The National Institute of Standards and Technology released its AI Risk Management Framework in January 2023, and organizations have been downloading PDFs ever since. The document is solid—comprehensive, authoritative, and built by experts who understand AI governance at scale. But here’s what I’ve observed in working with enterprises across finance, healthcare, and higher education: downloading the framework is not the same as implementing it.
Most teams stall at the same place: the framework is broad, abstract, and designed to apply across thousands of use cases. It doesn’t tell you what your governance meetings should look like, how to prioritize which AI systems to assess first, or what metrics actually matter for your business. You read through the four functions, realize they’re interdependent, and wonder where to start.
Let me walk you through how to move from “we have the NIST AI RMF document” to “we have a NIST AI RMF program that people actually follow.”
The four functions—and what they actually mean
NIST organizes AI governance around four core functions: Govern, Map, Measure, and Manage. They’re presented as a cycle, which is correct, but that visualization often obscures what each function is really asking you to do.
Govern: Build the structures and commitments
Govern is about institutional will. It’s asking: Do you have leadership commitment? Have you defined organizational goals for responsible AI? Do people across the business know what you care about?
In practice, this means:
- A documented AI governance policy that’s actually available to people who build and use AI systems
- Clear roles—who owns AI governance? Is it a dedicated team or distributed?
- An inventory of existing AI systems so you know what you’re actually governing
- Explicit values: what does “responsible AI” mean at your organization? Lower bias? Higher transparency? Reduced hallucinations?
The mistake here is making Govern too abstract. It’s not a philosophy conversation; it’s a governance decision. You’re answering: Which AI systems do we need to assess? What happens if one fails? Who decides if we deploy it?
Map: Understand your system and its context
Map is where you take a specific AI system and ask: What is it designed to do? Who does it affect? What could go wrong?
This isn’t threat modeling in the traditional security sense. It’s answering:
- What input data does the system use? Is it proprietary? Third-party? User-generated?
- Who are the stakeholders—employees, customers, candidates, patients?
- What decisions or actions does the system inform?
- What are the foreseeable harms? Not hypotheticals—realistic failure modes for your industry.
Organizations often rush this. They do a one-page risk assessment and move on. But the quality of everything downstream depends on how well you map here. If you miss a stakeholder or mischaracterize the decision the system supports, your risk controls will be misaligned.
Measure: Quantify the risks and controls
Measure is the technical layer. You’re measuring model performance, bias, robustness—whatever matters for your system.
But here’s where NIST and operational reality diverge. NIST suggests measuring dozens of things. Your enterprise can probably instrument 5 to 8 metrics well. So you prioritize:
- Accuracy and fairness metrics that directly map to your declared values from Govern
- Metrics tied to the specific harms you identified in Map
- Metrics you can actually collect and update over time
This is where a lot of organizations get stuck on LLMs. You measure hallucination rate, factuality, toxicity—but you also need to know: does the system make decisions that stick, or is it purely advisory? That changes what you measure and how much you worry about it.
Manage: Act on what you learn
Manage is governance in action. You’ve measured risks; now you decide: deploy as-is, deploy with guardrails, don’t deploy, or modify and remeasure.
Common patterns:
- Red flags trigger escalation. (e.g., bias above 5% triggers review by legal)
- Human review requirements scale with risk. A low-stakes recommendation system might have no human review. A hiring recommendation system has 100%.
- Incident response: what happens if the system fails in production?
- Monitoring: you don’t just measure once. You set up continuous monitoring and escalation thresholds.
Common implementation mistakes
Mistake 1: Treating it as a one-time exercise. The framework is cyclical for a reason. You measure, you learn, you adjust. If you’re not repeating Measure and Manage as systems evolve and more data comes in, you’re doing governance wrong.
Mistake 2: Measuring what’s easy instead of what matters. It’s easier to measure accuracy than fairness, hallucination rate than business impact. But a system can be accurate and still dangerous if it discriminates or fails at your stated values. Measure what you committed to in Govern.
Mistake 3: Skipping stakeholder engagement. Map requires understanding who the system affects. Too many organizations do this in a room with technologists only. You need product, operations, legal, and ideally someone who understands the downstream impact on users or customers.
Mistake 4: No escalation path. You measure, you find a problem, then what? If there’s no clear decision-making process, no escalation path, no way to say “this doesn’t deploy until we fix X,” you haven’t built a program—you’ve built a paperwork system.
Building a phased rollout plan
You can’t assess every AI system at once. So phase it:
Phase 1 (Month 1–2): Govern + inventory. Build your policy, define your values, and list every AI system in operation. You’ll probably find systems you didn’t know existed—legacy classifiers, embedded models, vendors running models on your data.
Phase 2 (Month 3–4): Map and prioritize. Map the systems that have the highest business impact or highest risk of harm. Rank them. You’re identifying which ones need urgent attention and which ones can wait.
Phase 3 (Month 5–6): Measure and assess. Deep dive into your top-priority systems. Instrument the metrics that matter. You’re not perfect—you’re probably missing things—but you’re establishing a baseline.
Phase 4 (Month 7+): Manage and scale. Make the go/no-go decisions. Implement controls. Set up continuous monitoring. Then repeat the cycle for your next tier of systems.
The key is consistency and momentum. A 50% implemented program across all systems beats a 100% implemented program on three systems.
Connecting NIST to existing enterprise risk management
NIST AI RMF doesn’t exist in a vacuum. You probably have enterprise risk management processes already—third-party risk assessments, compliance audits, security reviews. Map NIST to those.
Where does AI assessment fit in your vendor onboarding process? If a third party wants to deploy an AI system on your data, that triggers a Map + Measure. Where do you document that? Existing ISMS? Compliance platform? Security review workflow?
This isn’t about replacing your existing risk infrastructure. It’s about threading NIST in so that AI governance becomes a normal part of how you manage technology risk, not a separate initiative.
From download to program
The NIST AI RMF is well-designed. It’s comprehensive, flexible, and built by people who understand the problem space. But a document only matters if you implement it.
Start with Govern. Get leadership commitment. Define your values and your scope. Then move through Map and Measure systematically, with escalation paths and decision gates built in. And treat it as an ongoing cycle, not a compliance checkbox.
That’s the difference between downloading the framework and building a program that actually shapes how your organization builds and deploys AI.