Regulating AI: powers, people and evidence

An SME uses software to screen job applications. An employee challenges a rejection they do not understand. A municipality is preparing a decision-support tool. They all face the same question: whom should they contact? The answer depends less on the name of an authority than on what they need: an explanation, supervisory scrutiny or a decision they can challenge.

On 6 October 2026, before the House of Representatives’ Economy Committee, Minister Vanessa Matz stated: “Belgium has not yet made this formal designation.” She described two parallel initiatives, a federal law and a cooperation agreement with the federated entities, with a central role envisaged for the BIPT. That statement describes the position on that date. Readers should check the legislation in force when they read this article.[1]

On 15 October 2026, the BIPT opened a consultation on its draft strategic plan for 2027–2029, running until 6 November 2026 inclusive. The draft describes responsibilities relating to data markets and the supervision of AI systems as “in the process of being assigned”. A note states that the Institute will “very probably assume all or part of the responsibility”, but that “the final legislative texts on this subject have not yet been adopted”.[2]

On 20 October 2026, before the House of Representatives’ Social Affairs Committee, Minister David Clarinval announced that the bill intended to establish the AI Act governance framework would be introduced together with the minister responsible for digital affairs in the first half of 2027. This is the timetable announced in a provisional parliamentary report, not evidence that a law has been enacted or published. It supplements the account given on 6 October; it does not remove the need to check legislation adopted since then.[10]

Preparing is not the same as being empowered

The draft is right to prepare: capabilities, guidance for SMEs and an AI supervision laboratory. But a strategic plan cannot, by itself, confer the power to demand a document, impose a measure or levy a penalty. For the new responsibilities envisaged, those powers must rest on the legislation assigning them, without disregarding powers that already exist. The distinction protects everyone: businesses that need to understand their obligations, people seeking a remedy and officials who need to know the limits of their remit.

The opposite conclusion would be equally wrong. The absence of a designation on a particular date does not place AI use outside the law. Data protection, employment law and consumer law continue to apply, each with its own authorities.

Map the functions, then test the routes

My first proposal, drawn from the contribution I prepared for the consultation, is to publish a map of functions rather than a list of addresses. Who provides information? Who receives a report? Who investigates a complaint? Who decides? Each box would identify its legal basis and be updated when the legislation changes. A referral from one service to another would become understandable rather than feel like a closed door.

My second proposal is to give practical form, through real user journeys, to the BIPT’s ambition to be a ‘compass in the digital world’. The draft describes a guiding role, not a named software product. It also plans to examine the establishment of a digital economy assistance service in 2027–2029. The routes proposed here are our contribution to that discussion.[2] People should be able to describe their problem before knowing its legal classification. The route should then explain what the service contacted can do, which documents are useful and what happens next. An acknowledgment is not a decision; reporting a concern does not always open proceedings. Explaining this early prevents disappointed expectations.

These routes would benefit from testing before publication, with citizens, SMEs and people less comfortable with digital tools. The consultation is the right moment to discuss them.

The proposals in these four articles are the author’s own, drawn from the contribution he prepared for the BIPT consultation. They do not represent the Institute.

Choosing an authority is not enough. It is also necessary to decide who advises, who supervises, who decides and how expertise circulates between them. Four examples from other countries help clarify these choices, provided each is read in its actual legal context and stage of development.

Germany: enacted legislation. The KI-MIG of 22 July 2026 (BGBl. 2026 I No 223) gives the Bundesnetzagentur competence as the general rule, subject to exceptions for certain products and financial activities. An independent chamber of three members reports annually to the Bundestag. Its scope is specific: certain high-risk biometric systems used for law enforcement, border management, justice and democracy, together with the high-risk systems listed in points 6, 7 and 8 of Annex III. Section 2 nevertheless reserves supervision of systems placed on the market, put into service or used by public bodies of the Länder to the authorities competent under their respective laws. This chamber is therefore not a general authority for all AI systems.[3]

The Netherlands: a proposal. On 20 April 2026, the government presented a model based on existing authorities, coordinated by the Autoriteit Persoonsgegevens and the Rijksinspectie Digitale Infrastructuur. This is a proposal, not enacted legislation.[4]

Austria: a support service. Austrian legislation assigns the RTR an AI information and advisory service (“Informations- und Beratungsangebot”) for media, telecommunications and postal services.[5] This is a support function. It does not, by itself, create a general power to impose penalties.

The United Kingdom: sharing expertise. The Digital Regulation Cooperation Forum describes its AI audit group as an “internal, cross-regulatory forum”. Regulators exchange audit methods and compare the powers the law gives each of them.[6] The group creates neither a joint authority nor a new power.

Three functions that must remain distinct

These examples reveal three separate functions. Advice helps a business understand what is expected of it. Sharing audit capabilities helps supervisors work more effectively. Supervision and penalties require a specific legal power. An institution may perform all three, but the person dealing with it needs to know which function is being exercised.

The contribution I prepared for the consultation draws a practical conclusion: a business asking a question in good faith should not fear that the question will trigger an investigation into it. How the advisory and supervisory functions communicate, or do not communicate, determines the quality of the questions the Institute will receive.

Cooperate without blurring the roles

A letter dated 8 July 2026 illustrates the boundary in the Netherlands. The President of the Administrative Jurisdiction Division of the Raad van State comments on two provisions of the draft legislation. On Article 4.1, she considers accompanying a project in a regulatory sandbox incompatible with the judicial function. On Article 5.1, she expressly asks that the provision should not apply to the President of that Division: agreeing on a uniform interpretation with other authorities would be incompatible with the President’s independent judicial role. She nevertheless preserves the possibility of cooperation necessary for effective supervision, provided that it does not compromise that independence.[7] The same letter mentions a proposed joint expert team; it does not say that the team is already operating. This is a consultation response, not a judgment.

For Belgium, the lesson concerns method. Before celebrating cooperation, we need to identify the roles it brings together. Our proposal is to test the arrangements under consideration against two or three hypothetical cases: one combining a technical difficulty with an effect on a person, another exposing disagreement between services. The test would show the route followed, responsibility at each stage and what the user learns.

The proposals in these four articles are the author’s own, drawn from the contribution he prepared for the BIPT consultation. They do not represent the Institute.

A strategic plan comes to life in a case opened on a Monday morning, a difficult question or an explanation that needs to be given again. The 2027–2029 draft recognises this reality. It places continuing training, languages, cooperation between services and recruitment at the heart of the Institute’s work, particularly in the new digital fields.[2]

Bring different professions together around one case

Consider a hypothetical example: a business asks whether its new tool complies with the rules. The engineer wants to understand how it works. The lawyer examines obligations and powers. The economist explains the effects on the market. The person at the front desk needs to explain what can be answered immediately. These perspectives complement each other when they meet early.

This contribution therefore proposes workshops using fictional cases built from anonymised real files, temporary exchanges between services and training time included in workloads. Without that allocated time, training always gives way to the urgency of the next case. It also proposes that colleagues participating in a European group report what they have learned in a few dated lines.

What the Sewan case can teach us

Case law can support this shared work, provided it is read in full. Sewan’s dispute with the BIPT led to two judgments by the Market Court.

On 8 March 2023, the Court upheld the finding of infringements. It nevertheless held that the fine, which was criminal in nature, was “not properly reasoned”: the decision did not explain why certain mitigating circumstances had been rejected or why the announced amount remained unchanged despite the corrective measures taken by the operator. It annulled the decision on that point alone, remitted the determination of the amount to the BIPT and ordered the Institute to pay the costs.[8]

On 20 March 2024, in a second judgment, the Court confirmed the principle and proportionality of the fine but annulled the increase for an aggravating circumstance. Its reasoning distinguishes two things. First, knowledge of the rules, invoked in the decision: “There is nothing abnormal about Sewan knowing the rules; quite the contrary.” Second, awareness of breaching them, which the BIPT invoked only in its written submissions: that would imply a deliberate infringement, which the Court found unproven. Negligence, even inexcusable negligence, is not intent. As each party had partly failed in its claims, neither received a procedural indemnity.[9]

The lesson is not that an argument can never be developed before a court. It is more concrete: an aggravating circumstance must be proved, and the reasons for a fine must be explained step by step. A workshop could follow that chain through these two public judgments. Which fact has been established? Which rule applies? Why does that fact justify that amount?

These judgments say nothing about the general quality of a team. They address specific grounds of challenge in a specific case. What is upheld matters as much as what is corrected.

Measure what helps people work better

The draft provides for indicators and periodic evaluation while recognising that outcomes also depend on external factors.[2] For supervisory work, the contribution suggests indicators of how proceedings are conducted rather than their volume: time taken to handle a report, time between completion of the investigation and the decision, and follow-up to requested corrections. A figure never speaks for itself. Fewer reports may reflect better information or a route that has become harder to find.

The proposals in these four articles are the author’s own, drawn from the contribution he prepared for the BIPT consultation. They do not represent the Institute.

A search engine displays a penalty, a date and a title. The reader thinks they have the answer. But the decision may have been challenged, partly annulled and then reconsidered. The Sewan case illustrates the problem: without the second judgment, the first can mislead; without the first, the second is difficult to understand.

What EU TELECOMS aims to do

EU TELECOMS is the independent documentary research project I am developing. Its ambition is European and multilingual: to connect telecommunications regulators’ decisions with the proceedings in which they were reviewed, then with the relevant European texts, including dates and official documents. The project is under development. Its decision and appeal search is not yet open to the public. We will announce it only once it has been checked.

The tool speaks for no authority and does not replace official publications.

Four requirements for reliable research

Connect the right decision to the right appeal. Sewan involves two judgments a year apart. A record that merged them would be wrong. If a subsequent development has not been found, say so without concluding that none exists.

Separate what is upheld from what is corrected. “BIPT overruled” and “penalty upheld” are both inaccurate summaries of the same case.

Respect what the source does not disclose. The 2024 judgment states that certain amounts are confidential. A search engine must not reconstruct them.

Distinguish the interface language from the text’s language. A translated interface helps people find material; the decision is authoritative in its authentic language. A decisive passage should be reread in the original.

Support dialogue, not a league table

The draft plan recalls that BIPT decisions may be challenged before the Market Court, which may suspend, annul or replace them (page 22, section 2.1).[2] This does not mean that every appeal automatically suspends the decision. The court’s powers and the conditions governing them must be checked under the rules applicable to the decision being challenged. Appeals can help the Institute learn and practitioners understand. They are not a ranking system.

You can follow progress on the EU TELECOMS project page. Both Sewan judgments are published on the BIPT website.[8][9]

Trust begins when readers can check what they are being asked to believe. For a legal tool, showing the route to the evidence is worth more than a confident answer.

The proposals in these four articles are the author’s own, drawn from the contribution he prepared for the BIPT consultation. They do not represent the Institute.