By now you've built LLM apps and RAG systems (AIINFRA 300–301). This week we shift from building AI systems to governing them — and we do it through the three frameworks that show up in almost every job posting, RFP, and vendor questionnaire in this space: the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act. The single most important thing to learn this week is that these are not three flavors of the same checklist. They are three different kinds of things, and confusing them will cost you credibility with auditors, legal teams, and clients.
The NIST AI Risk Management Framework 1.0 (published as NIST AI 100-1 on January 26, 2023) is a voluntary framework — meaning nobody enforces it and there is no license or certificate at the end. It gives you a shared vocabulary and a repeatable method for thinking about AI risk, organized into four functions:
| Function | What it does | Example activity |
|---|---|---|
| GOVERN | Cross-cutting culture, policy, and accountability structures that touch the other three functions | Assigning an AI risk owner; writing an internal AI use policy |
| MAP | Establishes context: what is this system, who does it affect, what could go wrong | Documenting intended use, stakeholders, and foreseeable misuse |
| MEASURE | Analyzes, benchmarks, and monitors risks, including bias and fairness metrics | Running bias audits on model outputs; tracking drift |
| MANAGE | Prioritizes and responds to identified risks | Deciding to retrain, restrict, or retire a system based on measured risk |
Notice that GOVERN sits apart from the other three — it's the foundation that MAP, MEASURE, and MANAGE all depend on. As of this writing, there is no formal "AI RMF 2.0"; NIST's approach has instead been to add sector-specific Profiles (for example, generative AI) on top of the 1.0 core, rather than replacing the framework itself. The companion NIST AI RMF Playbook is where the framework becomes actionable: it offers concrete "suggested actions" under each function that you can select based on your organization's context — which is exactly what you'll do in this week's lab.
ISO/IEC 42001:2023 is the world's first certifiable AI management system standard. Like other ISO management-system standards (think ISO 27001 for information security), it follows a Plan-Do-Check-Act cycle and requires an organization to maintain documented policies, conduct AI risk and impact assessments, assign clear roles and responsibilities, and demonstrate continual improvement. It is still voluntary — no law requires it — but it is certifiable: an accredited third party can audit your organization and issue a certificate, the same way companies get ISO 27001-certified for security. A companion standard, ISO/IEC 42005:2025, adds detailed guidance specifically for AI system impact assessments, which pairs naturally with the MAP function in NIST's framework.
Unlike the other two, the EU AI Act is not voluntary at all — it is binding law that entered into force in 2024, and it reaches beyond the EU's borders (more on that below). It sorts AI systems into four risk tiers — unacceptable (banned outright), high-risk (heavily regulated), limited-risk (transparency obligations), and minimal-risk (largely unregulated) — and rolls out obligations on a staggered timeline:
| Date | Obligation takes effect |
|---|---|
| 2 Feb 2025 | Prohibitions on unacceptable-risk practices (Article 5) |
| 2 Aug 2025 | Obligations for General-Purpose AI (GPAI) model providers |
| 2 Aug 2026 | Full high-risk obligations: conformity assessment, CE marking, EU database registration |
| 2 Aug 2028 | Extended deadline (per the June 2026 Digital Omnibus) for high-risk AI embedded in already-regulated products |
Penalties are steep: up to €35 million or 7% of global annual turnover, whichever is higher, for the most serious violations (prohibited-practice breaches).
"These three are interchangeable AI compliance checklists." They are not — they differ in kind, not just scope. The EU AI Act is binding, extraterritorial law you must obey. ISO/IEC 42001 is a voluntary program you can choose to get certified against. NIST AI RMF is a voluntary day-to-day method with no certificate available at all. Treating them as three versions of the same document will lead you to under-invest in the one that's actually enforceable (the Act) or over-promise on the one that has no certification (NIST). "We should get 'NIST AI RMF certified.'" This phrase appears constantly in job postings and vendor pitches, and it describes something that does not exist. NIST does not certify organizations, products, or individuals against the AI RMF — there is no exam, no auditor, no certificate. If a vendor claims "NIST AI RMF certification," that is a red flag. Certification in this space belongs to ISO/IEC 42001, full stop. "The EU AI Act only applies to EU-based companies and only to the model provider." Also false. The Act is extraterritorial: it applies to any provider or deployer whose AI system affects people in the EU market, regardless of where the company is headquartered. It also explicitly covers deployers (organizations that use an AI system, even one built by someone else) in addition to providers (the organizations that build and place the system on the market). A U.S. company deploying a high-risk hiring tool used on EU-based candidates is squarely in scope. Holding these distinctions straight — one binding law, one certifiable standard, one voluntary method — is the foundation for everything else we build on in this course's governance unit.