AI Governance That Works: An Operating Model, Not a Policy Document

Most organisations now have an AI policy. Far fewer have AI governance that changes what anyone does on a Monday morning. Meanwhile teams buy tools with AI features switched on, build assistants over internal documents, and connect agents to systems that move money. The gap between the document and the practice is where the risk lives.

Our stance is simple. An AI governance operating model is a set of roles, decisions, records and routines that run at the speed of delivery. The policy is only one output of it. This article sets out what it contains, how ISO/IEC 42001 and the NIST AI RMF can serve as scaffolding, and how a mid-sized firm can do it without building a bureaucracy.

Why policy documents fail

A policy states intent. It does not say who approves a new use case, what evidence they need, or what happens when a supplier changes the model behind a product you already use. When those questions have no answer, two things happen. Cautious teams stall, because nobody will say yes. Less cautious teams go ahead, because nobody is positioned to say no. Both are governance failures, often in the same organisation at the same time.

The test of governance is whether a delivery team can answer three questions quickly: who decides, on what evidence, and what must stay true after go-live.

The AI governance operating model: roles, inventory and risk tiers

Roles and decision rights

• An accountable executive. One member of the leadership team owns AI risk overall. A person, not a committee.
• A business owner for every system. The person whose process the AI supports owns its outcomes, including its errors. IT can operate a system, but should not own the decision it influences.
• A small review group. Risk, privacy or legal, security, data and a delivery lead meeting often enough that review is never the bottleneck.

Then write down who can approve a low-risk use case alone, what needs the review group, and what must go to the executive. If most decisions reach the top tier, the tiers are wrong.

The AI inventory

Most organisations underestimate what they run. The inventory should cover systems you build, systems you buy, and AI features embedded in software you already license. For each entry record the purpose, the business owner, the data it touches, the supplier and model behind it, who is affected, and the risk tier. Keep it light enough to stay current, and feed it from procurement, architecture review and security assessment.

Risk tiering

Tier by use, not by technology. The same language model is a minor risk when drafting meeting notes and a serious one when screening job applicants. Ask who is affected, how consequential the decision is, how much autonomy the system has, and how reversible an error is. If you operate in or sell into the EU, map your tiers to the categories in the EU AI Act, which entered into force in August 2024 with obligations phasing in over the following years. Our EU AI Act compliance playbook covers classification, and you should confirm current timelines and your own position with counsel.

Three tiers are usually enough: registration and standard guardrails for low risk, a documented evaluation and owner sign-off for medium risk, and full review with human oversight and scheduled reassessment for high risk.

Controlling what runs: evaluation, suppliers and agents

Evaluation and monitoring

Approval at launch is the weakest form of control, because AI systems change after launch. Before go-live, each system above the lowest tier needs an evaluation set built from real examples of the task, including the awkward ones, with pass criteria agreed by the business owner. Higher tiers add bias testing and adversarial testing of the kind described in our piece on red teaming your AI.

After go-live, monitor a few things that matter: sampled output quality, override and complaint rates, incidents, cost, and any change in the underlying model. Rerun the evaluation set whenever the model, the prompt or the data source changes.

Governance that cannot say yes quickly, and no quickly, will be bypassed by the people it was meant to guide.

Third-party and model supplier risk

Most AI in a mid-sized firm is bought, so supplier governance carries much of the load. Extend normal due diligence with AI questions. Is your data used to train their models, where is it processed, and how long is it kept? Will you be told when the underlying model changes? Which model providers sit behind the product? Can you export your data and configuration if you leave? Watch for AI features that arrive through routine updates to software you already own. They rarely pass through procurement.

AI agents that take actions

Agents that call tools, update records or send messages change the question: the risk is now what the system does, not only what it says. As organisations move agentic AI from pilots to production, governance needs specific controls:

• Identity and least privilege. Each agent has its own identity and only the permissions its task needs.
• Action limits. Thresholds on value, volume and scope, with human approval above them.
• Full action logs. Every tool call is recorded with inputs and outputs, so an incident can be reconstructed.
• A stop mechanism. Someone can pause an agent immediately and knows they have the authority.
• Untrusted inputs. Agents read emails, documents and web pages that may contain planted instructions. Treat that content as data, never as commands.

Using ISO/IEC 42001 and the NIST AI RMF as scaffolding

ISO/IEC 42001 is an international standard for an AI management system. It follows the management system pattern many organisations know from information security and quality: leadership commitment, policy, defined roles, risk assessment, operational controls, monitoring, internal audit and continual improvement, applied to AI. Organisations can be certified against it.

The NIST AI Risk Management Framework is a voluntary framework from the US National Institute of Standards and Technology. It organises AI risk work around four functions, govern, map, measure and manage, and describes the characteristics of trustworthy AI. It is not a certification scheme, but it is freely available and practical.

The two are complementary: one helps you think about risk, the other gives you the management system that holds the work together. Neither guarantees legal compliance. Use them as a checklist of what a complete system contains, and adopt what fits your risk profile.

Proportionality: a 90-day start for mid-sized firms

A firm of a few hundred people does not need an AI ethics office. It needs the essentials.

• Days 1 to 30: see and name. Appoint the accountable executive. Build the first inventory from procurement records and a short survey. Publish a one-page acceptable use guide.
• Days 31 to 60: tier and decide. Agree three risk tiers and the decision rights for each. Give the review group a fixed weekly slot. Tier everything in the inventory.
• Days 61 to 90: control and monitor. Put evaluation sets and monitoring in place for higher-tier systems. Add AI questions to supplier due diligence. Report the inventory and open risks to the board.

After that, governance is maintenance.

Common questions

What is an AI governance operating model?

It is the working system behind an AI policy: named roles, decision rights, an inventory of AI systems, risk tiers, evaluation and monitoring routines, and supplier controls. It tells a delivery team who approves what, on what evidence, and what must be maintained after launch.

Do we need ISO/IEC 42001 certification?

Not necessarily. Certification is worth considering if customers, partners or regulators ask for independent evidence, or if you sell AI-enabled products to large enterprises. Many organisations gain most of the benefit by aligning to the standard without certifying.

Who should own AI governance in a mid-sized company?

One accountable executive, often the COO, CIO or chief risk officer, supported by a small cross-functional review group. Each system also needs a business owner who answers for its outcomes. Avoid placing it solely in IT or legal.

Where Ganexa can help

Ganexa helps organisations design governance that delivery teams can work with. Our AI governance and compliance service covers the operating model and its alignment to ISO/IEC 42001 and the EU AI Act, and connects to our work on the AI business case and operating model so that governance and value are designed together. To talk through your current position, book a free consultation.