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

A.9 Use of AI systems

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

An HR department buys a tool that ranks CVs by fit. According to the provider, it is meant to support pre-selection, and a person should check every rejection. Three months later, rejections below a certain score go out automatically because it saves time. A.9 asks the organization to use AI systems according to defined rules and objectives, and to keep them within their intended purpose and their instructions for use.

A.9 is the control for deployers’ daily work. It applies equally to in-house and bought-in systems and answers three questions: by which rules does the organization decide to use a system? Which objectives does it pursue in using it? And how does it make sure the system is used the way it was meant to be?

What does the standard require?

Annex A.9 of ISO/IEC 42001 contains three controls. The implementation guidance sits in Annex B.9.

  • Processes for responsible use (A.9.2). The organization defines and documents the processes it follows when using AI systems, whether developed in-house or bought in. It gets clear about the considerations that drive the decision to use a system and addresses them in policies. Annex B.9.2 gives examples: required approvals, cost including ongoing monitoring and maintenance, approved sourcing requirements, and the legal requirements that apply. Existing usage rules for other systems may be incorporated.
  • Objectives for responsible use (A.9.3). The organization identifies and documents objectives that guide use, such as fairness, accountability, transparency, explainability, reliability, safety, robustness and redundancy, privacy and security, and accessibility. It puts mechanisms in place to achieve them, for example by assessing whether a third-party or an in-house solution suits the intended use.
  • Human oversight as an objective. Annex B.9.3 expects the organization to decide in which stages it builds in meaningful oversight: people reviewing outputs with authority to override, oversight wherever the instructions for use require it, monitoring of accuracy, reporting concerns about outputs and about changes in performance, and weighing whether automated decision-making is appropriate at all. Those who exercise oversight are informed or trained and know the system’s documentation. The impact assessment can show how much oversight is needed.
  • Intended use (A.9.4). The organization ensures that the system is used according to its intended uses and accompanying documentation. This may require resources, for example for the planned oversight, and input data that matches the documentation. Operation is monitored. If even correct use in line with the instructions gives cause for concern about affected people or legal obligations, the organization communicates this to the relevant people internally and to the supplier. Logs or other records show that the system is used as intended; how long they are kept depends on the intended use, internal rules and the law.

In practice

Keep a catalogue of approved services. Employees can see at a glance which AI services they may use for which purposes and with which data. Anything missing is requested. Shadow AI, meaning private tools used for work data, is reported and either brought into the catalogue or stopped.

Record purpose and limits per system. Take the provider’s intended purpose into the AI register and add what you use the system for and which uses are excluded. Extending the use, for example from pre-selection to automatic rejection, is a change that touches classification, impact assessment and oversight again.

Plan oversight realistically. Whoever reviews outputs needs time, training and the authority to deviate. Check in samples whether oversight actually takes place or whether recommendations are adopted unreviewed. High adoption rates without a single deviation are a warning sign of automation bias.

Count costs over the lifetime. Beyond licence and rollout, monitoring, oversight, training and maintenance cost money continuously. Leaving this out of the decision tends to mean saving on oversight later, of all places.

Typical audit evidence

For A.9, auditors typically expect:

KPI

Share of AI systems in use that entered the AI register through the approval process

The target is 100%. Systems that only surface during an inventory show that the approval process is being bypassed or is too cumbersome.

Supporting KPIs:

  • Share of high-risk systems with named and trained oversight
  • Number of uses found outside the intended purpose
  • Share of outputs changed or discarded by oversight, as an indication that review actually happens

A.9 largely corresponds to the deployer obligations under Article 26 of the AI Act (Regulation (EU) 2024/1689). Deployers of high-risk systems use them in accordance with the instructions for use (paragraph 1), assign human oversight to competent, trained people with the necessary authority (paragraph 2), ensure that input data under their control is relevant and sufficiently representative for the purpose (paragraph 4), monitor operation and report risks (paragraph 5) and keep the logs (paragraph 6). Before use at the workplace, they inform workers and their representatives (paragraph 7).

For all AI systems, the prohibited practices under Article 5, the duty to support the AI literacy of staff (Article 4) and the transparency obligations under Article 50 apply. Anyone who substantially modifies a high-risk system, or repurposes another system so that it becomes high-risk, is considered a provider under Article 25 (→ EU AI Act).

A.9 puts the rules into practice where AI actually does its work:

Sources

Frequently asked questions

Do I need an approval for every AI service?

For every new AI service used for work, yes. The approval can be lean: purpose, permitted data classes, a check against prohibited practices and the contractual position on data protection. What matters is that the service then appears in the catalogue of approved services and in the AI register.

What applies if we want to use a system differently from what the provider intended?

Then you are changing the intended purpose. Before deployment, revisit classification, impact assessment and oversight, and clarify with the provider whether the system is suitable. If the new purpose turns a system into a high-risk system, you take on the obligations of a provider under Article 25 of the AI Act.

How do I show that human oversight works?

Through records showing that outputs were reviewed and changed where needed, such as review notes, logs of overrides or sample evaluations. Oversight that never deviates draws attention in an audit. Training records and a description of the oversight staff's authority also help.

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.