Enterprise AI has moved beyond isolated pilots. Organisations now run multi-step agents, customer-facing assistants, and automated decision flows connected to ERP, CRM, and cloud data. As change speeds up, governance built around periodic spreadsheet reviews and static checklists tends to fall behind the systems it is meant to control.
Why Static Policy Documents Fall Behind
When engineering teams change prompts, swap models, or extend retrieval indices, a point-in-time assessment can be out of date within days. The EU AI Act (Regulation (EU) 2024/1689) expects risk management for high-risk systems to be a continuous process. Article 9 applies to high-risk AI systems, and the obligations that apply to a given system depend on its role and classification.
Continuous Assurance in Practice
We treat AI governance as part of the engineering system. For each production AI system, we keep four records up to date:
- System register: model provider and version, intended purpose, users, EU market exposure, and the risk classification with its reasoning.
- Runtime checks: input and output checks for prompt injection, schema validation, and personal-data handling, sized to a latency budget agreed for each system.
- Transparency: clear disclosure that a person is talking to an AI system (Article 50(1)), and machine-readable marking of generated content where Article 50(2) applies.
- Logs: event logs kept for at least six months where the Act requires it. Article 19 applies to providers of high-risk systems; Article 26(6) applies to deployers.
Relating the EU AI Act to ISO/IEC 42001
Customers and auditors increasingly ask for a single evidence trail. ISO/IEC 42001 describes an AI management system, and it is useful for organising controls. It does not, by itself, establish conformity with the EU AI Act. Mapping the two, alongside ISO/IEC 27001 and SOC 2 criteria, lets one set of technical evidence serve several reviews. The mapping needs checking against the current text of each standard and the Act before it is relied on.
A Practical First 90 Days
A workable start is to list every internal and third-party AI use, classify each one against the prohibitions in Article 5, the Annex III categories, and the Annex I product rules, add logging and transparency measures where they are needed, and set up a small committee with a clear route for incidents. Legal review should follow the classification, not precede it.
Last reviewed 17 November 2026*
