring dashboards, incident log |
How to build an AI risk register (with examples)
Every framework ends up needing the same thing: a register that lists each AI system, its risks, the controls and an owner. None of the articles we compared showed a filled-in one. Here’s a starting template with three realistic rows. Scores use 1–5 for likelihood (L) and impact (I).
| System | Risk | L | I | Controls | Framework refs | Owner |
|---|---|---|---|---|---|---|
| Customer-support chatbot (LLM API) | Confabulated refund or policy terms given to customers | 4 | 3 | Retrieval from approved policy docs only; refusal on uncovered topics; weekly sample review; “you’re talking to AI” notice | NIST 600-1 confabulation; OWASP LLM09; AI Act Art. 50(1) | Head of Support |
| CV screening model (vendor) | Discriminatory ranking of candidates | 3 | 5 | Vendor bias test results; human review of every rejection; quarterly adverse-impact analysis; logs kept | AI Act Annex III (employment) — Art. 26 deployer duties; ISO 42005 impact assessment | HR Director + DPO |
| Coding agent with repo and shell access | Prompt injection via issue text leads to secret exfiltration | 3 | 4 | Sandboxed runner; no production credentials; tool allow-list; human approval for pushes | OWASP LLM01 prompt injection, LLM06 excessive agency; MITRE ATLAS | Engineering lead |
Keep the register small and alive. One row per real risk, reviewed when the system or its model changes, beats a 200-row spreadsheet nobody opens. For agents, pair it with the checks in our AI agent testing checklist and agent observability guide.
What about LLM and agent risks? OWASP and MITRE ATLAS
The big frameworks are deliberately technology-neutral, so for LLM apps and agents you need a threat catalogue on top:
- OWASP Top 10 for LLM Applications (2025): LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, LLM03 Supply Chain, LLM04 Data and Model Poisoning, LLM05 Improper Output Handling, LLM06 Excessive Agency, LLM07 System Prompt Leakage, LLM08 Vector and Embedding Weaknesses, LLM09 Misinformation, LLM10 Unbounded Consumption. OWASP released a 2026 edition on August 3, 2026. Check it before you standardise on the 2025 labels.
- MITRE ATLAS: a knowledge base of real adversary tactics and techniques against AI systems, modelled on ATT&CK. Version 5.1 (November 2025) lists 16 tactics and 84 techniques. Use it for red-team planning.
In NIST terms, these feed MAP (what can go wrong) and MEASURE (how you test it). Our AI agent security guide covers the agent-specific controls.
Banks: where does SR 11-7 fit?
US banks have run model risk management under the Federal Reserve’s SR 11-7 guidance (issued with the OCC in 2011) for over a decade: model inventory, independent validation, ongoing monitoring. It maps cleanly onto MAP and MEASURE. The gap is generative AI. SR 11-7 assumes models with defined outputs that can be back-tested, while LLMs need the confabulation, information-integrity and value-chain risks from NIST AI 600-1. The pragmatic path is to extend the existing model risk policy, not to start a second framework.
A 90-day plan to get started
- Days 1–15 — inventory. List every AI system you build, buy or embed, and flag anything that might be high-risk under AI Act Annex III.
- Days 15–30 — govern. Name an accountable owner, write a two-page AI policy, set up AI literacy training (an AI Act duty since February 2025).
- Days 30–60 — map and measure. Fill the risk register for your top five systems, run an impact assessment on anything touching people’s rights, and test against OWASP LLM risks.
- Days 60–90 — manage. Decide on treatments, set up monitoring and an incident log, and decide whether ISO 42001 certification makes business sense.
This order also protects you if something goes wrong. Documented risk management, cooperation and self-reporting are all factors that lower an AI Act fine. See how AI Act fines are calculated.
FAQ
What is the difference between NIST AI RMF and ISO 42001?
The NIST AI RMF is a free, voluntary US framework organised into four functions (Govern, Map, Measure, Manage) and cannot be certified. ISO/IEC 42001 is an international management-system standard with 38 Annex A controls that accredited bodies can certify.
Does ISO 42001 certification mean EU AI Act compliance?
No. ISO/IEC 42001 is not a harmonised standard under the AI Act, so it gives no presumption of conformity. The dedicated European standard for AI Act risk management, prEN 18228, was still a draft in September 2026.
Is the NIST AI RMF mandatory?
No. It is voluntary guidance. NIST is revising version 1.0 after the July 2025 AI Action Plan, but as of September 2026 no new version has been published.
When does EU AI Act Article 9 apply?
For high-risk systems listed in Annex III, from December 2, 2027; for AI in regulated products (Annex I), from August 2, 2028. The Digital Omnibus (Regulation 2026/1744) moved both dates from August 2026.
Which AI companies are ISO 42001 certified?
Among large providers: Anthropic (January 2025), AWS (November 2024), Microsoft for Copilot and Copilot Chat, and Google for Google Cloud, Workspace and the Gemini app.
What risks does the NIST Generative AI Profile list?
Twelve: CBRN information, confabulation, dangerous or hateful content, data privacy, environmental impacts, harmful bias, human-AI configuration, information integrity, information security, intellectual property, obscene or degrading content, and value-chain and component integration.
Which AI risk framework should a startup use first?
Usually the NIST AI RMF with its Generative AI Profile plus the OWASP Top 10 for LLM Applications. They are free and quick to adopt. ISO 42001 certification makes sense once enterprise customers ask for it.
Polish readers: zarządzanie ryzykiem AI — NIST, ISO 42001 i AI Act po polsku.
Sources & methodology
- NIST AI 100-1 — AI Risk Management Framework 1.0 (January 2023).
- NIST AI 600-1 — Generative AI Profile (July 2024).
- NIST — AI RMF page (“being revised” status).
- NIST — AI RMF to ISO/IEC 42001 crosswalk.
- NIST — COSAiS control overlays concept paper (August 2025).
- ISO — ISO/IEC 42001:2023.
- ISO — ISO/IEC 42006:2025.
- ISO — ISO/IEC 23894:2023.
- ISO — ISO/IEC 42005:2025.
- EU AI Act Article 9 — risk management system.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — EUR-Lex.
- CEN-CENELEC — EN 18286 AI quality management (July 2026).
- European Commission JRC — analysis of ISO/IEC 42001 vs AI Act requirements.
- Anthropic — ISO 42001 certification (January 2025).
- AWS — ISO/IEC 42001 accredited certification.
- OWASP — Top 10 for LLM Applications 2025.
- MITRE ATLAS.
Method: on September 29, 2026 we extracted the top-ranking English pages for “ai risk management framework” (8 of 10 could be fetched), scored 24 practitioner questions against each with a local reranker, and had every candidate answer checked by an LLM judge. The crosswalk and risk register are DecodeTheFuture’s own simplification, not official mappings.
