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

A.5 Assessing impacts of AI systems

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

A city administration introduces a system that pre-sorts applications for housing benefit. Technically everything runs smoothly and the error rate is within plan. Only months later does someone notice that applications from people with limited German are sent for manual review far more often. A.5 asks you to examine, before deployment and throughout the life cycle, what an AI system means for the people it affects, and to record the answer.

The impact assessment adds the perspective of affected people to the risk assessment. The risk assessment asks what threatens the organization’s objectives; the impact assessment asks what can happen to individuals, groups and society. Clauses 6.1.4 and 8.4 of the standard require the process, and Annex A.5 describes its building blocks.

What does the standard require?

Annex A.5 of ISO/IEC 42001 contains four controls. The implementation guidance sits in Annex B.5.

  • A defined process (A.5.2). Starting from the intended purpose and use, the organization examines whether a system touches people’s legal position or life chances, their physical or mental well-being, universal human rights or society at large. The process settles at least when an assessment is due (for instance when the purpose or context is critical, the technology complex and highly automated, the data sensitive, or when any of these change significantly), which steps it consists of (identify, analyse, evaluate, treat, document and report), who carries it out, how results feed into design and use, and which groups of people are considered. Data, the techniques applied and the way the whole system works all belong in the picture.
  • Document and retain results (A.5.3). Results are recorded, updated when needed and kept for a defined period. According to Annex B.5.3, useful content includes intended use and foreseeable misuse, positive and negative effects, foreseeable failures with countermeasures, the demographic groups concerned, system complexity, the role of humans including oversight, and effects on employment and skills.
  • Effects on individuals and groups (A.5.4). The organization takes its governance principles, AI policy and objectives into account, considers what users expect in terms of trustworthiness and pays attention to groups with particular protection needs, such as children, people with impairments, older people and workers. Depending on the purpose, fairness, accountability, transparency and explainability, privacy and security, health and safety, financial consequences, accessibility and human rights come into play. Where needed, it brings in experts, researchers or users.
  • Societal effects (A.5.5). The assessment looks at possible effects on environmental sustainability, the economy, government, health and safety, and norms, culture and values, for better or worse. Annex B.5.5 offers prompts such as the energy and water demand of training and operation, misuse for disinformation and the entrenchment of historical disadvantage.

In practice

Screen everything, assess in full where needed. A short questionnaire clarifies for each system whether it influences decisions about people, processes sensitive data or reaches vulnerable groups. A writing assistant for internal texts gets by with a brief justification, while a system that pre-selects job applications needs the full assessment. Both are documented, including the short version.

Name the affected people precisely. “Customers” is too broad. Who exactly is affected, which groups among them are particularly vulnerable, and who is affected indirectly without ever using the system? For applicant pre-selection, this includes the applicants and also the business units that work with the resulting shortlist.

Results have to trigger something. An impact assessment that changes no decision has no effect. Set thresholds: above a certain severity, the system goes live only after management approval and with countermeasures in place. Material effects are carried into the AI risk register as scenarios so that they are treated alongside the other risks.

Build on existing assessments. If a data protection impact assessment already exists for the processing, refer to it and add what it does not cover. For bought-in systems, the provider often supplies information on limitations and risks. Check it against your own context of use before adopting it.

Typical audit evidence

For A.5, auditors typically expect:

KPI

Share of AI systems in the AI register with a valid, approved impact assessment

Valid means approved, renewed after the last significant change and younger than the defined repetition interval. The target is 100%. The figure shows whether the triggers actually work in day-to-day operation.

Supporting KPIs:

  • Number of systems whose assessment is still outstanding after a significant change
  • Share of effects above the threshold for which countermeasures are in place
  • Average time from trigger to approval of an assessment

The AI Act (Regulation (EU) 2024/1689) has two related duties. Certain deployers of high-risk systems carry out a fundamental rights impact assessment under Article 27 before first use and notify the market surveillance authority of the result. This covers bodies governed by public law, private entities providing public services, and deployers of systems for creditworthiness checks and for risk assessment and pricing in life and health insurance. Providers of high-risk systems run a risk management system under Article 9 that considers risks to health, safety and fundamental rights, including the effect on minors and other vulnerable groups.

An impact assessment under A.5 can carry both duties if it contains the required details. Article 27(4) allows the assessment to build on an existing data protection impact assessment under Article 35 of the GDPR (→ EU AI Act).

A.5 is closely tied to the controls that provide its facts or carry its results forward:

Sources

Frequently asked questions

Is the impact assessment under A.5 the same as the fundamental rights impact assessment under Article 27?

The two overlap but have different scopes. Article 27 applies only to certain deployers of high-risk systems, such as public bodies and deployers of systems for creditworthiness checks or for risk assessment in life and health insurance. A.5 applies to every AI system in scope, whatever its risk class. A single procedure can cover both if it includes the details required by Article 27(1).

How long do impact assessments have to be kept?

The standard asks for a defined period without naming a figure. Your own retention schedule and legal requirements decide. For high-risk systems it makes sense to align with the ten-year period for technical documentation under Article 18 of the AI Act.

Can I reuse a data protection impact assessment?

Yes, as a building block. Annex B.5.2 expects the organization to check whether assessments from other disciplines capture the AI aspects well enough. A DPIA under Article 35 GDPR covers privacy risks, while fairness, accessibility or societal effects usually appear only at the margins; you add those parts.

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.