Skip to main content
Praxikon

Who tests AI systems for the regulator? Union AI testing support structures under Article 84 of the AI Act

··12 min read

A market surveillance authority that wants to know whether an AI system does what its provider promises has to be able to test it. Not every Member State has the expertise and equipment to do that itself. Article 84 therefore provides that the Commission designates one or more Union AI testing support structures. They work at the request of authorities; in our reading providers and deployers cannot use them directly. Their work can still reach you: when a test report becomes part of an investigation into a system you supply or use.

We work from the consolidated text of 27 July 2026, including the amendments made by Regulation (EU) 2026/1744.

In short

The Commission designates structures that perform, in the area of AI, the tasks listed in Article 21(6) of Regulation (EU) 2019/1020: carrying out testing of products at the request of market surveillance authorities, the Network or the Commission; providing independent technical or scientific advice at the request of the Network; and developing new techniques and methods of analysis. They also provide independent technical or scientific advice at the request of the Board, the Commission or market surveillance authorities.

Article 84 has applied since 2 August 2026. On 19 November 2026 we found no designation of a Union AI testing support structure. That does not mean none will come.

What Article 84 provides

The Article has two paragraphs. Under paragraph 1 the Commission shall designate one or more Union AI testing support structures to perform the tasks listed under Article 21(6) of Regulation (EU) 2019/1020 in the area of AI. That is the Regulation on market surveillance, which Article 74(1) also applies to AI systems. Article 21(6) lists three activities, within the area of competence of the facility:

  1. carry out testing of products at the request of market surveillance authorities, the Network or the Commission;
  2. provide independent technical or scientific advice at the request of the Network;
  3. develop new techniques and methods of analysis.

The Network is the Union Product Compliance Network set up by that Regulation. Because Article 74(1)(b) provides that references to a product are to be understood as including all AI systems within the scope of the AI Act, the testing is in our reading the testing of AI systems.

Paragraph 2 adds a task. The structures shall also provide independent technical or scientific advice at the request of the Board, the Commission, or of market surveillance authorities. Under Regulation (EU) 2019/1020 only the Network can ask for advice; under the AI Act the Board, the Commission and the national authority can as well.

Recital 152 gives the purpose: to support adequate enforcement as regards AI systems and reinforce the capacities of the Member States. And Article 69(3) asks the Commission to ensure that the combination of support activities carried out by the testing support structures and by the experts of the scientific panel is efficiently organised and provides the best possible added value.

Has a structure been designated yet?

On 19 November 2026 we found no designation. We searched EUR-Lex for "AI testing support structures": that returns five results, all of them the AI Act itself, its consolidated text and preparatory documents. On that date the Publications Office database (Cellar, the source behind EUR-Lex) likewise contained no act with the AI Act as its legal basis that designates a testing support structure. The Commission's pages on the AI Office and on governance and enforcement of the AI Act do not mention a designation either, and our own register of implementing acts lists none.

Under Regulation (EU) 2019/1020 itself, the Commission does designate Union testing facilities, by implementing decision. Examples, both following a call for expression of interest, are Implementing Decision (EU) 2023/733 for radio equipment and toys and Implementing Decision (EU) 2026/1481 for textiles. We found no such decision for AI. Article 84 itself does not say in what form or when the Commission designates.

Not a test lab for providers

Three things look alike but are not the same:

  • Union AI testing support structures (Article 84) serve enforcement: they test and advise at the request of authorities. In our reading a provider cannot order a test from one; Article 84 and Article 21(6) of Regulation (EU) 2019/1020 name only market surveillance authorities, the Network, the Board and the Commission as requesters.
  • Testing and Experimentation Facilities (TEFs) come from the Digital Europe Programme. According to the Commission they are open to technology providers across Europe to test and experiment with AI solutions at scale, and they can support regulatory sandboxes. That is innovation support, not enforcement.
  • AI regulatory sandboxes (Article 57) are a controlled framework in which a provider develops and tests under the guidance of a competent authority. See the AI regulatory sandbox rules.

From when

Article 84 sits in Chapter IX and has applied since 2 August 2026, the general date of Article 113. Article 69 has applied since 2 August 2025 (Article 113, point (b)). The requirements for high-risk systems under Annex III apply from 2 December 2027, those for systems under Annex I from 2 August 2028 (Article 113, point (c)). In our reading a test of compliance with those requirements becomes most likely after those dates.

When a test report ends up in your file

Suppose it is 2028. A municipality uses a system that gives every application for special assistance benefit a score for the likelihood that the application is incomplete; a caseworker decides. The supplier placed the system on the market in 2028. The example is invented. In our reading the system falls under Annex III, point 5(a). Because it makes a prediction about each applicant based on personal data such as income, it performs profiling in our reading (Article 3, point (52), referring to Article 4, point (4), GDPR) and is therefore always high-risk, even if it were a preparatory task (last subparagraph of Article 6(3)). We assume it is not social scoring within the meaning of Article 5(1)(c). The supplier is the provider, the municipality the deployer.

A resident complains to the market surveillance authority (Article 85). The authority has sufficient reason to consider that the system presents a risk and evaluates it (Article 79(2)). The documentation does not settle how often applicants with fluctuating income receive a high score. The authority asks a testing support structure to test the system and receives a report. That report becomes part of the evaluation.

What the municipality must be able to show:

  • How it uses the system. In accordance with the instructions for use (Article 26(1)), with human oversight assigned to caseworkers who have the necessary competence, training and authority (Article 26(2)), and with input data that is relevant and sufficiently representative, to the extent it exercises control over that data (Article 26(4)). Also that it informed applicants that the system is applied to them (Article 26(11)). How caseworkers deal with the score does not, in our reading, follow from a test of the system; the municipality has to show that itself.
  • The logs under its control, kept for at least six months (Article 26(6)), and its registration in the EU database (Article 26(8) and Article 49(3)).
  • The fundamental rights impact assessment (Article 27) and what it did with its monitoring (Article 26(5)).

The municipality cooperates with the authority (Article 26(12) and Article 79(2)). If the authority then takes a measure, it must state the exact grounds, and the economic operator to whom the measure is addressed must first be given the opportunity to be heard within a period of not less than 10 working days, unless urgency prevents it (Article 18(1) and (3) of Regulation (EU) 2019/1020, applicable through Article 79(2)). Where the measure is addressed to the municipality, in our reading that is the moment to respond to the test report, for example where the test used a different configuration from the one the municipality uses; where it is addressed to the supplier, that response lies first with the supplier. The authority and the testing support structure must keep what they learn about the system confidential (Article 78); in our reading that may limit what the municipality gets to see of the report, for example where the supplier's trade secrets are involved. How the powers and document requests work in more detail is set out in market surveillance powers and in the procedure for AI systems presenting a risk.

Second example: a bank. A bank builds its own model that evaluates the creditworthiness of natural persons (Annex III, point 5(b)) and uses it for its own loans. This example is invented too. The bank is then both provider and deployer (Article 3, points (3) and (4)). The market surveillance authority is the authority responsible for the financial supervision of the bank, in so far as the placing on the market, putting into service, or the use is directly connected with the provision of the financial service (Article 74(6)), unless the Member State identifies another relevant authority under Article 74(7). If that authority asks a testing support structure for a test or advice, the bank as provider must be able to give the authority, where necessary, access to its technical documentation (Article 11) and the training, validation and testing data sets (Article 74(12)). As deployer the bank must also be able to show its fundamental rights impact assessment (Article 27(1) expressly covers Annex III, points 5(b) and (c)) and its logs, which it keeps as part of the documentation kept pursuant to financial services law (Article 26(6)). In our reading it helps if its own testing for accuracy and robustness (Article 15) is recorded in a repeatable way, with the model version and the test data. For an insurer with a system for risk assessment and pricing in relation to life and health insurance (Annex III, point 5(c)) it works the same way in our reading.

In the Netherlands. No market surveillance authority for Annex III has yet been designated by law; the Dutch AI Act Implementation Act (Uitvoeringswet AI-verordening) is a draft. What that draft proposes is set out in who enforces what. Whether and how a Dutch authority will call on a testing support structure is not yet settled.

Frequently asked questions about Union AI testing support structures

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

When the regulator knocks: Articles 79 to 82 of the AI Act on AI systems presenting a risk

What happens when a market surveillance authority considers that an AI system presents a risk? Article 79 sets out the national procedure: evaluation, corrective action within a period of at most 15 working days, and where needed a prohibition or recall. Article 80 deals with a system the provider wrongly classified as non-high-risk, Article 81 with the Union safeguard procedure and Article 82 with a system that complies with the Regulation and still presents a risk. With who will supervise in the Netherlands, from when it applies, and a municipality using a system for social assistance applications and a bank as examples.

EU AI ActArticle 79Article 80

Certificates and the emergency derogation: what Articles 44, 45 and 46 of the AI Act regulate

A certificate from a notified body is valid for four or five years at most and can be suspended or withdrawn. And very occasionally a high-risk AI system may be used while the conformity assessment is still under way. What Articles 44, 45 and 46 of the AI Act regulate, who will decide on such an authorisation in the Netherlands, why the GDPR simply continues to apply, and what a safety region, a municipality and a bank need to do.

EU AI ActArticle 44Article 45

Experts of the scientific panel for the supervisor: Article 69 AI Act

Article 69 allows Member States to call upon experts of the European scientific panel on AI. It is not the organisation itself but its supervisor that brings these experts in. What the Article provides, what the Digital Omnibus changed about the fees, how it relates to Articles 68, 78 and 84, and what a municipality or a bank notices during an investigation.

EU AI ActArticle 69Scientific panel