Governance that often arrives after the uses
AI governance is rarely the first workstream opened. It arrives when the uses are already there: AI functions activated in off-the-shelf software, assistants built by business teams, employees entering the organisation’s data into consumer tools. No one has the complete list, and no one can say which of these systems create an obligation.
Two classic responses fail. The first is the charter: a text of principles, signed, that nothing links to project decisions. The second is the single committee through which everything must pass: it becomes saturated, lead times lengthen, and teams learn to bypass it. In both cases, undeclared use grows.
Useful governance does something else. It knows which systems exist, classifies each according to its risk, appoints an owner and applies a proportionate level of control. A low-risk use must be able to be authorised quickly. A system that affects a person’s access to a job or to credit must be examined thoroughly.
What the AI Act and the GDPR require
Regulation (EU) 2024/1689 of 13 June 2024, known as the AI Act, entered into force on 1 August 2024. It was amended by Regulation (EU) 2026/1744 of 8 July 2026, which notably revised its timetable. Its application is staggered according to the categories of obligations. The text reasons in terms of levels of risk.
| Level | What the regulation covers | Consequence for your organisation |
|---|---|---|
| Prohibited practices (Article 5) | Prohibited uses | Check that no system, including purchased ones, falls under them |
| High-risk systems (Article 6, Annexes I and III) | Requirements on the system and obligations for providers and deployers | Identify these systems, determine your role, organise human oversight and documentation |
| Transparency obligations (Article 50) | Informing people about certain systems, for example those that interact directly with them | Identify the interfaces and content concerned |
| General-purpose AI models (Chapter V) | Obligations on the providers of these models | Know which models your systems rely on and what documentation their provider makes available |
Two notions structure the classification work. The first is role. The regulation distinguishes in particular the provider, who develops a system or has it developed and places it on the market or puts it into service under its own name, and the deployer, who uses it under its own authority in a professional context. The same organisation may hold both roles depending on the system. Article 25 also provides for cases where a deployer is considered the provider of a high-risk system, for example when it makes a substantial modification to it. The second notion is purpose: it is the intended use that determines the classification, not the technology employed. Annex III covers areas such as employment and the management of workers or access to certain essential services.
For high-risk systems, Article 26 sets the obligations of deployers, in particular: using the system in accordance with its instructions for use, assigning human oversight to competent persons, monitoring its operation and keeping the logs under their control. Article 4, as worded following Regulation (EU) 2026/1744, also asks providers and deployers to take measures to support the development of AI literacy among their staff.
The GDPR (Regulation (EU) 2016/679) applies as soon as a system processes personal data, in training as in use. Legal basis, data minimisation, informing individuals, an impact assessment where processing is likely to result in a high risk (Article 35), safeguards around automated individual decisions (Article 22): these requirements apply cumulatively with those of the AI Act. A system that falls under none of the rows in the table is therefore not outside the framework.
Our approach: four workstreams for applicable AI governance
Our consultants work within your teams; decisions remain yours.
The register of systems. We take stock of systems developed in-house, purchased solutions and AI functions built into software already in place. The sources are cross-checked: project portfolio, procurement, application inventory, interviews with the business. Each record describes the purpose, the owner, the supplier, the data processed and the persons concerned. The register of AI systems is linked to the record of processing activities required by Article 30 of the GDPR.
Classification. A question grid allows a project manager to place their system: purpose, persons concerned, degree of automation of the decision, role of the organisation. Each classification is justified in writing. Doubtful cases are examined with your lawyers and your data protection officer. The final legal qualification rests with your organisation and its legal advisers.
Roles and the approval process. Each system has a named business owner. The process has several tracks: rapid authorisation for low-risk uses, an in-depth review for sensitive systems. Procurement is involved, because part of the obligations is handled in contracts: documentation provided by the supplier, conditions for the use of data, incident notification.
Risk management over time. Assessment before going into service, monitoring in operation, incident handling, re-examination when the purpose, the data or the model changes. Frameworks such as the ISO/IEC 42001 standard (AI management system) or the AI risk management framework published by NIST can serve as a structure, without being an end in themselves.
This arrangement builds on what already exists. The rules on data classification, quality and access fall under data governance: AI governance built alongside it creates duplication and contradictions.
The mistakes we help you avoid
- Limiting the inventory to IT department projects. AI functions activated by a vendor in software already deployed and tools adopted directly by the business often represent the least visible part of the estate.
- Classifying once and for all. A system can change level of risk when its purpose evolves. An assistant used to summarise meeting minutes does not call for the same review as the same tool used to sort job applications.
- Reducing compliance to a file. Documentation that is accurate on the day it is written and never updated offers no protection. Controls must be built into the everyday running of projects.
- Prohibiting everything by default. A general ban shifts use towards personal tools, out of all visibility. Offering an authorised solution is a risk-control measure.
How to know whether governance is working
The indicators concern coverage, speed and durability over time:
- the share of systems in service that appear in the register, checked by cross-referencing with procurement and the application inventory;
- the share of systems classified with a written justification and a named owner;
- the decision time of the approval process, for each track;
- the share of high-risk systems with assigned human oversight and complete supplier documentation;
- the number of incidents reported and the time taken to deal with them;
- the number of systems discovered outside the register during periodic reviews.
The last indicator is the most telling. As long as systems appear outside the register, the approval process is perceived as an obstacle. A fall in this number shows that teams are declaring their uses because they see a benefit in doing so: a rapid answer and a clear framework.