Skip to content
Technology transformation expertise firm · ARDIANE Group Innovate, secure, succeed

Artificial intelligence

In what order should AI governance be implemented?

In brief

Implementing AI governance is the sequence of steps by which an organisation lists its AI systems, classifies them by risk level, designates who is accountable for them, sets up an approval process and organises their monitoring. Management Square describes these steps here in order, what each one produces and who is accountable for it, in the light of Regulation (EU) 2024/1689.

Page contents

In paragraph 2 of its Article 26, Regulation (EU) 2024/1689 on artificial intelligence states: “Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.” This provision, read in the consolidated version following Regulation (EU) 2026/1744, concerns high-risk AI systems. It is enough to show what implementing AI governance requires.

To apply it, an organisation must know which systems it uses, which of them are high-risk, who is accountable for them, who authorised their entry into service and who monitors them. These answers are established in this order: inventory, classification, roles, approval process, monitoring. Each step produces a verifiable document and has a named owner. The content of the set-up is presented on our AI governance page; this article describes how it unfolds.

Implementing AI governance: the order of the steps and its reason

The order comes from a dependency: each step consumes what the previous one produced. Only listed systems can be classified. An owner is designated for a system whose level of requirements is known. An approval process without roles has no one to convene.

One prerequisite comes before the inventory: the mandate. Senior management designates in writing the person who leads the approach, their reporting line and the body to which they report. Without this mandate, the departments asked to contribute to the inventory have no reason to respond.

Step What it produces Who is accountable Proposed exit criterion
Inventory The register of AI systems The person who leads the approach An additional source no longer reveals any unknown system
Classification For each system, the organisation’s role and the risk level, justified in writing The system’s business owner, with the legal or compliance function Each line of the register carries a dated classification
Roles The responsibility matrix and the designations The department to which the use belongs Each system has a named owner, who has accepted their role
Approval process The procedure, the file template and the record of decisions The chair of the body A first real file has been examined from start to finish
Monitoring Review minutes and the incident log The owner of each system, then the person who leads the approach for consolidation A first review has led to a change in the register

This table is a proposed method: the order of the steps, the allocation of responsibilities and the exit criteria are not drawn from the regulation, and are adjusted to the size and bodies of each organisation.

The inventory opens the register of systems

The inventory starts from the lists the organisation already has: software contracts, application repository, project portfolio. A short questionnaire is then sent to the departments, for uses that do not appear in any of these lists.

A common definition avoids debating each case. That of Article 3, point 1, of the regulation describes a machine-based system that “infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions”. Adopting it as it stands removes the need to invent a scope specific to the organisation.

Each line of the register is declared by a named person. The field devoted to the data processed is filled in with the owners designated by data governance. From this step onwards, a declaration channel is opened for new systems, otherwise the register ages as soon as it is closed.

Classification assigns each system a role and a risk level

Two questions arise for each line, in this order. The first concerns the role: is the organisation a provider or a deployer, within the meaning of Article 3, points 3 and 4? Article 25(1) adds cases in which a deployer is considered a provider of a high-risk system, for example when it modifies the intended purpose of a system in such a way that it becomes high-risk. The second concerns the level: prohibited practice (Article 5), high-risk system (Article 6, Annexes I and III), transparency obligation (Article 50), or none of these regimes. General-purpose AI models follow a separate regime, in Chapter V, which weighs on their providers. Our entry on the AI Act presents these categories.

The regulation itself gives the example of a reasoned classification. Under its Article 6(4), “a provider who considers that an AI system referred to in Annex III is not high-risk shall document its assessment before that system is placed on the market or put into service”. The method extends this discipline to every line of the register: a written, dated classification, accompanied by the events that would cause it to be reopened.

The business owner fills in the grid. The legal or compliance function validates it, with the data protection officer when personal data is processed. The legal qualification belongs to the organisation and its advisers. The application dates specific to each category are read in Article 113 of the consolidated version and carried over into the register.

Roles are designated system by system

Each system receives a business owner, who is accountable for its use and purpose, and a technical owner, who is accountable for its operation and the retention of logs.

For a high-risk system, the people responsible for human oversight are added. The designation takes up one by one the terms of Article 26(2): the expected competence, the training received, the authority, that is the power to set aside an output of the system, and the support, that is the time and recourse available to the person.

Other responsibilities are set at the level of the organisation. Article 4 requires providers and deployers to take measures to promote the development of AI literacy among their staff: a function is responsible for this. Article 26(7) requires deployers who are employers to inform workers’ representatives and the workers concerned before putting a high-risk system into service in the workplace: the matrix indicates who does this.

The approval process hooks onto existing milestones

The process plugs into the decisions the organisation already takes: a project’s commitment milestone, the signature of a contract by procurement, the activation of an AI feature in software already in service, a change of purpose. The use cases selected by the AI strategy arrive there with their business case. The process adds the classification and the conditions of use.

The file is proportionate to the classification. For a system outside the regulation’s regimes, the register record and its classification are enough, and the decision is taken by delegation. For a high-risk system, the file brings together the provider’s documentation, the designation of human oversight and the required assessments: the data protection impact assessment under Article 35 of the GDPR where it applies, and the fundamental rights impact assessment that Article 27 of the regulation requires of certain deployers.

Each review concludes with one of four decisions: authorised, authorised with conditions, refused, sent back for classification. The record keeps the reason and the conditions.

Monitoring leads back to the previous steps

The owner of each system rereads its line in the register at a set interval and keeps the incident log. For a high-risk system, Article 26(5) requires the deployer to monitor operation on the basis of the instructions for use. If it detects a serious incident, it immediately informs the provider, then the importer or distributor and the market surveillance authorities. The procedure names the person responsible for this.

Monitoring also serves to restart earlier steps. A system discovered outside the register reopens the inventory. A change of purpose, model, data or provider terms reopens the classification. The departure of an owner reopens the roles. Requests that bypass the body lead to a review of the process. The indicators that substantiate these findings are on our service page.

The register can be opened as soon as the mandate is signed, from two lists that the organisation already has: its current software contracts and its applications in service.

Insights

Further reading

All insights

FAQ

Frequently asked questions

Another question? Ask us

Which step should AI governance implementation start with?

With the inventory of AI systems, once senior management has given the mandate designating the person who leads the approach. Classification, roles and the approval process concern listed systems: started before the inventory, they would apply to an unknown scope. The inventory begins from software contracts, the application repository and the project portfolio.

Does the European AI regulation require an internal register of AI systems to be kept?

The consolidated version of Regulation (EU) 2024/1689 contains no article requiring, under that name, a register internal to the organisation. Its Article 49 does, however, provide for registration in the EU database, in particular for providers of the high-risk systems listed in Annex III and for deployers that are public authorities. The internal register remains a practical necessity: a system that does not appear in it can be neither classified nor assigned to an owner.

Who is accountable for AI governance in an organisation?

Three levels share the responsibility. Senior management gives the mandate and designates the person who leads the approach, each system has a named business owner, and a body decides sensitive cases. For high-risk systems, Article 26(2) of Regulation (EU) 2024/1689 also requires the deployer to assign human oversight to persons who have the necessary competence, training, authority and support.

Should you rely on the ISO/IEC 42001 standard or on the NIST framework?

The NIST AI Risk Management Framework (AI RMF 1.0), published on 26 January 2023, is voluntary in use and is organised into four functions: govern, map, measure and manage. NIST states that it is under revision. This framework, like the ISO/IEC 42001 standard, can serve as a structure for the approach, without removing the need for the inventory or for classification under the European regulation.

Is the AI literacy obligation part of governance?

Article 4 of Regulation (EU) 2024/1689, as worded following Regulation (EU) 2026/1744, requires providers and deployers to take measures to promote the development of AI literacy among their staff. The same article states that this obligation does not require them to guarantee a specific level of literacy for any individual. In implementation, the roles step designates the function responsible for these measures and how a record of them is kept.

A project to bring to completion?

A first conversation is enough to understand your challenge, identify the right people and tell you frankly what we can bring.