Annex A Controls and the Statement of Applicability
ISO 42001 includes Annex A, which lists specific controls that organisations might use to manage AI risks. Annex A does not say you must use every control. Instead, it is a menu of controls to choose from. The idea is that different organisations face different risks and will select different subsets of these controls. Your job is to choose which controls fit your organisation and your risk appetite.
Annex A contains controls across several categories. Some controls address governance and processes. For example, one control is “establish an AI policy.” Another is “establish a process for assessing AI risks.” Some controls address data governance. One control is “document the origin and characteristics of training data.” Another is “establish a process for detecting and handling data quality issues.” Some controls address model development and testing. One control is “establish requirements for model validation and verification.” Another is “establish a process for detecting and mitigating model bias and drift.”
Some controls address AI system operation. One control is “establish a process for monitoring AI systems in production.” Another is “establish a process for managing AI-related incidents.” Some controls address human oversight. One control is “implement mechanisms for human review of AI decisions.” Another is “ensure people using AI systems understand their capabilities and limitations.”
Some controls address security and integrity. One control is “establish controls to protect model integrity.” Another is “establish a process for managing changes to AI systems.”
Some controls address documentation and transparency. One control is “establish requirements for documenting AI systems.” Another is “establish a process for disclosing when AI is being used.”

As you read through Annex A, you will see that many of these controls relate to things you have already learned about in Clauses 4 to 10. The Annex is not introducing new ideas. It is offering practical ways to implement the requirements of the main clauses. If Clause 6 requires you to assess risks, Annex A offers controls like “document the purpose, performance characteristics and limitations of each AI system.” If Clause 8 requires you to control AI system development, Annex A offers controls like “establish a process for testing model fairness” and “establish requirements for model documentation.”
A Statement of Applicability (SOA) is a document that says which controls you have decided to use and which ones you have decided not to use. For each control you use, your SOA says how you are implementing it. For each control you do not use, your SOA says why. The “why” might be: “This control is not applicable because our organisation does not build AI systems; we only purchase and deploy them” or “We considered this control but concluded that the risks it addresses are low in our organisation and our existing controls are sufficient.”
An SOA is important for several reasons. First, it makes your risk assessment and control decisions visible. A regulator or an auditor reading your SOA can see that you have thought through which risks apply to your organisation and how you are addressing them. Second, it protects you if something goes wrong. If an AI system causes harm and someone asks why you did not implement a particular control, your SOA allows you to point to documented reasoning. Third, it is an input to your internal audit. During an audit, you can check whether you are actually implementing the controls you said you would implement.
An SOA is often just a table. The table might have columns: Control ID (matching the control number in Annex A), Control Description, Applicable (Yes/No/Partial), Implementation Method (if applicable), and Justification (if not applicable). You fill in the table and then you keep it as a live document. As your organisation changes, you update your SOA. If you start a new line of business that uses AI differently, you might find that controls you had marked as not applicable are now applicable. If you retire a system, you might find that you no longer need a control.
The level of detail in your SOA should match your organisation’s size and complexity. A small organisation might have a one-page SOA. A large organisation with many AI systems might have a multi-page SOA or separate SOAs for different business units or AI systems.
Enterprise AI Readiness & Maturity Assessment
Benchmark your organization's data infrastructure, AI governance posture, and high-ROI implementation pathways.
