The sales team uses a chatbot for quotes, HR is trialling a tool that pre-sorts job applications, and IT is building its own assistant for customer service. Every department has its own rules, and none of them knows the rules of the others. A.2 asks for one AI policy that puts all of these projects under the same principles and fits with the rest of the policy set.
The AI policy is the governing document of the AI management system. It records what the organization uses AI for, where the limits are and who decides when a rule does not fit. Every other AI control builds on it: roles, impact assessments, the life cycle and suppliers.
What does the standard require?
Annex A.2 of ISO/IEC 42001 groups three controls. The implementation guidance sits in Annex B.2.
- A documented AI policy (A.2.2). The organization sets down in writing how it develops or uses AI systems. The policy draws on the business strategy, the organization’s values and appetite for risk, the risk posed by its own AI systems, legal and contractual duties, the wider risk environment and the effects on the people concerned. It contains guiding principles for all AI activities and says how deviations and exceptions are handled.
- Topics that need guidance of their own. Where it helps, the policy takes up individual topics or points to the policy that covers them. Typical examples are AI resources, impact assessments and the development of in-house systems.
- Alignment with other policies (A.2.3). AI touches quality, information security, occupational safety and privacy. The organization checks where existing policies overlap with the AI policy and either adjusts them or moves the rule into the AI policy. Direction from the governing body feeds into the AI policy.
- Review at intervals and when needed (A.2.4). The policy is reviewed on a schedule and whenever circumstances call for it, so that it stays suitable, adequate and effective. A role approved by management owns drafting, review and evaluation. Results of the management review flow into it.
In practice
Gather the inputs before you write. Put the AI register next to the business strategy: which systems are already running, which are planned, and how are they classified under the AI Act? Add management’s risk appetite and the duties from law and contracts. These ingredients decide how strict the policy has to be. An organization that only uses bought-in writing assistants needs different priorities from a provider of a high-risk system.
Settle exceptions before they arise. Decide who approves a deviation, what an exception has to state (system, rule affected, reason, compensating measure, deadline) and where it is recorded. There is no exception for the prohibited practices of Article 5 of the AI Act, and the text should say so explicitly.
Make overlaps visible. Go through the existing policies and mark every passage that touches AI: privacy for training and input data, information security for access to AI services, supplier management for bought-in models. Decide for each passage which document leads, and add a cross-reference in the other.
Name the role and set the date. Name the role that maintains the policy and put the next review in the calendar. Bring the results of the last management review into it: new objectives, incidents, feedback from interested parties.
Typical audit evidence
For A.2, auditors typically expect:
- Approved AI policy with date, version and sign-off by top management (→ AI policy in the AI Compliance Starter Kit)
- Derived rules of use that follow the principles of the policy (→ AI usage policy)
- Alignment with other policies, for example a list of cross-references to information security, privacy and suppliers
- Exception register with every approved deviation, its reason and its deadline
- Review records, including “no change required”, and the link to the management review
- Appointment of the responsible role for drafting and maintaining the policy
KPI
Share of AI systems in the AI register whose use matches the current, approved AI policy
This figure shows whether the policy actually steers the inventory. The target is 100%. Systems without a classification, without an impact assessment or with an expired exception pull the value down.
Supporting KPIs:
- Days since the last review of the AI policy (target: no more than 365)
- Number of open exceptions past their deadline
- Share of employees who have acknowledged the AI usage policy
Link to the AI Act
The AI Act (Regulation (EU) 2024/1689) does not ask for an “AI policy” by that name. Several duties do assume a settled line, though: the ban on certain practices in Article 5, AI literacy of staff under Article 4 and, for providers of high-risk systems, the quality management system under Article 17, which contains a strategy for regulatory compliance. An AI policy under A.2 is where that line is written down for the whole organization (→ EU AI Act).
Related controls
A.2 sets the frame for the other topics of Annex A:
- A.3 Internal organization: assigns the roles that put the policy into practice and sets up the channel for concerns.
- A.5 Assessing impacts of AI systems: delivers the policy’s promise to check the effects on people before use.
- A.5.1 Policies for Information Security and Privacy: the ISO 27001 counterpart the AI policy is aligned with.
Sources
- ISO/IEC 42001:2023, Annex A.2 and Annex B.2: Policies related to AI
- Regulation (EU) 2024/1689 (AI Act), Articles 4, 5 and 17
- ISO/IEC 42001 in the Cenedril wiki: structure of the standard and its relation to the AI Act