AI governance used to be something only big tech and regulators worried about. Now almost every organisation uses AI somewhere, often without a clear owner, and the rules are catching up fast.
The good news: you don't need to be a lawyer or a data scientist to start. Here's a practical starter framework, built on the three references most organisations are converging on.
The three references worth knowing
1. NIST AI Risk Management Framework (AI RMF 1.0)
Free, voluntary and vendor-neutral. It organises AI risk work into four functions:
- Govern: who is accountable, what the policies are, how decisions are escalated
- Map: what each AI system does, who it affects and what could go wrong
- Measure: how you test for accuracy, bias, robustness and other risks
- Manage: how you prioritise, respond to and monitor those risks over time
If you read only one thing, read this. The companion Playbook turns it into concrete actions.
2. The EU AI Act
The first comprehensive AI law, with obligations phasing in over several years. Its core idea is a risk-based approach:
- Unacceptable risk: banned (e.g. certain manipulative or social-scoring systems)
- High risk: strict requirements (e.g. AI used in hiring, credit, education, critical infrastructure)
- Limited risk: transparency duties (e.g. telling people they're talking to a chatbot)
- Minimal risk: largely unregulated
Even outside the EU, these categories are becoming the common way companies classify their AI use cases.
3. ISO/IEC 42001
The international management-system standard for AI. It follows the same structure as ISO 27001, so organisations that already run information-security management systems can extend what they have rather than starting from scratch.
A 5-step starter framework
Step 1: Build an AI inventory. List every AI system in use, including tools bought from vendors and features switched on inside existing software. You can't govern what you can't see. This step alone usually surprises leadership.
Step 2: Classify by risk. Use the EU AI Act tiers as a guide. Most systems will be minimal or limited risk. Focus your effort on the few that make or shape decisions about people.
Step 3: Assign owners. Every AI system needs a named business owner, not just an IT contact. They're accountable for how it's used and for responding when it goes wrong.
Step 4: Set proportionate controls. For high-risk systems: human oversight, documented testing (including for bias), clear records, and a way for affected people to challenge outcomes. For low-risk tools: a simple acceptable-use policy is often enough.
Step 5: Monitor and review. Models drift, vendors change products and laws change. Review high-risk systems on a schedule, and track incidents and complaints.
Where most organisations get stuck
- Governance as paperwork. Policies nobody reads don't reduce risk. Tie controls to real decisions and real owners.
- Ignoring vendor AI. Much of your AI risk sits in tools you bought, not ones you built. Ask vendors how they test and govern their models.
- All-or-nothing thinking. You don't need full ISO 42001 certification on day one. An inventory, risk tiers and named owners already put you ahead of most organisations.
Want to go deeper?
Our AI Ethics and Governance course walks through these frameworks with real-world scenarios, written for managers and professionals rather than engineers.
🎁 Free for the first 10 readers (expires November 17):
AI Ethics and Governance: free coupon
If the free spots are gone, you can still get it here:
AI Ethics and Governance on Udemy
Want to test your judgement? Try our AI Ethics & Governance: 100 Scenario-Based Questions practice set.
What's the hardest part of AI governance in your organisation? Share in the comments.
— Team SkillMint
Top comments (0)