An AI system that performs well in the lab can turn out differently in practice. That is why a provider often wants to try a system out with real applications, real customers or real pupils before it goes into use. The AI Act calls this testing in real-world conditions. Outside an AI regulatory sandbox, Article 60 governs who may carry out such a test and under which conditions. Article 61 governs the consent of the people taking part, and Article 76 what the market surveillance authority can do when things go wrong.
Suppose it is spring 2028. A Dutch municipality wants to use a supplier's AI system that assesses applications for special social assistance and gives the caseworker advice on each application. The system has not yet been placed on the market: the supplier only wants to market it after the trial. Before the system actually goes into use, the municipality wants a three-month trial run on live applications. May the municipality do that on its own? What does the supplier have to arrange? And is the consent that applicants give for the test also enough for the GDPR? The example is fictitious.
In short
Testing in real-world conditions is the temporary testing of an AI system for its intended purpose outside a laboratory or simulated environment. Such a test does not count as placing on the market or putting into service, but only where all the conditions of Article 57 or Article 60 are fulfilled (Article 3, point 57).
Outside an AI regulatory sandbox, only providers and prospective providers may test, on their own or in partnership with one or more deployers (Article 60(1), (2) and (4)). A municipality that will only use a supplier's system therefore does not test under Article 60 itself; it takes part alongside the supplier. If the municipality develops the system itself or has it developed, and puts it into service under its own name, it is the prospective provider.
The conditions are in Article 60(4): a plan, approval by the market surveillance authority, registration, establishment in the Union, at most six plus six months, protection of vulnerable subjects, an agreement with the deployer, informed consent under Article 61, oversight by competent people, and outputs that can be reversed.
The Digital Omnibus on AI extended Article 60 to products under Annex I, Section A, and inserted Article 60a for Section B. Consent under Article 61 is not consent within the meaning of the GDPR (recital 141): a legal basis and, where required, a DPIA remain necessary.
For Annex III systems, no market surveillance authority has yet been designated in the Netherlands; as long as that is the case, in our reading a provider cannot submit a plan and therefore cannot meet paragraph 4, points (a) and (b). For products under Annex I, Section A, the sectoral market surveillance authority is the authority (Article 74(3)), and for financial institutions the financial supervisory authority (Article 74(6)); which of the AFM and DNB is not yet settled. The draft Dutch AI Act implementation bill designates the Dutch Data Protection Authority (AP) for the social assistance example; that is not yet law.
This piece is about testing outside a sandbox. What an AI regulatory sandbox is, is covered in AI regulatory sandboxes, how the Dutch sandbox is being set up in the Dutch AI sandbox, and what the European implementing act on sandboxes covers in the consultation on AI regulatory sandboxes. Where the two routes differ is set out in a table further down.
What counts as testing in real-world conditions?
The definition is in Article 3, point 57: the temporary testing of an AI system for its intended purpose in real-world conditions outside a laboratory or otherwise simulated environment, with a view to gathering reliable and robust data and assessing whether the system complies with the Regulation. Three things follow.
It is not a free zone. Research, testing and development before placing on the market fall outside the AI Act, but testing in real-world conditions is expressly not covered by that exclusion (Article 2(8)). Anyone who tests a high-risk system in real-world conditions therefore does so within Article 57, 60 or 60a; otherwise the test counts as putting into service.
It is part of risk management. High-risk systems are tested before they are placed on the market or put into service, and that testing may include testing in real-world conditions under Article 60 (Article 9(7) and (8)). See the risk management system of Article 9.
Without the conditions it is no longer a test. A test only falls outside putting into service "provided that all the conditions laid down in Article 57 or 60 are fulfilled" (Article 3, point 57). According to the draft explanatory memorandum to the Dutch implementation bill (section 4.6.3), without those conditions the obligations for providers apply in principle.
The Regulation does not say where the simulated environment ends. In our reading, a trial on old, closed case files is not testing in real-world conditions. A trial in which the system runs alongside live applications, even where the caseworker does not have to follow its advice, does in our reading fall within the definition, because the system is then working for its intended purpose under real conditions. That is not settled: on 19 November 2026 we found no guidance or implementing act on this boundary.
Who may test: the supplier, the municipality, or both?
Article 60(1) names only one party that may test: providers or prospective providers of high-risk systems listed in Annex III or covered by the harmonisation legislation in Annex I, Section A. They may do so "on their own or in partnership with one or more deployers or prospective deployers", at any time before placing on the market or putting into service (paragraph 2). And paragraph 4 opens with: providers or prospective providers "may conduct the testing in real world conditions only where" all the conditions are met. The obligations therefore rest on the provider.
The deployer has a smaller role of its own. It is informed of all aspects of the testing relevant to its decision to participate, receives the instructions for use under Article 13, and concludes an agreement with the provider on roles and responsibilities (paragraph 4, point (h)). Through qualified people, it takes part in overseeing the test (paragraph 4, point (j)). And the market surveillance authority can also require the deployer to modify an aspect of the test (Article 76(3), point (b)).
| Situation | Role of the municipality | Who tests under Article 60 |
|---|---|---|
| The supplier develops the system, which has not yet been placed on the market or put into service, and the municipality takes part in the test | Prospective deployer | The supplier as prospective provider; the municipality through the agreement under paragraph 4, point (h) |
| The municipality develops the system itself or has it developed, and puts it into service under its own name | Prospective provider: a public authority that develops a system or has it developed and puts it into service under its own name is a provider (Article 3, point 3), and putting into service includes own use (point 11) | The municipality itself, with all the conditions of paragraph 4 |
| The municipality wants to test a supplier's system on its own, without the supplier | Deployer | Article 60 gives the deployer no route of its own. In our reading, the municipality cannot test this way under Article 60 unless it becomes a provider itself |
| The system has already been placed on the market or put into service, and the municipality starts a pilot | Deployer | No one: Article 60 only works before placing on the market or putting into service (paragraph 2). The pilot is ordinary use, with the obligations of Article 26 and, where applicable, Article 27 |
On the last row: a deployer becomes the provider of a high-risk system if it puts its name or trademark on it, makes a substantial modification or changes its intended purpose (Article 25(1)). When that happens is covered in when do you become a provider. The Regulation does not define prospective provider; in our reading, paragraph 2 makes clear that it means whoever will later place the system on the market or put it into service.
The conditions of Article 60(4)
All eleven conditions must be met at the same time.
| Point | Condition | What it means |
|---|---|---|
| (a) | Plan | The provider draws up a real-world testing plan and submits it to the market surveillance authority of the Member State where the testing takes place |
| (b) | Approval | That authority approves the testing and the plan. If it does not reply within 30 days, approval is deemed given; where national law does not provide for tacit approval, the testing remains subject to authorisation |
| (c) | Registration | Registration under Article 71(4) with a Union-wide unique single identification number and the information in Annex IX; law enforcement, migration, asylum and border control in the secure non-public section of the EU database (Article 49(4), point (d)); critical infrastructure (Annex III point 2) at national level (Article 49(5)) |
| (d) | Establishment | The provider is established in the Union or has appointed a legal representative established in the Union |
| (e) | Transfers | Testing data go to third countries only with appropriate safeguards under Union law |
| (f) | Duration | No longer than necessary and at most six months, extendable once by six months if the provider notifies the authority in advance with an explanation |
| (g) | Vulnerable groups | Subjects who are vulnerable due to their age or disability are appropriately protected |
| (h) | Deployers | They are informed, receive the instructions for use and conclude an agreement with the provider |
| (i) | Consent | Subjects give informed consent under Article 61; only in law enforcement may the test proceed without it, where seeking consent would prevent the testing, the test has no negative effect on them and their data are deleted afterwards |
| (j) | Oversight | Provider and deployer effectively oversee the test through people with the necessary qualifications, capacity, training and authority |
| (k) | Reversal | Predictions, recommendations and decisions of the system can be effectively reversed and disregarded |
The plan. A real-world testing plan describes the objectives, methodology, geographical, population and temporal scope, monitoring, organisation and conduct of the test (Article 3, point 53). The Commission is to specify the detailed elements in an implementing act (Article 60(1), second subparagraph). We did not find one on 19 November 2026, and the draft Dutch explanatory memorandum says it did not yet exist when the memorandum was drafted (section 4.6.4). Until it does, how detailed a plan must be is not settled.
The registration. What is registered is set out in Annex IX: the identification number, the name and contact details of the provider and the deployers involved, a brief description of the system and its purpose, a summary of the plan, and information on suspension or termination. That information is only accessible to market surveillance authorities and the Commission, unless the provider consents to making it public (Article 71(4)). The registration is therefore not public by default, not even at a municipality.
Reversal and disregard. Point (k) is in practice the heaviest condition. In the social assistance example it means, in our reading, that the caseworker decides, that the system's advice can be disregarded, and that a decision that relied on the advice after all can be put right. If that is not possible, the trial is not allowed in that form.
During and after the test: paragraphs 5 to 9
- Withdrawal. A subject or their legally designated representative may withdraw at any time, without detriment and without having to give a reason, by revoking their consent. They may then request the immediate and permanent deletion of their personal data. What has already been done remains valid (paragraph 5).
- Inspections. Member States give their market surveillance authorities the power to require information, to carry out unannounced remote or on-site inspections and to check the conduct of the test (paragraph 6, with Article 75).
- Serious incidents. A serious incident during the test is reported under Article 73: immediately and no later than 15 days, 10 days in the event of death and 2 days for a widespread infringement or a serious and irreversible disruption of critical infrastructure (Article 3, point 49(b)). The provider takes immediate mitigation measures or suspends the test until they are in place, or terminates it, and has a procedure for the prompt recall of the system in that case (paragraph 7). See reporting a serious incident.
- Reporting the end. The provider notifies the market surveillance authority of the suspension or termination and of the final outcomes (paragraph 8).
- Damage. Paragraph 9: the provider or prospective provider "shall be liable under applicable Union and national liability law for any damage caused in the course of their testing in real world conditions". Paragraph 9 says nothing about the deployer; what the municipality and the supplier each bear between them belongs in the agreement under paragraph 4, point (h).
One difference from the sandbox: inside a sandbox, the authorities impose no administrative fines for infringements of the Regulation where the participant observes the plan and the terms of participation and follows the authority's guidance in good faith (Article 57(12)). Article 60 contains no such rule.
Article 61: informed consent
Informed consent is a subject's freely given, specific, unambiguous and voluntary expression of their willingness to take part in a particular test, after having been informed of all aspects relevant to that decision (Article 3, point 59). A subject is a natural person who takes part in the test (point 58).
Consent is obtained before participation, after the subject has received concise, clear, relevant and understandable information on (Article 61(1)):
- the nature and objectives of the test and the possible inconvenience;
- the conditions of the test, including the expected duration of their participation;
- their rights and guarantees, in particular the right to refuse and the right to withdraw at any time, without detriment and without having to give a reason;
- the arrangements for requesting the reversal or disregard of a prediction, recommendation or decision of the system;
- the Union-wide unique single identification number of the test and the contact details of the provider or its legal representative.
Consent is dated and documented, and a copy is given to the subject or their legal representative (paragraph 2).
At the municipality, the applicants for special social assistance are the subjects. Anyone who refuses may suffer no detriment (Article 61(1), point (c), and Article 60(5)). In our reading, that means the ordinary procedure, within the same time limit, for everyone who does not take part. Whether the caseworkers who work with the system are subjects too (Article 3, point 58), the Regulation does not say. They exercise oversight on behalf of the deployer (paragraph 4, point (j)). If their work with the system is also being tested, in our reading it makes sense to seek their consent under Article 61 as well, paying attention to voluntariness in an employment relationship. That is not settled.
Consent under Article 61 is not GDPR consent
Recital 141 is clear on this: the consent of subjects to take part in a test "is distinct from, and without prejudice to" consent for the processing of personal data under data protection law. The AI Act leaves the GDPR unaffected, except where Article 4a and Article 59 provide otherwise (Article 2(7)). For a test outside the sandbox, that means six things.
1. A legal basis of its own under the GDPR. Processing for the test needs a ground under Article 6(1) GDPR. For a municipality, a task carried out in the public interest is the obvious candidate (Article 6(1)(e) GDPR), and that basis must be laid down in Union or national law (Article 6(3) GDPR). Consent within the meaning of the GDPR is difficult for a public body: it is not a valid ground where there is a clear imbalance between data subject and controller, "in particular where the controller is a public authority" (recital 43 GDPR). Whether the municipality's task also covers testing a new system, and whether reusing existing case files is compatible with the original purpose (Article 6(4) GDPR), the municipality must substantiate itself.
2. Article 59 only helps inside the sandbox. Article 59 provides a basis for further processing of personal data collected for other purposes, to develop certain AI systems in the public interest, under ten cumulative conditions. That basis applies "in the AI regulatory sandbox" (Article 59(1)). Outside it, so for a test under Article 60, the AI Act provides no such basis. And even inside the sandbox, Article 59 does not cover processing that leads to measures or decisions affecting the data subjects (Article 59(1), point (f)).
3. Special categories of personal data. Special categories are governed by Article 9 GDPR. Article 4a AI Act exceptionally allows providers, and deployers of high-risk systems, to process such data to detect and correct bias, where that is strictly necessary and six conditions are met. See special category data for bias testing.
4. A DPIA before the test. A DPIA is carried out prior to the processing where it is likely to result in a high risk (Article 35(1) GDPR), and in any event for a systematic and extensive evaluation of personal aspects on which decisions are based that produce legal effects or similarly significant effects (Article 35(3)(a) GDPR). The AP's list of processing operations that require a DPIA (Staatscourant 2019, 64418) includes profiling, financial situation and fraud prevention, each with its own conditions. In the social assistance example, a DPIA is in our reading needed before the test starts. The deployer uses the Article 13 information for it (Article 26(9)), and it already receives that information in the test (Article 60(4), point (h)). The municipality seeks the advice of its data protection officer (Article 35(2) and Article 37(1)(a) GDPR). See when a DPIA is required.
5. Who is the controller? The municipality is the controller for handling the applications. The test itself serves the supplier's purpose: assessing whether its system complies (Article 3, point 57). For the processing for the test, the supplier is therefore in our reading a controller itself (Article 4, point 7, GDPR), or a joint controller together with the municipality (Article 26 GDPR), with a basis of its own; the municipality needs a basis to disclose data to it. Record this in the agreement under Article 60(4), point (h).
6. Rights, decisions and transfers. Withdrawing consent to the test gives the right to request deletion (Article 60(5)), and the rights of the data subject under the GDPR remain (recital 141). Where a decision is based solely on automated processing, Article 22 GDPR applies; see Article 22 GDPR and rubber-stamping algorithms at municipalities. Transfers outside the Union follow, in addition to Article 60(4), point (e), the rules of Chapter V of the GDPR (recital 141).
And the FRIA? A public body carries out a fundamental rights impact assessment "prior to deploying" a high-risk system (Article 27(1)). The Dutch language version says "in gebruik wordt gesteld", the defined term for putting into service, and a test that complies with Article 60 does not count as putting into service (Article 3, point 57). Read that way, the FRIA does not appear to be a condition for the test itself. The English word "deploying", however, is not tied to that defined term. Other language versions diverge too: the German follows the Dutch, the Italian speaks of "before using", and paragraph 2 links the duty to the first use. Whether a test on live applications already triggers the FRIA duty is therefore not settled. In our reading, a municipality would be wise to draw up the FRIA before the test and update it with the test results. For actual use the duty applies in any case: for Annex III from 2 December 2027, and for a system already in use before that date only upon a significant change and by 2 August 2030 at the latest (Article 111(2)). See the FRIA for municipalities.
Article 76: what the market surveillance authority can do
| Paragraph | Power |
|---|---|
| 1 | Market surveillance authorities have the powers to ensure that testing in real-world conditions complies with the Regulation. For a test based on Article 60a, read the authority under the product legislation of Annex I, Section B, for market surveillance authority (second subparagraph, inserted by the Omnibus) |
| 2 | Where testing in real-world conditions takes place within a sandbox, the authority checks Article 60 as part of its supervision of the sandbox. It may then allow derogations from the maximum duration and the protection of vulnerable groups (Article 60(4), points (f) and (g)) |
| 3 | After a serious incident, or where it has other grounds to consider that Articles 60 and 61 are not complied with, the authority can suspend or terminate the test, or require providers and deployers to modify an aspect of it |
| 4 | A decision under paragraph 3 and an objection to the plan under Article 60(4), point (b), state the reasons and how the provider can challenge them |
| 5 | Following a decision under paragraph 3, the authority informs the authorities of the other Member States where the testing takes place under the plan |
Paragraph 2 deserves attention. Within a sandbox, the authority can therefore give room on the two points where Article 60 is strictest: the six-plus-six-month clock and the protection of vulnerable groups. Outside the sandbox, that is not possible.
In the Netherlands, according to the draft explanatory memorandum, approval of a plan is a decision within the meaning of Article 1:3 of the Dutch General Administrative Law Act (Awb), as is a decision to suspend, terminate or modify the test; objection and appeal are available against it (section 4.6.4).
Inside or outside the sandbox?
| In an AI regulatory sandbox (Articles 57 and 58) | Outside a sandbox (Article 60) | |
|---|---|---|
| Plan | The sandbox plan incorporates, where applicable, the real-world testing plan (Article 57(5)) | A separate plan, approved by the market surveillance authority (Article 60(4), points (a) and (b)) |
| Conditions | The authority agrees the terms and safeguards for the test with the participants (Article 58(4)) | The eleven conditions of Article 60(4) |
| Duration and vulnerable groups | The authority may allow derogations (Article 76(2)) | At most six plus six months; protection mandatory |
| Fines | No administrative fines where plan, terms and guidance are followed in good faith (Article 57(12)) | No comparable rule |
| Reuse of personal data | The Article 59 basis for systems in the public interest, including efficiency and quality of public administration (Article 59(1), point (a)(v)), but not for processing that leads to decisions affecting the data subjects (point (f)) | No basis in the AI Act; only the GDPR |
| Available | Each Member State has a national sandbox by 2 August 2027 (Article 57(1)) | Chapter VI has applied since 2 August 2026; see below for where to submit a plan in the Netherlands |
For a municipality that wants to develop or test a system in the public interest using data it already holds, the sandbox may therefore be a route worth considering. Note, however: Article 59 does not cover processing that leads to decisions affecting the data subjects (Article 59(1), point (f)); for a trial that runs alongside live applications, that basis therefore does not help. The draft Dutch implementation bill gives the market surveillance authorities joint responsibility for setting it up, and has the AP advise where an applicant wants to use the Article 59 basis (Article 4.1(1) and (5) of the draft).
What the Digital Omnibus on AI changed
Regulation (EU) 2026/1744 entered into force on 27 July 2026 (Article 4 of Regulation (EU) 2026/1744). For testing in real-world conditions, it makes these changes:
- Article 60 also applies to products under Annex I, Section A (Article 1, point 24, of Regulation (EU) 2026/1744; below only the point). Paragraphs 1 and 2 previously named only Annex III; they now also cover high-risk systems under the harmonisation legislation in Annex I, Section A, such as medical devices and toys. According to recital 26 of the Omnibus this is to foster innovation, "subject to sufficient safeguards", and it leaves other Union or national law on testing such products unaffected.
- New Article 60a for Annex I, Section B (point 25), such as vehicles, aviation and, since the Omnibus, machinery. Member States may allow testing outside a sandbox there to assess compliance with Articles 8 to 15. Those that do adopt a framework and notify it to the Commission in advance. That framework contains a plan agreed between the provider and the competent authority, and ensures compliance with Article 60(2) and (3), (4), points (d) to (j), and (5) to (9). For these systems, only Article 6(1) and Articles 102 to 112 of the AI Act otherwise apply, and Articles 57 to 59 only to the extent the requirements have been integrated into the product legislation (Article 2(2)). According to recital 27, those sectoral acts will in due course incorporate the requirements of Articles 8 to 15.
- Machinery moved to Section B (point 41). Since 27 July 2026, the Machinery Regulation (EU) 2023/1230 is no longer listed in Section A but in Section B of Annex I. AI in machinery therefore runs through Article 60a, not Article 60.
- Article 76(1) gained a second subparagraph (point 33), for tests based on Article 60a.
- The sandbox plan incorporates the testing plan (point 22(d)). Article 57(5) now provides that the sandbox plan incorporates, where applicable, the real-world testing plan referred to in Articles 60 and 60a. According to recital 23 of the Omnibus, this is a procedural simplification.
- The sandboxes themselves. The date by which each Member State must have a national sandbox moved from 2 August 2026 to 2 August 2027 (point 22(a)), paragraph 10 was replaced (point 22(f)): the involvement now applies to the competent data protection authorities; that they are involved was already in the original text, and the Article 58 implementing act now also covers their involvement (point 23).
Article 61, Article 60(3) to (9) and Article 76(2) to (5) are original text.
Two points remain open. First, Article 60(4), point (c), refers for registration to Article 71(4) and Annex IX, which concern Annex III systems. How a test with a product under Annex I, Section A, is to be registered is in our reading not settled. Paragraph 4, point (c), nevertheless remains a condition; put the route to the sectoral market surveillance authority in advance (Article 74(3)). Second, the draft Dutch explanatory memorandum was written before the Omnibus and still says that only Annex III systems can be tested outside the sandbox (section 4.6.3). Since 27 July 2026 that is out of date.
Who approves a testing plan in the Netherlands?
That is not yet laid down in law. The draft Dutch AI Act implementation bill (Uitvoeringswet AI-verordening) was open for public consultation until 1 June 2026. According to the Dutch government's planning letter on digital affairs, submission to the House of Representatives is planned for the second quarter of 2027 (Parliamentary Paper 26643, no. 1554). What the draft provides:
- Who. For Chapter VI, which contains Article 60, and Chapter IX, Sections 1 to 4, which contain Article 76, the AP is the market surveillance authority for, among others, Annex III points 1, 3, 4 and 5(a). For creditworthiness and life and health insurance (points 5(b) and (c)) these are the AFM, DNB and the AP, with a division by category of operator by ministerial regulation (Article 2.2(1)(c) and (2) of the draft). See who will supervise the AI Act in the Netherlands.
- Conditions attached. Conditions may be attached to the approval, and acting contrary to them is prohibited (Article 4.2(1) and (2) of the draft). Rules on costs and on the elements of the plan may follow by ministerial regulation (paragraph 3).
- Tacit approval. The Awb rules on deemed approval where no timely decision is taken (section 4.1.3.3 Awb) apply to the application (Article 4.2(4) of the draft). The draft thereby opts for the tacit approval after 30 days in Article 60(4), point (b).
- Fines. The draft gives the market surveillance authorities the power to fine up to the maximum of Article 99(4) for infringements of Article 60(7) and (8), Article 61 and the conditions attached to the approval (Article 3.7(2), points (b) and (d), of the draft). Whether a municipality, as a public body, can be fined is for each Member State to decide (Article 99(8)).
That is a draft and not yet law. For Annex III systems, no market surveillance authority has yet been designated in the Netherlands; as long as that is the case, in our reading a provider cannot submit a plan and therefore cannot meet paragraph 4, points (a) and (b). For products under Annex I, Section A, the sectoral market surveillance authority is the authority (Article 74(3)), and for financial institutions the financial supervisory authority (Article 74(6)); which of the AFM and DNB is not yet settled.
Main example: the municipality and the social assistance advice
The classification. A system intended to be used by or on behalf of a public authority to evaluate whether someone is eligible for essential public assistance benefits falls under Annex III point 5(a). If the system only performs a narrow procedural or preparatory task and does not profile, it may fall outside high risk (Article 6(3)). We assume a high-risk system. More on the GDPR and FRIA side of such systems is in AI welfare fraud risk scoring.
The roles. The supplier is the prospective provider and bears the conditions of Article 60. The municipality is the prospective deployer and the controller under the GDPR.
The steps, in order:
- Agreement (paragraph 4, point (h)). Supplier and municipality record roles and responsibilities: who writes the plan, who oversees, who reports incidents, who keeps the consent records, what each bears in case of damage and who has which role under the GDPR, including the municipality's basis for disclosing data to the supplier.
- DPIA and legal basis. The municipality substantiates its ground under Article 6 GDPR and carries out a DPIA before the start, with the advice of its data protection officer (Article 35(2) and Article 37(1)(a) GDPR).
- Plan and approval (paragraph 4, points (a) and (b)). The supplier submits the plan to the market surveillance authority; in the draft, that is the AP for this system. As long as that authority has not been designated in law, in our reading this step cannot be taken. It starts only after approval, or after 30 days of silence where national law allows that.
- Registration (paragraph 4, point (c)). The supplier registers the test in the EU database. The information is not public unless it consents (Article 71(4)).
- Consent (Article 61). Each applicant who takes part receives the information in Article 61(1) in advance and gives dated, documented consent. Anyone who refuses gets the ordinary procedure.
- A person decides (paragraph 4, points (j) and (k)). The caseworker decides and can disregard the advice; a decision that relied on wrong advice is put right. Those who oversee are trained and authorised to do so.
- Clock and incidents (paragraph 4, point (f), and paragraphs 7 and 8). Three months fits within the six; it can be extended once by six months after a reasoned notification in advance. A serious incident goes to the authority under Article 73, and at the end the supplier reports the outcomes.
- Afterwards. For actual use, the ordinary obligations follow: the municipality carries out a FRIA (Article 27(1)) and registers its use in the EU database (Article 49(3)), and the supplier has completed its conformity assessment. See the obligations of the deployer.
If the municipality builds the system itself. Then the municipality is the prospective provider and bears all conditions itself: plan, approval, registration, consent and notifications. If it also wants to reuse case files collected for ordinary administration, the sandbox with the Article 59 basis is worth considering, although the national sandbox only has to be operational by 2 August 2027 and Article 59 does not cover processing that leads to decisions affecting the data subjects (Article 59(1), point (f)).
Example 2: a bank and an insurer
Suppose it is 2028. This example is fictitious too.
The bank builds its own. A bank develops its own model that assesses the creditworthiness of consumers (Annex III point 5(b); systems used to detect financial fraud are excluded). The bank puts it into service under its own name and is therefore the prospective provider. It wants to run it alongside real applications for four months. Then all conditions of Article 60(4) apply to the bank itself. Each applicant who takes part consents under Article 61, and anyone who refuses gets the ordinary assessment. During the test, the credit decision follows the existing process, so that the model's output can be disregarded (paragraph 4, point (k)). Under the GDPR, the bank chooses a ground from Article 6(1) GDPR; consent under Article 61 does not automatically count for that. A credit score can itself be a decision within the meaning of Article 22 GDPR where it plays a determining role (CJEU, SCHUFA, C-634/21). For a financial institution, the financial supervisory authority is the market surveillance authority (Article 74(6)); which of the AFM and DNB is not yet settled. The draft Dutch implementation bill names the AFM, DNB and the AP, with a division by ministerial regulation. See also AI in credit decisions.
The insurer buys in. An insurer wants to test a supplier's pricing model for health insurance (Annex III point 5(c)). Here the insurer is the prospective deployer, just like the municipality: the supplier bears the conditions, the insurer concludes the agreement under paragraph 4, point (h), and takes part in oversight. Because the system is used by a financial institution, the financial supervisory authority is in our reading the market surveillance authority (Article 74(6)); which of the AFM and DNB is not yet settled. Before the insurer then actually uses the system, it carries out a FRIA (Article 27(1)).
Briefly on education. If a school tests a system that evaluates learning outcomes (Annex III point 3(b)), the subjects are often minors. The Regulation provides for the legal representative (Article 60(5) and Article 61(2)) and requires protection of those who are vulnerable due to age (paragraph 4, point (g)). A test is without prejudice to the prohibitions of Article 5 (paragraph 1), so emotion recognition in education remains prohibited, except for medical or safety reasons (Article 5(1), point (f)).
From when?
- Articles 60 and 61 are in Chapter VI. That chapter is not among the exceptions in the third paragraph of Article 113 and has therefore applied since the general date of 2 August 2026 (second paragraph). Article 60a is in the same chapter.
- Article 76 is in Chapter IX and has also applied since 2 August 2026.
- The Omnibus amendments entered into force on 27 July 2026; for Chapter VI they have applied since 2 August 2026.
- The requirements a test checks against (Chapter III, Sections 1 to 3) apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I (Article 113, third paragraph, point (c), as replaced by Article 1, point 40, of Regulation (EU) 2026/1744).
- The national sandbox must be operational by 2 August 2027 (Article 57(1)).
- The implementing act on the testing plan was not found by us on 19 November 2026; the Regulation sets no date for it.
What to do now
As a municipality, bank or insurer buying a system:
- Find out which trials are in fact tests in real-world conditions. If a system that has not yet been placed on the market runs alongside real applications or customers, that is testing in real-world conditions and Article 60 or Article 57 applies. If the system is already on the market, a pilot is ordinary use, with the obligations of Article 26 and, where applicable, Article 27.
- Do not test alone. As a deployer, you test together with the provider. Ask for the plan, the approval and the identification number, and conclude the agreement under paragraph 4, point (h).
- Handle the GDPR separately. A legal basis, a DPIA before the start, the data protection officer's advice and a clear division of roles with the supplier. Consent under Article 61 does not replace any of that.
- Design for no. Anyone who refuses or withdraws gets the ordinary procedure, and every piece of advice from the system must be capable of being disregarded and put right.
As a provider:
- Choose the route. Inside a sandbox with tailored terms, protection from fines when acting in good faith and possibly the Article 59 basis, or outside a sandbox with the fixed conditions of Article 60.
- Plan the time. Allow 30 days for approval and at most six plus six months for the test.
- Record consent properly. The information in Article 61(1), dated, documented and with a copy for the subject.
- Build in a brake. A procedure for serious incidents, suspension and prompt recall (paragraph 7).
Want to know which AI Act and GDPR obligations come into view for your use case and which role you have? With the case check you describe the use case in a few questions.
Frequently asked questions about testing in real-world conditions
Sources
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