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

A.6.2 AI system life cycle

Updated on 7 min Reviewed by: Cenedril Editorial
ISO/IEC 42001EU AI Act

A retailer uses a model that predicts returns and steers purchasing. After a vendor update, the hit rate drops, and nobody notices: no metric is being watched, and no log shows since when the system has been calculating differently. A.6.2 defines what has to be in place in every stage of an AI system, from requirements through testing and deployment to operating rules, technical documentation and event logs.

A.6.2 is the largest control group in Annex A. It covers in-house developments as well as bought-in systems, because anyone who deploys a third-party model still plans its rollout, monitors its operation and logs its use. For providers, requirements, design and validation weigh considerably more.

What does the standard require?

Annex A.6.2 of ISO/IEC 42001 contains seven controls along the life cycle. The guidance in Annex B.6.2 is extensive; here are the key points.

  • Requirements and specification (A.6.2.2). For new systems and material enhancements, the organization records the requirements, including the rationale (for example a business case, a customer request or a government mandate) and how the model can be trained and the data needs met. Requirements span the whole life cycle and are revisited when the system does not work as planned or new insights emerge, up to a decision to stop a project for financial reasons.
  • Document design and development (A.6.2.3). The design decisions are recorded: learning approach, algorithm and model type, training route and required data quality, evaluation and refinement, hardware and software, security threats such as data poisoning, model stealing or model inversion, interface and presentation of outputs, ways for people to intervene, and interoperability and portability. The end result is documentation of the final system architecture.
  • Verification and validation (A.6.2.4). The organization defines test measures and the criteria for using them: test methods and tools, selection and representativeness of test data, release criteria. This includes an evaluation plan for risks that affect people, for instance with acceptable error rates, operational factors and methods for checking whether affected people can interpret the outputs correctly. If the system misses the criteria, the organization reconsiders the intended use, the performance requirements or the countermeasures.
  • Deployment (A.6.2.5). A documented deployment plan accounts for differences between development and target environments, the separate rollout of software and model, and the perspective of the interested parties concerned. Before release, the release criteria are met, such as passed tests, achieved metrics, completed user testing and management approvals.
  • Operation and monitoring (A.6.2.6). At minimum, the organization arranges for repairs, updates, user support and the monitoring of both system and performance. Annex B.6.2.6 adds drift caused by changing production data, continuously learning systems, checking unforeseen uses, AI-specific attacks and task-specific performance metrics with justified target values.
  • Technical documentation (A.6.2.7). The organization determines which technical documentation each group of interested parties needs, such as users, partners or supervisory authorities, and provides it in a suitable form. Typical content includes description and purpose, usage instructions, technical assumptions and limitations, monitoring and control functions, design, data and test records, changes during operation, the impact assessment, and operating documents such as a failure plan, health monitoring, standard operating procedures and roles. The documents are current, accurate and approved by the relevant management.
  • Event logs (A.6.2.8). The organization decides in which stages logging takes place, at least while the system is in use. Logs serve to trace the system’s functioning and to detect performance outside the intended operating conditions. They are kept for as long as the intended use requires, within retention rules and legal requirements.

In practice

Hang the life cycle on the AI register. Each system carries a status in the register (conception, development, validation, operation, decommissioning) together with the evidence for that stage. A change of status requires the release criteria to be met. This shows which systems are running without a deployment plan or validation.

Include bought-in systems. A purchased service also has requirements, a deployment plan and operating rules. Check vendor updates before adopting them for changes in behaviour, record which version is running, and request the technical documentation you need for your users.

Monitor drift and attacks in operation. A model that performed well at release can deteriorate as input data changes. Evaluate the metrics continuously, review samples from a domain perspective and decide at what deviation the model is retrained, restricted or switched off. Extend the threat analysis to AI-specific attacks such as manipulated training data or crafted inputs.

Design logs for a purpose. Log the time of use, the inputs processed or references to them, outputs outside the expected range and interventions by human oversight. Agree content and retention with the privacy team, because logs often contain personal data.

Typical audit evidence

For A.6.2, auditors typically expect:

KPI

Share of AI systems in production with defined performance metrics that are evaluated at the planned frequency

The target is 100%. The figure shows whether operation is actually monitored or whether performance losses only surface through complaints.

Supporting KPIs:

  • Share of systems with a deployment plan and documented approval
  • Number of deviations above the alert threshold and mean time to respond
  • Share of systems with logging active while in use

For high-risk systems, A.6.2 mirrors several provider duties under the AI Act (Regulation (EU) 2024/1689): technical documentation under Article 11 and Annex IV, automatic logging under Article 12, instructions for use under Article 13, and accuracy, robustness and cybersecurity under Article 15. Article 15(5) explicitly names measures against data poisoning, model poisoning, adversarial examples and confidentiality attacks. Providers keep automatically generated logs for at least six months under Article 19, and deployers do the same under Article 26(6) to the extent the logs are under their control.

Deployers monitor operation on the basis of the instructions for use under Article 26(5). For remote biometric identification systems under point 1(a) of Annex III, Article 12(3) prescribes additional log content, an example of the special duties Annex B.6.2.8 points to. Substantial modifications can trigger a new conformity assessment under Article 43(4) (→ EU AI Act).

A.6.2 is the operational core that many controls connect to:

Sources

Frequently asked questions

Does A.6.2 also apply to bought-in AI systems?

Yes, for the most part. Requirements, the deployment plan, operation and monitoring, documentation for your own users and event logs all concern deployers too. Design and validation sit with the provider; you request the evidence and check it in your own context of use.

How long do event logs have to be kept?

The standard asks for retention as long as the intended use requires, within your own retention rules and legal requirements. For high-risk systems, the AI Act sets a minimum of six months, for providers in Article 19 and for deployers in Article 26(6). Data protection deletion duties continue to apply alongside.

What is the difference between verification and validation?

Verification checks whether the system meets its specified requirements, in other words whether it was built right. Validation checks whether it fulfils its purpose in the intended use, in other words whether the right thing was built. A.6.2.4 requires defined measures and criteria for both.

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.