An online shop uses a chatbot to answer questions about deliveries. A customer only realizes after ten minutes that she is writing to a machine. When the bot gives her the wrong return deadline, she finds no way to report it. A.8 makes sure that people know what they are dealing with, can report problems and learn in good time what happened when an incident occurs.
A.8 looks outward. The other controls generate knowledge about an AI system: resources, impacts, limitations, incidents. A.8 governs who learns what, through which channel and when. The standard places weight on understandability: information has to be readable and easy to find for its audience, including people who rely on accessibility features.
What does the standard require?
Annex A.8 of ISO/IEC 42001 contains four controls. The implementation guidance sits in Annex B.8.
- Information for users (A.8.2). The organization determines which information the users of an AI system need and provides it. Depending on the context, this ranges from a simple notice that AI is involved, or the labelling of generated images, to technical details. Annex B.8.2 lists, among other things, purpose, how to use the system, how and when to override it, technical requirements and limitations including expected lifetime, the need for human oversight, accuracy and performance, relevant impact assessment results, revised claims about benefits, updates and maintenance, contact details and learning material. The criteria for selecting information are documented, for instance intended use, foreseeable misuse, the users’ expertise and possible impacts. The organization checks that the information is accessible, complete, current and correct.
- External reporting (A.8.3). Users and other external parties get a way to report adverse impacts of the system, such as unfair outcomes. This complements the organization’s own monitoring of operation.
- Communication of incidents (A.8.4). The organization sets out and documents a plan for informing users about incidents. Incidents can concern the AI system itself or information security and privacy, such as a data breach. Contracts and law can dictate which incidents must be reported, by when, to which authorities and with which details. AI incidents may be folded into general incident management as long as AI-specific requirements remain in view, for example separate notification duties when personal data in training data is affected.
- Reporting obligations towards interested parties (A.8.5). The organization determines and documents which information about the AI system it has to report to whom, such as customers or supervisory authorities. This can include technical documentation (with data sets, justification of algorithmic choices and test records), risks, impact assessment results and logs. The organization also knows the rules on disclosures to law enforcement authorities.
In practice
Match audiences to content. For each system, decide who needs which information: end users a clear AI notice and the route for complaints, operators usage instructions and ways to intervene, customers the limits of the domain of use, authorities the technical documentation. Record the mapping in the AI communication plan.
Offer an external reporting channel. A form on the website, an address in the instructions for use or a function in the system itself will do, as long as the route is easy to find and reports are handled. Route incoming reports into the same process as internal observations so that they are assessed and, where needed, treated as incidents.
Write the incident plan in advance. Anyone who starts working out whom to inform once the incident is under way loses time. For each type of incident, set who is notified, through which channel, within which deadline and with which core message. Align the deadlines with legal requirements, such as the 72 hours for data breaches under the GDPR and the deadlines for serious incidents under the AI Act.
Keep information current. Every significant change to the system triggers a review of the user information. If performance figures or benefit claims change, update the details and inform the people concerned.
Typical audit evidence
For A.8, auditors typically expect:
- User information per system with purpose, limitations, oversight and contact (→ AI usage policy in the AI Compliance Starter Kit, section 5)
- Documented selection criteria for the information provided
- External reporting channel with a description and evidence of handling
- Communication plan with audiences, channels, deadlines and content for incidents (→ AI communication plan)
- Template for reports to authorities (→ Serious incident report)
- Overview of reporting obligations towards customers and authorities (→ Legal compliance register)
- Instructions for use where the organization is a provider (→ AI provider policy, section 6)
KPI
Share of AI systems that reach users or affected people for which user information, an external reporting channel and incident communication are defined
The target is 100%. The figure shows whether every system that reaches people has the three basic building blocks of A.8. If one is missing, either the information, the way back or the crisis communication is left open.
Supporting KPIs:
- Average time from detecting an incident to informing users
- Number of external reports and the share answered within the target time
- Share of user information updated after the last significant change
Link to the AI Act
A.8 brings together several information duties of the AI Act (Regulation (EU) 2024/1689). Article 50 requires people to be told when they interact with an AI system, unless this is obvious, and artificially generated or manipulated content to be disclosed. Providers of high-risk systems supply instructions for use under Article 13 covering purpose, accuracy, limitations, oversight measures and maintenance. Deployers inform people about whom an Annex III high-risk system makes or supports decisions under Article 26(11), and on request explain the system’s role in a decision under Article 86.
Article 73 governs incidents: providers report serious incidents to the market surveillance authority within staggered deadlines of no more than 15 days, 2 days for widespread infringements or disruption of critical infrastructure, and 10 days in the event of a death. Deployers inform providers and authorities under Article 26(5). For data breaches, Articles 33 and 34 of the GDPR apply alongside (→ EU AI Act).
Related controls
A.8 carries knowledge from other controls to the outside world:
- A.3 Internal organization: the internal channel for concerns, which can be combined with the external one from A.8.3.
- A.5 Assessing impacts of AI systems: supplies the results that users and authorities learn about.
- A.6.2 AI system life cycle: produces the technical documentation the information is drawn from.
- A.5.24 Incident Management Planning and Preparation and A.5.26 Response to Information Security Incidents: the ISO 27001 counterparts into which AI incidents can be integrated.
Sources
- ISO/IEC 42001:2023, Annex A.8 and Annex B.8: Information for interested parties; clause 7.4
- Regulation (EU) 2024/1689 (AI Act), Articles 13, 26, 50, 73 and 86
- Regulation (EU) 2016/679 (GDPR), Articles 33 and 34
- ISO/IEC 42001 in the Cenedril wiki: structure of the standard and its relation to the AI Act