AI Governance: A Practical Guide for Security Leaders

What is AI governance?

AI governance is the set of policies, roles, processes and evidence that control how an organization chooses, builds, buys, uses and retires AI systems. In plain terms, it answers five questions:

• What AI do we use? That is the inventory.
• Who is accountable for each system? That is ownership.
• What could go wrong, and how bad would it be? That is risk.
• What rules apply, and are we following them? That is policy and compliance.
• Can we prove it? That is evidence.

It is not a ban on AI and it is not a one-time document. It is an operating program, like vulnerability management or third-party risk.

Why it lands on the security leader

AI risk overlaps with work security leaders already own.

• Data exposure. Staff paste customer data, code and contracts into public AI tools.
• Third-party risk. Vendors keep adding AI features, often switched on by default.
• New attack surface. Prompt injection, model manipulation, insecure plugins and over-permissioned AI agents.
• Regulatory and contractual pressure. Customers add AI clauses to questionnaires, and laws such as the EU AI Act set obligations.
• Board questions. Directors want to know the organization's exposure and who is in charge.

Someone has to own this. In most organizations security, risk and compliance share it, and the security leader is asked first.

The frameworks that matter

You do not need to adopt all of them. Pick a primary framework and map the others to it.

ISO/IEC 42001:2023, the management system standard

ISO/IEC 42001 specifies requirements for an AI management system. It follows the same high-level structure as ISO 27001, so organizations already working to ISO 27001 can extend processes they have. It is certifiable, which helps when customers want independent assurance. It suits organizations that sell to enterprises. Our executive guide to ISO/IEC 42001 goes deeper.

NIST AI Risk Management Framework (AI RMF 1.0)

A voluntary US framework organized around four functions: Govern, Map, Measure and Manage. It focuses on identifying and managing AI risk across the lifecycle and is not certifiable. It suits teams that want a practical, risk-based structure, especially those already using NIST CSF.

EU AI Act

A binding EU regulation that sorts AI systems by risk level and sets obligations for each category, including prohibited practices, high-risk systems and transparency duties. Provisions apply in phases, and it can reach organizations outside the EU whose AI is used in the EU. Treat it as a legal question and confirm current dates and scope with counsel.

How they fit together

• ISO/IEC 42001 is a certifiable standard. Use it to prove that a management program exists.
• NIST AI RMF is a voluntary framework. Use it to structure the risk work.
• EU AI Act is law. Use it to identify the legal duties that apply to you.

A sensible approach is to build the program on one backbone, map controls to the others and reuse the same evidence for each.

The eight building blocks of an AI governance program

1. AI inventory

List every AI system and AI-enabled tool: internal models, vendor AI features and tools staff use on their own. Record the owner, purpose, data used and vendor. You cannot govern what you cannot see.

2. Roles and accountability

Name an executive owner and a cross-functional group covering security, legal, privacy, data and the business. Assign an accountable owner to every system.

3. Policies

At minimum, write an AI acceptable use policy, a data handling rule for AI and an approval process for new AI tools. Keep them short enough that people read them.

4. Risk assessment and classification

Assess each system by impact: what data it touches, what decisions it influences, who is affected and what failure looks like. Classify systems into tiers so effort matches risk.

5. Third-party and vendor AI review

Add AI questions to vendor due diligence: where data goes, whether it trains models, what controls exist and how incidents are reported.

6. Technical and security controls

Apply access control, logging and data loss prevention, test for misuse such as prompt injection, and limit what AI agents may do on their own.

7. Monitoring and incident response

Watch for new tools, drift and misuse. Extend incident response to cover AI failures and data leakage.

8. Evidence and reporting

Keep records that show the program works: approvals, assessments, reviews and exceptions. Report status to leadership on a regular schedule.

A 90-day plan

Days 1 to 30: see and own

• Name an owner and form the working group.
• Build the first AI inventory, including shadow AI.
• Publish an interim acceptable use policy.

Days 31 to 60: assess and decide

• Classify systems by risk tier.
• Assess the highest-risk systems first.
• Add AI questions to vendor reviews.
• Choose a primary framework.

Days 61 to 90: operate and prove

• Set up approval and exception workflows.
• Start monitoring and add AI to incident response.
• Collect evidence and produce the first leadership report.

Common mistakes

• Banning AI. Staff use it anyway and you lose visibility.
• Writing policy without an inventory. The rules do not match reality.
• Treating it as a one-time project. Tools and rules keep changing.
• Leaving it to one team. Security alone cannot judge the legal, ethical and business impact.
• Collecting no evidence. If you cannot show it, an auditor or customer will treat it as not done.

My CISO Partner's perspective

Start small and make it real. An owner, an inventory, a short policy and a first risk review beat a large framework nobody operates. Size the program to how much the organization actually depends on AI today, and add structure as that dependence grows.

Readiness results are an assessment of your program. They are not a certification or a regulatory determination, and this guide is general information, not legal advice.