Our main guide, AI Governance for Enterprises: A Practical Framework for IT, Cybersecurity, and Risk Leaders, makes a point worth repeating: not every framework in this space is the same kind of thing. Some are legally binding. Some are certifiable management standards. Some are voluntary guidance. Some are just practical security references. Treating them as interchangeable is one of the fastest ways to misjudge what your organization actually needs to do.
This post goes deeper on each major framework, standard, and regulation — what it is, who publishes it, who should care about it, and where it fits in a governance program. As always: this is original commentary based on publicly available information, not a reproduction of any standard’s text. For binding requirements, always confirm current details against the official source before making compliance decisions.
First, the Taxonomy
| Type | What it means | Example |
|---|---|---|
| Regulation/Law | Legally binding, enforceable | EU AI Act |
| Management-system standard | Certifiable standard for running an ongoing program | ISO/IEC 42001 |
| Guidance standard | Structured but non-certifiable ISO/IEC guidance | ISO/IEC 23894, ISO/IEC 38507 |
| Voluntary framework | Government/industry-published, adopted at will | NIST AI RMF, NIST CSF 2.0 |
| Principles/soft law | High-level normative commitments, not directly enforceable on companies | OECD AI Principles, UNESCO Recommendation on AI Ethics |
| Security knowledge base | Practical technique-level reference for testing and defense | MITRE ATLAS, OWASP LLM Top 10, CIS Controls |
NIST AI Risk Management Framework (AI RMF)
What it is: A voluntary framework published by the U.S. National Institute of Standards and Technology (NIST), built around four functions — Govern, Map, Measure, Manage — for incorporating trustworthiness (validity, safety, security, accountability, fairness, explainability, privacy) into AI systems.
Who should use it: Any organization, regardless of location, looking for a structured way to think about AI risk. It’s not limited to U.S. entities.
Mandatory or voluntary: Voluntary. It is not a law, though some U.S. federal agencies reference it in their own internal requirements.
Notable extension: The Generative AI Profile (NIST AI 600-1), published in 2024, adds GenAI-specific risks like confabulation and prompt-based data leakage. NIST has also been revising the core framework as part of broader federal AI policy work — worth checking for updates before relying on version specifics.
Where it fits: Best used as the backbone of your AI risk management process, typically once basic AI governance ownership and inventory are in place.
NIST Cybersecurity Framework (CSF) 2.0 — and its emerging AI profile
What it is: NIST’s general-purpose, voluntary cybersecurity framework, organized around six functions (Govern, Identify, Protect, Detect, Respond, Recover). It isn’t AI-specific, but AI systems and the infrastructure around them fall inside its scope.
Notable development: NIST released a preliminary draft “Cybersecurity Framework Profile for Artificial Intelligence” in December 2025, addressing securing AI systems, using AI for cyber defense, and defending against AI-enabled attacks. As of this writing it remains in draft/comment status — treat it as an emerging reference, not a finalized standard, and check NIST’s site for its current state.
Where it fits: The cybersecurity governance backbone that AI-specific security work should extend rather than duplicate.
ISO/IEC 42001 — AI Management System
What it is: The first internationally certifiable standard for an “AI management system” (AIMS) — the organizational structures, policies, and continual-improvement processes for responsibly developing, providing, or using AI. Published by ISO/IEC.
Who should use it: Organizations that develop AI, embed it in products or services, use it operationally, or manage third-party AI — particularly those wanting an external, auditable governance credential.
Mandatory or voluntary: Voluntary; certification is performed by independent accredited bodies, not ISO itself.
Where it fits: The organizational “operating system” layer — the management structure that other, more specific frameworks plug into. Most useful once an organization is past initial ad hoc governance and ready to formalize.
ISO/IEC 23894 — Guidance on AI Risk Management
What it is: Non-certifiable ISO/IEC guidance that adapts general enterprise risk management principles specifically to AI risk sources.
Where it fits: Frequently used alongside ISO/IEC 42001 — 42001 defines the management system, 23894 provides deeper risk-methodology guidance to support it. Useful earlier in maturity, when formal AI risk assessment processes are first being built.
ISO/IEC 27001 and ISO/IEC 27701 — Information Security and Privacy Management
What they are: ISO/IEC 27001 is the globally recognized certifiable standard for an information security management system (ISMS). ISO/IEC 27701 extends it to privacy information management. Neither is AI-specific.
Where they fit: AI systems, training data, model endpoints, and AI vendor relationships all sit inside the scope of an organization’s existing information security and privacy management. AI governance should reuse these controls (access management, supplier security, data subject rights) rather than build a parallel structure.
ISO/IEC 38500 and ISO/IEC 38507 — Governance of IT and AI
What they are: ISO/IEC 38500 sets board-level principles for governing IT generally (evaluate, direct, monitor). ISO/IEC 38507 is a targeted extension applying those same governing-body responsibilities specifically to AI adoption. Neither is certifiable — they’re governance-principle standards aimed at boards and executives, not technical control catalogs.
Where they fit: The clearest anchor for tying AI oversight back to existing corporate and IT governance structures, making them especially relevant to the “AI IT Governance” side of a program.
OECD AI Principles and the UNESCO Recommendation on the Ethics of AI
What they are: Intergovernmental principles, not laws. The OECD AI Principles (2019, updated 2024) were the first intergovernmental AI principles adopted by member governments. UNESCO’s Recommendation on the Ethics of Artificial Intelligence (2021) is a normative instrument adopted by all UNESCO member states, covering human rights, sustainability, and social impact.
Mandatory or voluntary: Not directly enforceable against companies — governments commit to these at a policy level, and they’ve influenced later instruments (including aspects of the EU AI Act’s thinking).
Where they fit: Framing conversations about AI ethics and values with leadership and boards, rather than shaping day-to-day technical controls.
EU AI Act
What it is: The one genuinely binding regulation on this list — Regulation (EU) 2024/1689. It entered into force in August 2024 with a risk-tiered structure: prohibited practices, high-risk systems, limited-risk/transparency obligations (including generative AI disclosure requirements), and minimal-risk systems.
Key dates (verify current status before relying on these): Prohibited-practice provisions applied from February 2025; general-purpose AI governance obligations applied from August 2025; most high-risk-system obligations phase in through December 2027, with certain embedded high-risk products extending to August 2028. An “AI Omnibus” simplification package adjusted some transition periods and SME documentation burdens in mid-2026.
Who it applies to: It can apply extraterritorially — organizations outside the EU whose AI systems are placed on the EU market, or whose outputs are used within the EU, may be in scope regardless of headquarters location.
Where it fits: The legal floor that must be met when in scope, regardless of which voluntary frameworks an organization also chooses to adopt.
CIS Controls and AI Companion Guides
What they are: The Center for Internet Security’s prioritized cybersecurity control catalog, now supplemented with dedicated AI & LLM and AI Agents companion guides mapping the existing controls to AI-specific implementation guidance.
Where they fit: A practical way to operationalize AI security controls without inventing a parallel control set from scratch — useful once an organization is ready to move from principles to specific technical safeguards.
OWASP LLM, GenAI, and Agentic Top 10 Guidance
What it is: Community security guidance from the OWASP Gen AI Security Project identifying the most critical application-layer risks in LLM and generative AI systems — things like prompt injection, insecure output handling, training-data poisoning, and excessive agency in AI agents. Not a certifiable standard.
Where it fits: The primary AI/LLM application-security reference for development and security teams once an organization has AI applications in production and needs to test and harden them specifically.
MITRE ATLAS
What it is: A structured knowledge base of real-world adversary tactics and techniques against AI/ML systems, developed by MITRE with industry contributors, modeled on the structure of MITRE ATT&CK. It’s a reference for understanding attacks, not a compliance framework.
Where it fits: AI-specific threat modeling, red-team scenario design, and detection engineering — most valuable at more advanced security maturity, once basic AI security controls are already in place.
Singapore’s Model AI Governance Framework and AI Verify
What it is: A voluntary national framework and companion testing toolkit published by Singapore’s Info-communications Media Development Authority (IMDA), extended over time to address generative and agentic AI. It’s one of the earliest and most widely referenced national AI governance frameworks internationally.
Why it’s worth knowing: It’s a useful reminder that AI governance is a genuinely global discipline, not one shaped by a single region — and it’s frequently referenced by other regulators as a practical model.
Putting It Together
| Framework | Type | Mandatory? | Best fit in a program | Maturity level |
|---|---|---|---|---|
| NIST AI RMF | Voluntary framework | No | Structuring AI risk management | Developing–Defined |
| NIST CSF 2.0 (+ draft AI profile) | Voluntary framework | No | Cybersecurity governance backbone | Defined–Managed |
| ISO/IEC 42001 | Management-system standard | No (certifiable) | Organizational AI management foundation | Defined–Managed |
| ISO/IEC 23894 | Guidance standard | No | AI risk methodology | Developing–Defined |
| ISO/IEC 27001 / 27701 | Management-system standards | No (certifiable) | Information security and privacy backbone | Any (existing programs) |
| ISO/IEC 38500 / 38507 | Governance-principle standards | No | Board-level AI oversight | Aware–Defined |
| OECD Principles / UNESCO Recommendation | Principles/soft law | No | Ethics and values framing | Aware |
| EU AI Act | Regulation | Yes (where in scope) | Legal compliance floor | Any (when applicable) |
| CIS Controls + AI guides | Security knowledge base | No | AI security control implementation | Managed–Optimized |
| OWASP LLM/GenAI/Agentic Top 10 | Security knowledge base | No | AI/LLM application security | Managed–Optimized |
| MITRE ATLAS | Security knowledge base | No | AI threat modeling and red teaming | Advanced–Optimized |
None of these substitute for the others. The practical approach — and the one we recommend to clients — is to combine layers deliberately: a management-system foundation, a risk-management method, a cybersecurity backbone, AI-specific security references, and applicable legal obligations, matched to your organization’s actual sector, geography, and risk profile. For a step-by-step way to get there, see our AI Governance Maturity Model and our main guide’s starting roadmap.