ISO 42001 Clause 8 explained: operation

Everything up to this point in ISO 42001 has been preparation: context, leadership, a plan, and the support to run it. Clause 8, “Operation,” is where it all meets reality. This is the clause where you run the processes you designed in Clause 6, and, specifically, where you actually perform your AI risk assessment, risk treatment and impact assessment, not once at build time, but on a schedule and whenever things change. If Clause 6 is the recipe, Clause 8 is the cooking.

Clause 8 mirrors Clause 6

The structure of Clause 8 is deliberately a reflection of Clause 6:

• 8.1 Operational planning and control puts the plan into practice.
• 8.2 AI risk assessment performs the process defined in 6.1.2.
• 8.3 AI risk treatment performs the process defined in 6.1.3.
• 8.4 AI system impact assessment performs the process defined in 6.1.4.

The split is the point. Clause 6 asks whether you have designed these processes; Clause 8 asks whether you actually run them. A system can pass 6 on paper and fail 8 in practice, which is exactly the gap an auditor goes looking for.

8.1 Operational planning and control

Clause 8.1 requires you to plan, implement and control the processes needed to meet the system’s requirements and to carry out the Clause 6 actions. Concretely, that means:

• operating your AI systems and controls according to defined criteria;
• keeping documented information sufficient to show the processes were carried out as planned;
• controlling planned changes and reviewing the consequences of unintended ones, taking action where needed; and
• controlling externally provided processes, products and services that affect the system.

That last point matters more for AI than for almost anything else. Most organisations build on models and services they did not create, and 8.1 makes clear those dependencies are part of your operation: you have to manage them, not assume them away. This is where operation meets supplier and third-party assurance, you need enough visibility of external AI components to keep your own risk picture valid.

8.2 AI risk assessment

Clause 8.2 requires you to perform the AI risk assessment defined in 6.1.2 at planned intervals and when significant changes occur. The recurring-and-triggered pattern is the whole idea. AI risk is not static: a new system, a new use of an existing one, a model or data change, or a shift in the law or operating environment can all change the picture. So you assess on a regular schedule and whenever something material moves. You also retain the results as documented information, so the current risk position is always evidenced.

8.3 AI risk treatment

Clause 8.3 requires you to implement the AI risk treatment plan you produced under 6.1.3. Assessing risk and then not acting on the treatment is one of the most common and most damaging gaps: the risk is known, a treatment was decided, and it never got implemented. Clause 8.3 closes that loop, and, like 8.2, its results are retained as evidence. The controls listed in your Statement of Applicability are what actually get operated here.

8.4 AI system impact assessment

Clause 8.4 requires you to perform the AI system impact assessment from 6.1.4, again at planned times and when significant changes occur. Because AI’s effects on people can change as a system is used in new ways or on new populations, the impact assessment is a living activity, not a launch gate. The mechanics of running one are the same as at design time; Clause 8.4 is the requirement to keep doing it, and to feed what you learn back into treatment and oversight.

Common mistakes

• Designed but not performed. Processes defined in Clause 6 that are never actually run, the single most common way to pass 6 and fail 8.
• One-and-done assessment. Assessing risk or impact once at launch and never again, when 8.2 and 8.4 require intervals and change triggers.
• Treatment that stalls. Deciding on a risk treatment and never implementing it, so the risk stays live (8.3).
• Uncontrolled suppliers. Building on third-party models and services without controlling them, contrary to 8.1.
• No operational evidence. Running the processes but retaining nothing to show it, so the operation cannot be demonstrated in an audit.

The short version

Clause 8 is where ISO 42001 stops being a plan and starts being an operation. It asks you to put the Clause 6 processes into practice and keep them under control (8.1), and then to actually perform your AI risk assessment (8.2), your risk treatment (8.3) and your impact assessment (8.4), on a planned schedule and whenever significant changes occur, retaining evidence throughout. The recurring theme is that none of this is a one-time, build-time activity: AI systems, models, uses and rules change, so the operation has to be continuous. Design without operation is the gap Clause 8 exists to close, and it is the gap auditors most reliably find.

Running an AIMS in practice? The ISO 42001 guide walks the full standard, the checklist covers the controls you operate here, and our ISO 42001 consulting helps you turn a designed system into an operated one. For a quick baseline, the free AI governance check takes about ten minutes, no email.

Previous: Clause 7, support. Next: Clause 9, performance evaluation.