Join our Newsletter — 33% off our NHI Course
Home› Guides› NIST AI RMF vs OECD AI Principles: Turning…
Guide AI Governance & Risk

NIST AI RMF vs OECD AI Principles: Turning Policy Values Into Risk Management Outcomes

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 14 November 2026 17 min read
On this page

Many AI governance programmes quote the OECD AI Principles in their policy and the NIST AI Risk Management Framework in their risk process, without ever working out how the two fit together, so values sit in one document and controls in another with nothing connecting them. The OECD Principles are an intergovernmental statement of what trustworthy AI should be. The AI RMF (NIST AI 100-1) is a voluntary method for deciding what to do about AI risk inside an organisation. They are not rivals: NIST built its AI system definition, its idea of AI actors and its lifecycle figure on OECD work, and the OECD definition of an AI system also shaped the EU AI Act. This guide compares the two on purpose, structure, audience and evidence, gives a principle to subcategory crosswalk, traces how the OECD definition flows into regulation, and shows how to run both in one programme. It is for AI governance leads, GRC and security teams and the executives who sign off AI policy. Status checked on 14 November 2026: the OECD text in force is the May 2024 update, and NIST states that AI RMF 1.0 is being revised under the White House AI Action Plan.

Key takeaways

  • Short answer: OECD tells you what to aim for, the AI RMF tells you how to manage it. Adopt the OECD Principles as the value statement in your AI policy and use the AI RMF functions to turn each value into owned, measured outcomes.
  • Neither is binding on companies. The OECD Recommendation commits adherent governments politically and the AI RMF is intended for voluntary use. Legal force arrives through laws that reuse their concepts, such as the EU AI Act.
  • The AI RMF is partly built on OECD work. NIST adapted its AI system definition from the 2019 OECD Recommendation and ISO/IEC 22989, took the AI actor concept from the OECD, and modified its lifecycle figure from the OECD Framework for the Classification of AI Systems.
  • The definition is the bridge to regulation. The OECD revised its AI system definition in November 2023 to align with work in the EU and Japan, and Article 3(1) of the AI Act uses closely matching wording. Scoping your inventory with that definition makes later legal mapping easier.
  • Map once, evidence once. A single crosswalk from the five OECD values to AI RMF subcategories, with one evidence item per row, serves policy, audit and regulatory questions without three parallel programmes.

What each one is

OECD AI Principles

The OECD Recommendation on Artificial Intelligence was first adopted in 2019 and updated in May 2024, when the revisions adopted at the OECD Ministerial Council Meeting responded to general-purpose and generative AI and addressed privacy, intellectual property, safety and information integrity more directly. It has two parts. Five values-based principles apply to all AI actors: inclusive growth, sustainable development and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; and accountability. Five recommendations are addressed to governments: investing in AI research and development, fostering an inclusive AI-enabling ecosystem, shaping an enabling interoperable governance and policy environment, building human capacity and preparing for labour market transition, and international co-operation for trustworthy AI. The OECD lists 47 adherents, and G20 members committed to G20 AI Principles drawn from the OECD text.

The principles are outcome statements. Principle 1.4 expects mechanisms so that AI systems that risk undue harm or show undesired behaviour can be overridden, repaired and/or decommissioned safely as needed. Principle 1.5 expects traceability of datasets, processes and decisions, and a systematic risk management approach at each phase of the lifecycle, scaled to each actor's role, context and ability to act. They do not say who in your organisation does this, how often, or what evidence proves it. For the full text and a control view of each principle, see the OECD AI Principles Guide.

NIST AI RMF 1.0

NIST released AI RMF 1.0 on 26 January 2023 and describes it as intended for voluntary use. Its Core has four functions, GOVERN, MAP, MEASURE and MANAGE, with 6, 5, 4 and 4 categories, each broken into subcategories that read as outcomes, for example MAP 1.1 on intended purposes and context-specific laws, or MANAGE 2.4 on mechanisms to supersede, disengage or deactivate an AI system. NIST adds seven characteristics of trustworthy AI: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. Companion material includes the Playbook, a set of community crosswalks and the Generative AI Profile, NIST AI 600-1, released on 26 July 2024. For depth on the Core and profiles, see the NIST AI RMF Guide.

Side-by-side comparison

DimensionOECD AI PrinciplesNIST AI RMF 1.0
NatureIntergovernmental Recommendation; political commitment by adherent governmentsVoluntary framework from a US standards agency
Primary audienceGovernments (section 2) and all AI actors (section 1)Organisations that design, develop, deploy, evaluate or use AI
Unit of contentFive values and five policy recommendationsFour functions, 19 categories, subcategories as outcomes
Level of abstractionWhat trustworthy AI should achieveWhat an organisation should have in place to manage AI risk
Risk methodExpects a systematic risk management approach, does not prescribe oneIs the method: context mapping, measurement, treatment, governance
Definition of AI systemRevised November 2023: adds content outputs, implicit objectives and adaptiveness after deploymentAdapted from the 2019 OECD Recommendation and ISO/IEC 22989:2022
Classification aidOECD Framework for the Classification of AI Systems (February 2022), five dimensionsLifecycle and key dimensions figure modified from that OECD framework
Generative AI coverage2024 update addresses general-purpose and generative AIGenerative AI Profile, NIST AI 600-1
Status in November 2026May 2024 text current; 47 adherentsVersion 1.0, under revision per NIST
Evidence it producesNone by itself; reporting via government monitoringProfiles, risk records and measurement results if you implement it

Threat model and failure modes of each

Each framework used alone fails in predictable ways, which is why mature programmes use both.

Using only the OECD Principles

  • Values without owners: a policy that says "our AI is transparent and accountable" with no named role, review frequency or test.
  • No measurement: the principles do not define metrics, so fairness or robustness claims rest on assertion.
  • Override as a slogan: principle 1.4 asks for safe override and decommissioning, but without a tested procedure and a credential that can actually be revoked, an AI agent keeps running. See the AI Agent Kill Switch and Containment Guide.

Using only the AI RMF

  • Method without a mandate: the AI RMF tells you to set risk tolerance but does not say what values the organisation is committing to. Boards and regulators outside the US recognise OECD language more readily.
  • Checklist drift: NIST states that Core actions are not a checklist, yet many programmes score every subcategory yes or no and stop there.
  • Parochial framing: a programme expressed only in AI RMF terms can be harder to explain to EU, Japanese or Korean stakeholders, whose national guidance draws on the OECD text.
  • Revision exposure: NIST says AI RMF 1.0 is being revised. Tying policy wording to subcategory numbers means rework when the revision lands; tying it to OECD values does not.

Crosswalk: OECD principles to AI RMF outcomes

NIST published a crosswalk on 26 January 2023 that relates AI RMF trustworthiness characteristics to the OECD Recommendation, the then proposed EU AI Act, Executive Order 13960 and the Blueprint for an AI Bill of Rights. It predates the 2024 OECD update and the final AI Act, and NIST notes that listing a crosswalk does not mean either document fully covers the other. The table below is an NHI Mgmt Group working mapping from the current OECD text to AI RMF subcategories, with an evidence item per row. Treat it as a starting point and record your own rationale.

OECD principleAI RMF characteristicAI RMF subcategoriesEvidence to keep
1.1 Inclusive growth, sustainable development and well-beingFair, with harmful bias managed (partly)MAP 1.1 intended purposes and beneficial uses; MAP 5.1 likelihood and magnitude of impacts; MEASURE 2.12 environmental impactImpact assessment naming affected groups, intended benefits and compute footprint
1.2 Human rights and democratic values, including fairness and privacyPrivacy-enhanced; fair, with harmful bias managedMEASURE 2.10 privacy risk; MEASURE 2.11 fairness and bias; GOVERN 5.1 external feedbackDPIA, bias test results by group, feedback channel log
1.3 Transparency and explainabilityAccountable and transparent; explainable and interpretableMEASURE 2.8 transparency and accountability risks; MEASURE 2.9 model explained and output interpreted in contextSystem card, user notice, explanation and contest procedure
1.4 Robustness, security and safetyValid and reliable; safe; secure and resilientMEASURE 2.5, 2.6, 2.7; MANAGE 2.4 supersede, disengage or deactivateTest reports including adversarial testing, tested shutdown runbook, rollback record
1.5 AccountabilityAccountable and transparentGOVERN 1.4 transparent policies and procedures; GOVERN 1.6 AI inventory; GOVERN 2.1 roles; GOVERN 6.1 third-party risk; MANAGE 4.1 post-deployment monitoring; MANAGE 4.3 incident communicationAI register with owners, RACI, supplier assessments, monitoring plan, incident log
2.3 Interoperable governance (government recommendation)Not a characteristicGOVERN 1.1 legal and regulatory requirements understood, managed and documentedObligations register listing laws that reuse OECD terms

Principle 1.1 is broader than any AI RMF characteristic, so evidence it through the impact assessment. Principle 1.5 is where the frameworks overlap most, because the AI RMF GOVERN function is in effect a detailed answer to "systematic risk management at each phase". For a ready structure for the inventory evidence, use the AI Register and AI Inventory Template.

How the OECD definition flows into the EU AI Act

The definition of "AI system" decides what enters your inventory, your risk process and, in the EU, your legal scope. Its history explains why the OECD text matters even to teams that never read the Recommendation.

  1. 2019: the OECD Recommendation defines an AI system. NIST adapts that definition, with ISO/IEC 22989:2022, for AI RMF 1.0.
  2. February 2022: the OECD publishes the Framework for the Classification of AI Systems with five dimensions: People and Planet, Economic Context, Data and Input, AI Model, and Task and Output. NIST modifies it for the AI RMF lifecycle and key dimensions figure.
  3. November 2023: OECD members approve a revised definition. The OECD explained that it added "content" to cover generative AI, accepted "explicit or implicit" objectives, and added adaptiveness after deployment, with timing intended to foster alignment with policy processes in the European Union and Japan. An explanatory memorandum followed in March 2024.
  4. 2024: the EU AI Act, Regulation (EU) 2024/1689, defines an AI system in Article 3(1) as a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers from its input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Recital 12 says the notion should be closely aligned with the work of international organisations working on AI, for legal certainty, international convergence and wide acceptance, and names the capability to infer as a key characteristic.

The OECD also states that its definition and lifecycle are used by the EU, the Council of Europe, the United States and the United Nations in legislative and regulatory frameworks. Practical consequence: if you scope your AI inventory using the OECD definition, the boundary between "AI system" and conventional software in your register will closely track the AI Act boundary. Note that AI RMF 1.0 still uses the older 2019 based wording, without "content" or "implicit" objectives, so use the current OECD or AI Act wording for scoping. For the legal duties that follow, see the EU AI Act Compliance Guide and the NIST AI RMF vs EU AI Act.

Architecture in words: one programme, two layers

The cleanest design treats the OECD Principles as the policy layer and the AI RMF as the operating layer, joined by the crosswalk.

  • Policy layer: the AI policy states the five OECD values as commitments, adopts the current OECD or AI Act definition of an AI system for scope, and names an accountable executive. Start from the AI Governance Policy Template.
  • Operating layer: an AI RMF profile selects the subcategories that apply to your context and sets risk tolerance. GOVERN outcomes become procedures, MAP outcomes become the intake and impact assessment, MEASURE outcomes become test plans, MANAGE outcomes become treatment, monitoring and incident handling.
  • Join: the crosswalk table, version controlled, with one evidence pointer per row. When a regulator, customer or auditor asks about "transparency", you answer from the same row whether they used OECD, AI RMF or AI Act vocabulary.

Where the organisation already runs an ISO management system, the operating layer can sit inside it; see the NIST AI RMF vs ISO/IEC 42001 for that choice.

Where identity fits

Three OECD expectations depend on identity controls in practice, and the AI RMF outcomes that answer them are only real if the AI system's own identity is governed.

  • Override and decommission (1.4, MANAGE 2.4): shutting down an agent means revoking its tokens, API keys and workload identity, not only stopping a process. If the credential survives, so does the capability.
  • Traceability (1.5, MEASURE 2.8): decisions are traceable only if each agent acts under a distinct, attributable identity, with logs that separate agent actions from the human who delegated them.
  • Third parties (1.5 responsible business conduct, GOVERN 6.1): model APIs and agent platforms reached with long-lived secrets are supplier risk. Inventory those non-human identities alongside the AI systems that use them.

See the AI Agent Identity vs Service Account for how to give agents distinct identities.

Decision criteria

Your situationLead withWhy
Board wants a short, recognisable statement of AI valuesOECD PrinciplesFive values, widely adopted internationally, stable wording
Risk team needs a repeatable process with testable outcomesAI RMFFunctions and subcategories map to procedures and evidence
US federal customers or US-centric supply chainAI RMF, with OECD in policyUS buyers recognise AI RMF vocabulary
Selling into the EU, Japan or other adherent marketsOECD for scope and policy, AI RMF for operations, plus the local lawNational rules reuse OECD terms; frameworks do not replace legal duties
Small organisation, few AI systemsOECD values plus a light AI RMF profile (GOVERN 1.6, 2.1; MAP 1.1; MEASURE 2.7; MANAGE 2.4)Gets inventory, ownership, security testing and shutdown in place without a full profile
Large enterprise with many business unitsBoth, with a central crosswalk and local profilesConsistent values, context-specific risk tolerance; see the Centralised vs Federated AI Governance

Worked scenario: a mid-sized insurer adopting both

An insurer with operations in the UK, the EU and the US uses machine learning for claims triage and a generative AI assistant for customer emails, and is piloting an AI agent that can update claim records.

  1. Policy: the board approves an AI policy that commits to the five OECD values and adopts the AI Act Article 3(1) definition for scope. An executive owner is named.
  2. Inventory and classification: each system is registered with the five OECD classification dimensions recorded (GOVERN 1.6, MAP 1.1). The agent is flagged because it acts on records, not only advises.
  3. Profile: the risk team writes an AI RMF profile per system type. The generative assistant also takes actions from NIST AI 600-1. Claims triage is reviewed by legal for AI Act risk class.
  4. Measurement: bias testing on triage outcomes (MEASURE 2.11, OECD 1.2), adversarial testing of the assistant and agent (MEASURE 2.7, OECD 1.4), and a system card for each (MEASURE 2.9, OECD 1.3).
  5. Override: the agent runs under its own workload identity with scoped permissions. A shutdown runbook revokes its credentials and is tested quarterly (MANAGE 2.4, OECD 1.4).
  6. Monitoring and incidents: post-deployment monitoring and incident communication follow MANAGE 4.1 and 4.3, with incidents classified under the AI Incident Reporting Guide.
  7. Reporting: one quarterly crosswalk report per OECD value answers both EU customer questionnaires and US AI RMF questions.

Implementing both in phases

  1. Phase 1, foundation (first 60 days): adopt the OECD values in policy, choose the scoping definition, name owners, build the AI register. Test: can every production AI system be found in the register with an owner?
  2. Phase 2, crosswalk: adapt the table above, agree evidence per row, and store it where audit can reach it.
  3. Phase 3, profiles: write AI RMF profiles by system type, set risk tolerance, and start MEASURE activities on the highest-impact systems first.
  4. Phase 4, operate: monitoring, shutdown drills, incident handling and feedback loops (GOVERN 5.1). Review the crosswalk when NIST publishes the AI RMF revision or the OECD updates its text.
  5. Rollback: if a profile proves too heavy, reduce it to the minimum set in the decision table rather than dropping the crosswalk; the policy layer stays unchanged.

Detection and measurement

  • Register coverage: percentage of AI systems found by discovery (procurement, cloud billing, API gateway logs) that are already in the register.
  • Crosswalk evidence freshness: share of crosswalk rows with evidence reviewed in the last 12 months.
  • Measurement coverage: percentage of high-impact systems with current fairness, robustness and security test results.
  • Shutdown readiness: percentage of agents and automated systems with a shutdown test passed in the last quarter, including credential revocation.
  • Incident loop time: median days from AI incident to updated risk record and communication to affected actors.

For a fuller metric set, see AI Governance Metrics.

Framework mapping

NeedNIST CSF 2.0NIST SP 800-53 Rev 5AI RMF
Laws that reuse OECD definitions are trackedGV.OC-03PM-9 Risk Management StrategyGOVERN 1.1
Named roles for AI riskGV.RR-02PM-9GOVERN 2.1
Policy states AI values and scopeGV.PO-01PM-9GOVERN 1.4
Risk assessment per systemGV.OV-01RA-3 Risk AssessmentMAP 5.1
AI suppliers assessedGV.SC-07SR-6 Supplier Assessments and ReviewsGOVERN 6.1

Common failure patterns

  • Two documents, no link: an OECD-flavoured policy and an AI RMF risk register owned by different teams, never reconciled. Fix with the crosswalk and one owner.
  • Stale definition: scoping with the 2019 wording and missing generative or adaptive systems that the November 2023 OECD definition and the AI Act include.
  • Treating framework adoption as compliance: neither framework produces AI Act conformity assessment, registration or role duties. See the NIST AI RMF vs EU AI Act.
  • Scoring subcategories as a checklist: yes or no answers with no tolerance, context or evidence, the opposite of how NIST describes the Core.

Practitioner checklist

  • Adopt the five OECD values as commitments in the AI policy, with an accountable executive.
  • Scope the AI register with the current OECD or AI Act Article 3(1) definition, not the 2019 wording.
  • Record the five OECD classification dimensions for each system at intake.
  • Write an AI RMF profile per system type and set risk tolerance explicitly.
  • Maintain a version-controlled crosswalk with one evidence pointer per row.
  • Give every AI agent a distinct identity and test credential revocation in shutdown drills.
  • Run fairness, privacy, robustness and security measurement on high-impact systems first.
  • Track laws that reuse OECD concepts in an obligations register under GOVERN 1.1.
  • Review the crosswalk when the AI RMF revision is published or the OECD text changes.

Standards and references

Related NHI Mgmt Group resources: NIST AI RMF Guide · OECD AI Principles Guide · AI Governance Guide · AI Governance vs AI Risk Management · EU AI Act vs ISO/IEC 42001

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 14 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.