Picture a company that buys software to rank job applicants. Its risk team runs a NIST review, its AI management system is certified to ISO/IEC 42001, and its policy cites the OECD principles. Even so, none of that answers the first question a regulator would ask: does EU law treat this tool as high-risk?
The four instruments do different jobs, and swapping one for another leaves companies holding paperwork that proves less than they think. Below, you’ll find what each one does, how they fit together, and what changed in the EU this year. Dates are accurate as of 21 November 2026.
Governance versus the instruments that feed it
Your AI governance is the internal system of policies, roles, controls and reviews you run every day. External frameworks, standards and laws shape that system, while regulation sits outside it as the legal layer. To keep the two ideas apart, this article calls OECD, NIST, ISO, and the AI Act “instruments.”
| Instrument | Legal status | Use it to | Typical output | What it can’t prove |
|---|---|---|---|---|
| OECD AI Principles | Non-binding policy principles | Set values and commitments | A written policy statement | Legal compliance |
| NIST AI RMF | Voluntary guidance | Run risk work on one system | Risk assessment, owner, evaluation plan, deploy decision | Certification or legal compliance |
| ISO/IEC 42001 | Voluntary standard, optional certification | Build and improve a management system | Records of ownership, approvals, monitoring, review | Product approval or EU AI Act compliance |
| EU AI Act | Binding EU regulation where it applies | Determine legal duties | Classification memo, conformity documents where required | That voluntary practices are enough |
Start with the system

Record what the system does, which jurisdictions it touches, and whether your company acts as provider or deployer. Then pick the guidance that helps. The hiring tool makes a useful example (an illustration, not a compliance checklist):
| Step | What the team produces |
|---|---|
| Check legal scope | An inventory entry and classification memo: intended use, where it runs, whether rankings influence hiring, provider or deployer |
| Apply NIST to the risk work | A named owner, a risk assessment (inaccurate rankings, disparate effects), an evaluation plan, a deploy-or-restrict decision |
| Add ISO/IEC 42001 if buyers want assurance | Records showing how this use case runs inside the management system |
| Use the OECD principles for policy | A commitment, for instance, that applicants can challenge a decision |
Why hiring tools get extra scrutiny
In this example, the tool ranks applicants and recruiters review only those who make the shortlist. Hiring tools draw attention for a reason: AI ranks job applicants by statistical likelihood of success, and these systems can learn and amplify past hiring patterns. A tool’s intended purpose and how its output influences hiring are what make classification worth checking before purchase.
Where the Article 6(3) exception stops
Annex III, point 4(a), covers AI tools used for hiring, including tools that sort and screen applications or assess candidates, and Article 6 governs whether a particular system is high-risk. Article 6(3) offers a narrow exception. Even so, an Annex III system that performs profiling of natural persons is always high-risk. A provider that relies on the exception must document its assessment before placing the system on the market or putting it into service, and must register under Article 49(2). The Digital Omnibus text kept that duty and simplified the information providers submit, so the Omnibus didn’t create the requirement.
Whether a ranking tool profiles people is a legal call, so settle it in the classification memo before anyone buys the software. Meanwhile, evidence travels between steps: a risk assessment built for NIST can feed an ISO management review. The legal determination still needs its own memo.
The OECD AI Principles: values on paper
The OECD Council adopted its Recommendation on AI on 22 May 2019, the first intergovernmental standard on AI, and the principles were updated in May 2024. The OECD AI Principles give you a defensible statement of values. Still, they carry no legal duty.
They work best once they land in a written company AI usage policy with a named owner. There, a line like “affected applicants can challenge a decision” becomes something a manager can enforce.
NIST AI RMF: the working risk method
NIST released AI RMF 1.0 on 26 January 2023 for organizations that design, develop, deploy or use AI. Its core has four functions: govern, map, measure, and manage. Govern runs through the other three, and measure covers assessing, analyzing, and tracking the risks you identified. Teams apply the functions iteratively across the lifecycle, so they don’t run as a fixed sequence.

The framework carries no enforcement and relies on organizational commitment. Also, NIST’s AI RMF page says AI RMF 1.0 is being revised. For generative systems, teams can pair it with the Generative AI Profile (NIST AI 600-1, released 26 July 2024).
ISO/IEC 42001: a management system, audited
What certification covers
ISO and IEC published the standard in December 2023. It specifies how to establish, implement, maintain, and continually improve an AI management system. Certification is voluntary and covers that system, so no product or service earns approval through it. Before relying on a certificate, read its stated scope.

Where EN 18286 fits
ISO/IEC 42001 certification does not by itself create a presumption of conformity under the AI Act. Instead, Article 40 grants that presumption through harmonised standards whose references appear in the Official Journal. In October 2026, the Commission cited the first one, EN 18286:2026, through Implementing Decision (EU) 2026/2274, which appeared in the Official Journal on 14 October 2026. Providers that conform to EN 18286:2026 benefit from a presumption of conformity, but only for the Article 17 requirements the standard covers. That presumption does not establish conformity with separate system requirements such as risk management or cybersecurity.
The Commission’s explanation of the standard adds that it is meant to sit on top of existing standards such as ISO/IEC 42001, not replace them. In short, ISO/IEC 42001 can support your management system, but it can’t replace assessing the requirements that apply to a particular AI system.
The EU AI Act: what applies now and what moved
What applies and when

The Act entered into force on 1 August 2024. Then the Digital Omnibus on AI (Regulation (EU) 2026/1744) reached the Official Journal on 24 July 2026 and entered into force on 27 July. The Commission’s announcement confirms the new high-risk dates.
| Date | What applies |
|---|---|
| 2 February 2025 | AI literacy and the original prohibited-practice rules begin to apply |
| 2 August 2025 | General-purpose AI obligations begin to apply; models already on the market before that date have until 2 August 2027 |
| 2 August 2026 | Article 50 transparency duties generally apply; see the limited marking transition below |
| 2 December 2026 | Providers of covered generative systems placed on the market before 2 August 2026 must meet the Article 50(2) machine-readable marking duty; two new bans take effect on AI practices that generate or manipulate non-consensual intimate material or child sexual abuse material |
| 2 December 2027 | High-risk rules for Article 6(2)/Annex III systems (originally 2 August 2026) |
| 2 August 2028 | High-risk rules for Article 6(1)/Annex I systems (originally 2 August 2027) |
What the Omnibus changed besides dates
The Article 111(4) transition applies only to Article 50(2). It does not delay the separate Article 50(1) disclosure duty for systems intended to interact directly with people, where that duty applies. Also, the Omnibus narrowed “safety component”: AI used solely for non-safety purposes no longer counts as one, but systems whose failure or malfunction would endanger health and safety still do.
Because the Commission’s consolidated pages can lag an amending regulation, check the Official Journal text of Regulation (EU) 2026/1744 before you plan around a date.
A note on US jurisdictions
On 11 December 2025, an executive order directed the Justice Department to build a task force to challenge state AI laws, and the Attorney General established it by memo on 9 January 2026. The order did not itself invalidate state laws.
Colorado shows why dates need care. SB 26-189 generally takes effect on 1 January 2027 and applies to consequential decisions made on or after that date. Separately, an April 27 federal order restricts the attorney general from initiating enforcement, including investigations, for certain alleged violations, running through 14 days after the court rules on the preliminary-injunction motion. That order doesn’t move the statutory date. The bill text has the current language.
Owners, evidence and review triggers
Every instrument leaves operating detail to you, so three questions turn any of them into a working control.
Who owns it? Name one accountable person per AI system. A committee diffuses the blame, whereas a named owner gives the board something concrete to hold accountable.
What proves the control ran, and what proves it worked? An approval log shows who signed off, and an evaluation report shows whether the system met its acceptance criteria. Where human review applies, an audit trail for human review should show who reviewed the decision, what they considered, and whether they could change the outcome. Keep only the personal data that purpose needs.
What restarts the assessment? A new use case, a vendor or model change, an incident, or entry into a new jurisdiction should each reopen it.
FAQs
Q. Which instrument should a first-time team start with?
Start with the legal question: where does the system operate, and does it fall under a law? Next, add NIST for risk work, and bring in ISO and the OECD principles when policy and assurance needs appear.
Q. Do we need ISO/IEC 42001 certification?
Certification is voluntary. Consider it if customers, procurement requirements, or your organization call for independent assurance about the management system.
Pick one AI system and record its intended use, jurisdiction, and your organization’s role before deciding which framework to apply.
Related: AI Transformation Is a Governance Problem (Not Tech) — 2026 Truth
| Disclaimer: This article is for informational purposes only and does not constitute legal or regulatory advice. AI laws, standards, and compliance requirements may change. Organizations should verify current requirements through official sources and consult qualified legal professionals before making compliance decisions. |
