Zum Hauptinhalt springen
ISO/IEC 42001 · Annex A · AI control

A.3 Internal organization

Updated on 6 min Reviewed by: Cenedril Editorial
ISO/IEC 42001EU AI Act

A customer service chatbot gives a customer wrong information about her contract. The business unit points to IT, IT points to the vendor, and the privacy team says it is not their area. An employee who has known about the problem for weeks has no idea whom to tell. A.3 makes sure that every AI task has someone who acts on it, and that concerns travel along a route that actually reaches somebody.

Annex A.3 is the organizational control of the AI management system. It carries the roles from clause 5.3 over into the AI topics that really need an owner, and adds a reporting channel that brings problems to light early. Without this groundwork, the policy, the risk assessment and the impact assessment stay on paper, because nobody answers for them.

What does the standard require?

Annex A.3 of ISO/IEC 42001 contains two controls. The implementation guidance sits in Annex B.3.

  • Define and assign AI roles as needed (A.3.2). Which roles exist depends on what the organization needs. When handing them out, it draws on its AI policy, its AI objectives and the risks it has identified, so that no area is left without an owner. It may set its own priorities for the order of assignment.
  • Typical areas. Annex B.3.2 lists examples: risk management, impact assessment, assets and resources, information security, safety, privacy, development, performance, human oversight, supplier relationships, showing lasting legal compliance, and data quality across the whole life cycle.
  • Describe responsibilities in enough detail. Each role is spelled out precisely enough for the person holding it to perform the tasks.
  • A route for concerns (A.3.3). The organization sets up and runs a process through which concerns about its role in relation to an AI system can be raised, from planning to decommissioning.
  • Eight functions of the channel. Annex B.3.3 describes what the channel should deliver: confidentiality, anonymity or both at the reporter’s choice; access and promotion for employees and contracted persons; staffing with qualified people; powers to investigate and resolve; timely reporting and escalation to management; effective protection from reprisal for reporters and investigators; reports that follow the investigation and escalation rules while preserving anonymity and business confidentiality; and a response within a reasonable time. Existing channels may be used.

In practice

A list of AI tasks is the starting point. Take the twelve areas from Annex B.3.2 and strike out whatever does not apply to you. An organization that only runs bought-in systems usually drops development, while supplier management and human oversight stay central for it. For every remaining area, decide who decides, who executes and who is kept informed.

Attach roles to existing functions. Many tasks already have a home. Information security governs access to models and interfaces, privacy checks the legal basis for training and input data, and procurement manages suppliers. For each of these interfaces, settle who has the final say on an AI system and who is only consulted. A RACI matrix shows this on a single page.

Name the people who exercise human oversight. For systems whose outputs affect people, and especially for high-risk systems, a job title rarely suffices. Name the individuals, describe their authority to intervene and to stop the system, and make sure they are trained and have time for the task.

Promote the channel and try it out. A channel nobody knows about falls short of B.3.3, which expects it to be promoted. Point to it in the AI usage policy, in onboarding and in contracts with service providers. Set target times for acknowledgement and response. An annual test report has proven useful to check whether a report reaches the right people and whether escalation works.

Typical audit evidence

For A.3, auditors typically expect:

KPI

Share of AI systems in the AI register for which every required role is filled by a named person

This figure shows whether the distribution of roles covers the actual inventory. The target is 100%. Typical gaps are new systems without a risk owner and high-risk systems without named oversight.

Supporting KPIs:

  • Average time to acknowledge a report (target: no more than 7 days)
  • Share of reports closed within the set deadline
  • Number of AI roles without a named deputy

The AI Act (Regulation (EU) 2024/1689) distributes duties by role, such as provider, deployer, importer and distributor, and assumes that the organization passes them on internally. Deployers of high-risk systems assign human oversight under Article 26(2) to people with the necessary competence, training and authority. Providers describe an accountability framework in their quality management system under Article 17(1)(m), setting out the responsibilities of management and staff. The duty to support the AI literacy of staff (Article 4) is hard to meet in a targeted way without clear roles.

Article 87 matters for the reporting channel: it makes the Whistleblower Directive (EU) 2019/1937 applicable to reports of infringements of the AI Act. In Germany, the Whistleblower Protection Act (HinSchG) transposes that directive, and many organizations already run an internal reporting office under it. A.3.3 goes further, because a concern can be raised even when no law has been broken (→ EU AI Act).

A.3 connects the principles of the policy with the people who put them into practice:

Sources

Frequently asked questions

Do I need an AI officer?

The standard prescribes no particular job title. It asks for AI roles to be defined and assigned according to the organization's needs, supported by clause 5.3. A coordinating role that maintains the AI register and acts as the point of contact has proven useful; in small organizations it often sits with privacy or information security.

Can the internal whistleblowing channel under the German HinSchG also take AI concerns?

Yes, Annex B.3.3 explicitly allows existing reporting channels to be reused. Check that the channel meets all eight functions, above all access for contracted persons and timely escalation to management. Also make sure it accepts concerns that involve no breach of the law.

How detailed do the responsibilities have to be?

Detailed enough for the person to carry out their duties with them; that is the yardstick in Annex B.3.2. A single line such as 'responsible for AI' rarely suffices. Tasks, decision rights, the AI systems concerned and a deputy per role have proven to work.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.