AI Governance Frameworks Compared: EU AI Act, NIST AI RMF, ISO 42001 and 2026 Additions

Updated: November 2026

  • The EU AI Act is binding law, NIST AI RMF is voluntary guidance, and ISO 42001 is a certifiable management standard.
  • All three address risk assessment, human oversight, documentation and incident handling, so one set of controls can serve all three.
  • Document each control once, map it to every article, subcategory and clause it satisfies, and expand the mappings as new rules arrive.

The EU AI Act, NIST AI RMF and ISO 42001 share enough common ground that one governance program can satisfy all three. The overlap removes most duplicated work, but only if you design the program around shared controls from the start.

Trustible’s AI Compliance Frameworks module was built for this. You map your controls once and cover the EU AI Act, NIST AI RMF, ISO 42001 and 10+ other frameworks. This guide explains how the three compare, where they differ, and how to build one program that covers them.

What is an AI governance framework?

An AI governance framework is a structured set of policies, controls and processes that makes AI compliance repeatable, auditable and defensible. It differs from data governance, which covers data quality, lineage and privacy. An AI governance framework adds model-specific concerns such as bias testing, explainability, human oversight and regulatory compliance across the AI lifecycle.

Three frameworks appear together most often in enterprise programs. They come from different traditions, serve different functions and carry different obligations. But they overlap heavily, and that overlap is what makes a controls-based approach practical.

The three frameworks every organization should know

EU AI Act

The EU AI Act is a binding regulation. Under Article 2, it applies to providers that place AI systems on the EU market or put them into service there, wherever those providers are established. It applies to deployers established or located in the EU. And it reaches providers and deployers in third countries when the output of the AI system is used in the EU. The consolidated regulation text on EUR-Lex (version of July 27, 2026, including the AI Omnibus) is the most prescriptive AI-specific law in force anywhere.

The Act sorts AI systems into four risk tiers.

  • Unacceptable risk. Prohibited outright under Article 5. Examples include social scoring by public or private actors that leads to unjustified or disproportionate treatment, and real-time remote biometric identification in publicly accessible spaces for law enforcement, subject to narrow exceptions.
  • High risk. Includes AI used in hiring, credit decisions, critical infrastructure, education and law enforcement. These systems need a conformity assessment before deployment, ongoing monitoring, and detailed documentation of technical properties, training data and human oversight.
  • Limited risk. Article 50 transparency duties apply, including telling people when they’re interacting with an AI system such as a chatbot.
  • Minimal risk. No requirements beyond existing law.

Articles 9 through 15 set the requirements for high-risk systems. They cover risk management (Article 9), data governance (Article 10), technical documentation (Article 11), record-keeping (Article 12), transparency and information for deployers (Article 13), human oversight (Article 14), and accuracy, robustness and cybersecurity (Article 15). Providers must also put a quality management system in place under Article 17, and show that the system meets these requirements through a conformity assessment before it reaches the EU market.

Most Annex III systems use internal control, so the provider assesses itself and no notified body is involved. Biometric systems in Annex III point 1 are different. If the provider applied harmonised standards, it can choose internal control or a notified body. If it didn’t, or none exist, it must use a notified body (Article 43). Once the assessment is complete, the provider affixes the CE marking and registers the system in the EU database.

Provider and deployer obligations

The Act assigns different duties to providers and deployers.

Providers build AI systems or place them on the market. For high-risk systems, they carry the heaviest burden. They must run a quality management system (Article 17), keep technical documentation, complete a conformity assessment, register the system, monitor it after launch (Article 72) and report serious incidents (Article 73).

Deployers use those systems in their operations. Under Article 26, they must use the system according to the provider’s instructions, assign human oversight to people with the competence, training and authority to do it, monitor the system, keep the logs they control for at least six months, and inform workers before using a high-risk system in the workplace. A financial institution using an AI credit scoring tool is a deployer. It didn’t build the model, but it’s responsible for how the model is used, who oversees its decisions and whether affected people can contest them.

Under Article 25, a deployer becomes a provider if it puts its own name or trademark on a high-risk system, makes a substantial modification to one, or changes the intended purpose of a system so that it becomes high-risk.

Penalties

National authorities in each member state enforce the Act. Penalties are tiered under Article 99. In each tier the fine is the higher of the two figures, except for SMEs, which pay the lower.

ViolationMaximum fine
Prohibited practices35 million euros or 7% of global annual turnover
Other obligations, including high-risk system requirements and transparency duties15 million euros or 3% of global annual turnover
Incorrect, incomplete or misleading information given to authorities7.5 million euros or 1% of global annual turnover

Providers of general-purpose AI models face fines of up to 15 million euros or 3% under Article 101, imposed by the European Commission. For organizations with significant EU exposure, these figures are real.

Timeline, including the 2026 AI Omnibus

Regulation (EU) 2026/1744, the AI Omnibus, was published on July 24, 2026 and entered into force on July 27, 2026. It delayed most high-risk obligations. It did not delay prohibited practices, general-purpose AI obligations or the Article 50 transparency duties. It did soften the AI literacy duty in Article 4. Providers and deployers must now take measures to support AI literacy among their staff, instead of ensuring a sufficient level.

DateWhat applies
February 2, 2025Prohibited practices and AI literacy duties
August 2, 2025General-purpose AI model obligations
August 2, 2026Article 50 transparency duties and general application of the Act
December 2, 2026Two new prohibitions apply, covering AI systems that generate child sexual abuse material or non-consensual intimate content depicting identifiable people. Article 50(2) machine-readable marking also becomes mandatory for generative AI systems that were on the market before August 2, 2026.
December 2, 2027High-risk obligations for stand-alone Annex III systems
August 2, 2028High-risk obligations for AI embedded in regulated products

Generative AI faces two separate sets of duties. Providers of general-purpose AI models, including large language models, have had documentation and transparency obligations since August 2, 2025. Anyone deploying AI that generates or manipulates content must meet Article 50 labeling and disclosure duties. Track both separately from the risk-tier classification.

NIST AI RMF

The NIST AI Risk Management Framework is a voluntary U.S. framework built on four core functions.

  • Govern sets up policies, roles, accountability and culture.
  • Map identifies and contextualizes risks by examining intended use, stakeholders and potential harms.
  • Measure quantifies and evaluates those risks with metrics, testing and evaluation criteria.
  • Manage prioritizes and treats risks through controls, mitigations and response plans.

The functions form a continuous cycle, not a one-time exercise. Each contains categories and subcategories, such as GOVERN 2.1 (documented roles and responsibilities) or MAP 3.5 (human oversight processes). NIST’s AI RMF Playbook adds suggested actions for each. The framework itself is NIST AI 100-1, and the official framework page hosts it with the Playbook and crosswalks.

Is there a NIST AI RMF 2.0? No. Version 1.0 (NIST AI 100-1, January 2023) is still current. NIST’s framework page says AI RMF 1.0 is being revised as part of the White House AI Action Plan, and no successor has been published. Some third-party pages describe a 2.0, and none of them is an official NIST release. For generative AI, use the companion Generative AI Profile (NIST AI 600-1), published in July 2024, which names 12 risks unique to or exacerbated by generative AI and maps suggested actions to the four functions. NIST also released a concept note in April 2026 for a profile on trustworthy AI in critical infrastructure. See the NIST AI RMF FAQ for more.

The framework is voluntary, and its influence runs through other frameworks. Treasury’s Financial Services AI Risk Management Framework, released in 2026, adapts the NIST AI RMF for financial institutions. Buyers also use it as a baseline for vendor due diligence.

The framework is deliberately non-prescriptive. It doesn’t say which controls to implement. It gives a structure for thinking about risk and a common language for documenting governance decisions. That makes it a workable structural base for multi-framework compliance. You can use NIST as the structure and add EU AI Act obligations on top. NIST also hosts crosswalks to other frameworks, including one to ISO/IEC 42001 that was prepared against the final draft of the standard.

ISO 42001

ISO/IEC 42001 is a certifiable international management system standard for AI. The ISO catalog entry describes it as a framework for establishing, implementing, maintaining and continually improving an AI management system. NIST gives a risk approach and the EU AI Act sets requirements for specific systems. ISO 42001 certifies that the organization itself has the structures, processes and management systems to govern AI.

Like other ISO management system standards, it follows Plan-Do-Check-Act.

  • Plan covers AI policy, objectives, risk assessments and treatment plans.
  • Do puts the policies, controls and processes into operation.
  • Check covers internal audits, management reviews and performance evaluation.
  • Act covers corrective actions and continual improvement.

Organizations that already hold ISO 27001 or ISO 9001 will recognize this structure and can extend their audit infrastructure, documentation practices and review cadence to AI without building a parallel system.

Certification requires a Stage 1 and Stage 2 audit by an accredited certification body, followed by annual surveillance audits over a three-year certificate cycle. Published practitioner guidance puts preparation at 3 to 9 months from kickoff to certificate for most organizations, and up to 12 for larger or less mature ones. Auditors look for a documented AI policy with objectives, a risk assessment process that identifies AI-specific risks, risk treatment plans with controls mapped to the standard’s Annex A controls, evidence of internal audits and management reviews, and records of corrective actions.

Certificates are issued by bodies that audit against ISO/IEC 42001. ISO/IEC 42006, published in July 2025, sets the requirements for those bodies and builds on ISO/IEC 17021-1, so check that your certification body is accredited and works to it.

ISO 42001 vs ISO 27001

ISO 42001 isn’t a subset of ISO 27001. ISO 27001 covers information security management. ISO 42001 covers AI-specific risks that 27001 wasn’t designed to address, including bias, transparency, explainability and human oversight of automated decisions. Both follow Plan-Do-Check-Act, so an organization with 27001 certification will find the structure familiar. But 42001 adds new control domains, and an organization handling AI and sensitive data typically needs both.

Does ISO 42001 satisfy the EU AI Act?

Not on its own. ISO 42001 certifies your management system. The EU AI Act sets requirements for specific AI systems. Certification covers much of the groundwork, including risk processes, roles, documentation and monitoring. But 42001 isn’t a harmonised standard under the Act and doesn’t replace a conformity assessment. Map your 42001 controls to the Act’s articles to find the remaining gaps.

How do AI governance frameworks differ?

The three frameworks differ in legal force, geographic scope, certification, structure and enforcement.

DimensionEU AI ActNIST AI RMFISO 42001
TypeRegulation (law)Framework (guidance)Standard (certifiable)
Legal forceMandatory for the EU marketVoluntaryVoluntary (certification optional)
Geographic scopeEU, plus providers and deployers whose output is used in the EUU.S.-focused, globally referencedInternational
CertifiableNo certification scheme. High-risk systems go through conformity assessment and CE marking.NoYes (third-party audit)
EnforcementNational authorities, and the AI Office for general-purpose AI models. Fines up to 35 million euros or 7% of revenue.NoneCertification bodies, no direct penalties
StructureFour risk tiersGovern, Map, Measure, ManagePlan-Do-Check-Act
Best forAny organization deploying AI in or for the EUU.S. organizations wanting flexible risk guidanceOrganizations needing certifiable proof
Primary audienceLegal and compliance teamsRisk and technical teamsQuality and management system teams
Requires risk assessmentYes (Article 9)Yes (GOVERN 1.3, MAP 5.1)Yes (clause 6.1.2)
Requires human oversightYes (Articles 14 and 26)Yes (MAP 3.5)Yes (controls on responsible design and use)
Requires documentationYes (Article 11 and Annex IV)YesYes (clause 7.5)
Addresses incident handlingYes (Articles 26 and 73)Yes (MANAGE 4.3)Yes (communication of incidents control)
Key datesProhibitions Feb 2, 2025. GPAI Aug 2, 2025. Transparency Aug 2, 2026. Annex III high-risk Dec 2, 2027. Annex I Aug 2, 2028.Version 1.0 current, revision in progressCertify any time. Three-year certificate with annual surveillance audits.
Mapped in TrustibleYesYesYes

The biggest difference is mandatory versus voluntary. The EU AI Act is the only one of the three with direct legal penalties. NIST and ISO 42001 are voluntary, but voluntary doesn’t mean irrelevant. Sector frameworks build on NIST, and customers ask about ISO 42001 during third-party due diligence.

The frameworks also treat core governance topics differently.

  • Risk assessment. The EU AI Act prescribes risk categories with defined obligations per tier. NIST leaves categorization to the organization. ISO 42001 requires a documented process without dictating the taxonomy.
  • Documentation. The EU AI Act requires technical documentation on training data provenance, system architecture and accuracy metrics for high-risk systems (Annex IV). NIST’s expectations are principles-based. ISO 42001 requires documented policies, objectives, risk treatment plans and audit records without prescribing technical depth.
  • Human oversight. The EU AI Act requires high-risk systems to be designed so people can oversee them (Article 14) and places duties on deployers (Article 26). NIST addresses oversight through MAP 3.5 and GOVERN 3.2. ISO 42001 handles it through its controls on responsible design and use.

Certification is unique to ISO 42001. When a customer asks how they know your AI governance is real, a third-party certificate answers that question where self-attestation can’t.

Where the frameworks overlap

All three address risk assessment, human oversight, documentation of AI system properties, incident handling and accountability structures. This isn’t a coincidence. These are the foundations every major framework has reached independently.

The alignment goes beyond the surface.

  • Transparency. The EU AI Act requires providers to give deployers the information they need to interpret and use a high-risk system, including its capabilities and limitations (Article 13). NIST treats accountability and transparency as a trustworthiness characteristic (section 3.4) and covers its risks in MEASURE 2.8. ISO 42001 includes controls on system documentation and information for users and interested parties.
  • Accountability. The EU AI Act requires an accountability framework inside the provider’s quality management system (Article 17) and competent overseers at the deployer (Article 26). NIST’s Govern function sets accountability structures (GOVERN 2.1 through 2.3). ISO 42001 requires top management commitment and defined roles under clauses 5.1 and 5.3.
  • Incident management. The EU AI Act sets serious incident reporting (Article 73). NIST addresses incident communication and recovery in MANAGE 4.3. ISO 42001 includes a control on communication of incidents.

A single human oversight control, properly documented, can satisfy EU AI Act Articles 14 and 26, NIST AI RMF MAP 3.5, and the relevant ISO 42001 controls. A single risk assessment process generates evidence for all three. A technical documentation package covering architecture, data provenance, performance metrics and known limitations supports EU AI Act Annex IV, NIST’s Map and Measure evidence, and ISO 42001’s documented information requirements under clause 7.5.

What changed for these frameworks in 2025 and 2026

All three frameworks changed in the last 18 months, and two of the changes affect 2027 planning.

EU AI Act. The AI Omnibus (Regulation (EU) 2026/1744) entered into force on July 27, 2026. Stand-alone Annex III systems, such as hiring, credit and education tools, now have until December 2, 2027. AI embedded in regulated products has until August 2, 2028. Prohibited practices, general-purpose AI obligations and Article 50 transparency duties were not delayed, though generative systems already on the market have until December 2, 2026 for machine-readable marking. The Omnibus softened the AI literacy duty and adds two new prohibitions that apply from December 2, 2026.

NIST AI RMF. Version 1.0 remains current. NIST is revising it under the AI Action Plan, with no successor published. Use the Generative AI Profile (NIST AI 600-1) for generative systems.

ISO/IEC 42001. The standard hasn’t changed, but two supporting standards were published in 2025. ISO/IEC 42005 (May 2025) explains how to run the AI system impact assessment that 42001 clauses 6.1.4 and 8.4 require. ISO/IEC 42006 (July 2025) sets requirements for certification bodies.

Colorado. Colorado SB 26-189, signed May 14, 2026 and effective January 1, 2027, replaced SB 24-205. It dropped the risk management program and impact assessment requirements and moved to notices and consumer rights for automated decision-making technology. The Attorney General must adopt rules by January 1, 2027. See Colorado AI governance.

Which AI governance framework to start with

Start from regulatory exposure and business priorities, not from which framework is theoretically most complete.

If your organizationStart withWhy
Sells or deploys AI in the EUEU AI ActLegal mandate with fines up to 35 million euros or 7% of revenue
Is U.S.-based with no EU exposureNIST AI RMFFlexible, voluntary, and the base for sector frameworks such as Treasury’s FS AI RMF
Must prove AI governance to clients or partnersISO 42001The only certifiable option, with third-party validation
Operates globallyEU AI Act plus NIST AI RMFNIST gives the structural base and the EU AI Act adds legal requirements
Already holds ISO 27001ISO 42001Extends an existing management system

For most mid-size and large organizations in regulated sectors, all three eventually apply. Rules also keep diverging by jurisdiction. Colorado, for example, replaced its 2024 AI Act in 2026 with a narrower notice-and-rights law. The sequencing question is where to direct attention first, not which frameworks you can avoid.

A common sequence starts with the framework under the most business pressure, usually the EU AI Act for global enterprises or NIST for U.S.-focused ones. Build the controls library around that framework. Risk assessments, human oversight procedures, documentation practices and incident reporting become the foundation. When you add the second framework, you map existing controls to new articles and fill the gaps. The third follows the same pattern with a smaller gap.

A global enterprise with EU exposure might start with the EU AI Act and build controls for risk classification, conformity assessment, technical documentation and human oversight. Adding NIST then means mapping Govern to the organizational governance already in place. Pursuing ISO 42001 means formalizing and auditing a program that mostly exists.

How to satisfy multiple AI governance frameworks without duplicating work

Short answer. Map each internal control to every framework requirement it satisfies, document it once, and let updates to the control flow to every mapping.

The alternative is three trackers, one per framework, each with its own evidence folder and its own owner. Every audit means reconciling them by hand, and every new regulation means a fourth tracker.

We recommend a controls-based approach that maps internal controls to all three frameworks from the start. Build these controls first, in this order, because each serves all three frameworks.

  1. AI inventory and use case intake
  2. Risk assessment
  3. Human oversight
  4. Technical documentation
  5. Incident handling
  6. Vendor and model due diligence

Take risk assessment. An organization sets up a structured process for identifying, analyzing and evaluating AI-specific risks before any system enters production. That one process, documented once, supports EU AI Act Article 9, NIST GOVERN 1.3 and MAP 5.1, and ISO 42001 clauses 6.1.2 and 6.1.3. One process, one set of documentation, one audit trail, three frameworks. The same pattern applies to human oversight, incident handling, transparency documentation and accountability.

Trustible’s framework mapping builds on a use case intake process that captures the governance data needed to drive controls-based compliance. This front-loaded work pays off as rules change. When a new framework or sector standard arrives, adding it is a mapping exercise, not a documentation rebuild.

How Trustible helps

Trustible is a purpose-built AI governance platform. It maintains the mapping, evidence generation and reporting for the EU AI Act, NIST AI RMF, ISO 42001 and 10+ other frameworks from a single control library.

AI Compliance Frameworks holds bespoke control mappings across all active frameworks, maintained by Trustible’s AI policy and legal experts. A risk assessment satisfies Article 9, the relevant NIST subcategories and the relevant ISO 42001 clauses at the same time, and when a new regulation arrives the mappings are updated so you don’t start over. AI Inventory centralizes every AI use case, so one inventory feeds reporting for all three frameworks. Automated Workflows route use cases through the assessments each one needs. Reporting & Dashboards produces audit-ready evidence packages and shows readiness for each active framework.

Customers report 4X more AI use cases approved, 10X faster AI intake, a 60% reduction in governance cycle times and 100% audit-ready use cases.

Go deeper on each framework.

The question isn’t whether you need AI governance. It’s whether you can afford to run three frameworks by hand when one controls-based approach covers them together. The frameworks will keep multiplying. Your governance effort doesn’t have to.

Request a demo

FAQs

What is an AI governance framework?


A structured set of policies, controls and processes that ensures AI is built and used responsibly and in line with applicable rules. In practice, it’s what makes AI governance repeatable and auditable instead of ad hoc.

Which AI governance framework is best?


No single one. Start with the EU AI Act if you deploy AI in the EU, because it’s mandatory. Start with NIST AI RMF if you’re U.S.-based and need flexible risk guidance. Start with ISO 42001 if customers require certification. Most mid-size and large enterprises end up needing all three.

Are AI governance frameworks mandatory?


The EU AI Act is mandatory for any organization deploying AI in the EU market, with fines up to 35 million euros or 7% of global revenue. NIST AI RMF and ISO 42001 are voluntary, though buyers and regulators increasingly expect them.

Can you comply with multiple AI governance frameworks at once?


Yes. Map internal controls to every applicable article, subcategory and clause from the start, and satisfying one control updates your posture across all mapped frameworks.

What are the penalties for not complying with the EU AI Act?


Up to 35 million euros or 7% of global turnover for prohibited practices, up to 15 million euros or 3% for other violations including high-risk requirements, and up to 7.5 million euros or 1% for incorrect information to authorities. Non-compliant systems can also be pulled from the EU market.

How does ISO 42001 differ from ISO 27001?


ISO 27001 covers information security. ISO 42001 covers AI-specific risks such as bias, transparency, explainability and human oversight. It isn’t a subset of 27001, and organizations handling both AI and sensitive data typically need both.

What’s the difference between AI governance and data governance?


Data governance manages data quality, lineage, access and privacy. AI governance manages the AI lifecycle, including risk assessment, bias testing, explainability, human oversight and compliance. They’re complementary programs.

What are provider and deployer obligations under the EU AI Act?


Providers build systems or place them on the EU market. For high-risk systems, Article 16 requires a risk management system, technical documentation, human oversight by design, a conformity assessment, registration and post-market monitoring. Deployers use those systems. Article 26 requires following the provider’s instructions, assigning trained overseers, checking input data, monitoring, keeping logs, informing affected workers and reporting serious incidents. Under Article 25, a deployer becomes a provider by putting its name on a system, substantially modifying it or repurposing it for a high-risk use.

Did the EU delay the AI Act?


Partly. The AI Omnibus, in force since July 27, 2026, moved high-risk obligations to December 2, 2027 for Annex III systems and August 2, 2028 for AI embedded in regulated products. Prohibited practices, AI literacy, general-purpose AI obligations and Article 50 transparency duties were not delayed.

Does ISO 42001 certification satisfy the EU AI Act?


Not on its own. It covers much of the groundwork, including risk processes, roles, documentation and monitoring, but it isn’t a harmonised standard under the Act and doesn’t replace a conformity assessment. Map your 42001 controls to the Act’s articles to find the remaining gaps.

Is the NIST AI RMF mandatory?


No. It’s voluntary guidance, though federal agencies, regulators and enterprise buyers reference it. Financial institutions should also look at the Treasury’s FS AI RMF, which adapts NIST’s framework for financial services.

Is there a NIST AI RMF 2.0?


No. Version 1.0 is still current. NIST is revising the framework under the July 2025 AI Action Plan and hasn’t published a successor. Treat any “2.0” label as unofficial.

Does ISO 42001 require a third-party audit?


Only for certification. The standard requires internal audits and management reviews, and you can implement it without an external auditor. Certification requires a Stage 1 and Stage 2 audit by an accredited body, then annual surveillance audits over a three-year cycle.

What changed in Colorado’s AI law?


Colorado SB 26-189, effective January 1, 2027, replaced SB 24-205. It dropped risk management program and impact assessment requirements in favor of notices and consumer rights.

In this article

    AI Clarity Starts Here

    AI clarity is a growth strategy

    See how Trustible helps governance teams approve more AI, faster.