Lifecycle, data, and relationship controls: A.6 to A.10

Lesson 3 of 4 in Annex A Controls and Annexes B, C, D: The Control Framework.

Now the controls that touch the systems themselves. A.6 (AI system life cycle) is the largest grouping — nine controls — because it shadows the entire 22989 lifecycle. Two controls set direction: A.6.1.2 objectives for responsible development and A.6.1.3 processes for responsible design and development. Seven then follow the system through time: requirements and specification (A.6.2.2), documentation of design choices (A.6.2.3), verification and validation (A.6.2.4), deployment (A.6.2.5), operation and monitoring (A.6.2.6), technical documentation (A.6.2.7), and recording of event logs (A.6.2.8).

Those last two deserve a star in your margin. When the EU AI Act demands technical documentation (Article 11) and automatic logging (Article 12) from high-risk providers, A.6.2.7 and A.6.2.8 are the AIMS controls you will point at. Standards people call this the crosswalk dividend: implement the control once, satisfy several regimes.

A.7 (Data for AI systems) gives data its own five-control discipline: an overall data management process (A.7.2), then acquisition (A.7.3), quality (A.7.4), provenance (A.7.5), and preparation (A.7.6). Each is a named grave where an AI project has been buried. Clearview AI scraped billions of face images without consent — acquisition. Getty Images sued Stability AI over training data whose licensing was never established — provenance. Amazon’s recruiting model learned from a decade of male-skewed hiring files — quality and representativeness. The ISO/IEC 5259 data-quality series exists to put engineering detail under A.7.4.

The final three groupings govern relationships. A.8 (Information for interested parties) obliges you to tell people what they need to know: system documentation and information for users (A.8.2), external reporting (A.8.3) — a way for outsiders to report adverse impacts to you — communication of incidents (A.8.4), and information about interested parties’ obligations (A.8.5). Air Canada learned the cost of the gap in Moffatt v. Air Canada (2024): its chatbot invented a bereavement-fare policy, and the tribunal held the airline to it. Accurate user-facing information about what your AI can and cannot do is not courtesy; it is liability control.

A.9 (Use of AI systems) points the lens at your own usage: processes for responsible use (A.9.2), objectives for responsible use (A.9.3), and intended use (A.9.4) — use systems only per their documented intended purpose. When Samsung engineers pasted proprietary semiconductor code into a public chatbot in 2023, the failure was A.9: no responsible-use process governing tools the company had not built.

A.10 (Third-party and customer relationships) completes the chain: allocate responsibilities across the AI supply chain (A.10.2), run supplier processes (A.10.3) so that services, products, or materials you buy meet your AIMS requirements, and address customer expectations (A.10.4). Almost nobody builds their stack alone — your foundation model, your labelling vendor, and your cloud all sit inside your risk surface, and ‘the model came from upstream’ has never impressed a regulator.

Interactive sorting exercise: An obligation lands on your desk. Which Annex A control grouping does the treating control live in? Drag each obligation to its grouping.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.