Every Member State must have an AI regulatory sandbox by 2 August 2027. A sandbox is an agreed project in which you develop and test an AI system under the guidance of the supervisory authority before it is placed on the market. Article 57 makes sure the sandbox exists. Article 58 instructs the Commission to lay down the rules. It did so on 10 November 2026, with Implementing Regulation (EU) 2026/2487. The Regulation was published on 11 November 2026 and enters into force on the twentieth day after that. By our calculation that is 1 December 2026.
Suppose it is December 2026. The municipality of Waterveen wants to test, together with a software company of 30 employees, a system that helps assess whether a resident is entitled to special assistance benefits. Which of the two applies? What does it cost? And what use is the paper the municipality receives at the end? Then the same questions for a bank, a university of applied sciences and a start-up. The examples are invented.
In short
Article 58 is addressed to the Commission. It adopts implementing acts on four subjects: who is admitted and how the selection works, the procedures from application to exit, the terms and conditions for participants and, since the Digital Omnibus on AI, the governance of the sandboxes. Paragraph 2 lists what those acts must ensure, such as a decision within three months and access free of charge for SMEs.
Implementing Regulation (EU) 2026/2487 fills this in with ten articles. It applies to every sandbox established under Article 57: national, regional or local, that of the European Data Protection Supervisor and that of the AI Office.
The provider or prospective provider applies, if it wishes in partnership with a deployer. The system must not yet be on the market, or it will be substantially modified, or a modified intended purpose turns it into a high-risk AI system. The authority weighs five selection criteria and decides within three months. That decision can be appealed.
SMEs and public authorities pay nothing for guidance, advice and supervision. Other undertakings may be charged the full costs.
At the end the participant receives an exit report, no later than 45 working days after completing participation, and on request a written proof. Those documents are not a declaration of conformity and not an approval to place the system on the market. Market surveillance authorities and notified bodies in all Member States must recognise them as documentation.
Participation exempts no one from the law. The GDPR keeps applying, and the data protection authority is involved as soon as personal data are processed.
This piece is about the rules. What a sandbox is and how the Netherlands plans to set it up is covered in the Dutch AI sandbox. Testing outside a sandbox is covered in testing in real world conditions, and priority for small companies in SMEs, start-ups and small mid-caps.
What Article 58 itself says
In our reading, Article 58 gives an applicant no rights of its own. It is an instruction to the Commission, which adopts the rules together with a committee of the Member States (the examination procedure of Article 98(2)). The implementing acts include common principles on four issues (paragraph 1):
| Point | Subject |
|---|---|
| (a) | Eligibility and selection criteria for participation |
| (b) | Procedures for the application, participation, monitoring, exiting from and termination of the sandbox, including the sandbox plan and the exit report |
| (c) | The terms and conditions applicable to the participants |
| (d) | Governance, including the involvement of and supervision by the data protection authorities and the coordination and cooperation at national and Union level |
Point (d) was added by the Digital Omnibus on AI (Regulation (EU) 2026/1744), which also put the word governance in the introductory part of paragraph 1. The rest of Article 58 was not amended.
Paragraph 2 says what the implementing acts must ensure. In brief:
- the sandbox is open to any provider or prospective provider who fulfils transparent and fair criteria, and the applicant is informed of the decision within three months (point (a));
- access is broad and equal, and applications in partnership with deployers and other third parties are possible (point (b));
- national authorities keep as much flexibility as possible for their own set-up (point (c));
- access is free of charge for SMEs, including start-ups, apart from exceptional costs recovered in a fair and proportionate manner (point (d));
- the learning outcomes help with conformity assessment and with the codes of conduct of Article 95 (point (e));
- other actors can be involved, such as notified bodies, standardisation organisations, testing facilities and researchers (point (f));
- procedures are simple and streamlined across the Union, and participation is mutually and uniformly recognised across the Union (point (g));
- participation lasts as long as is appropriate to the complexity and scale of the project and may be extended (point (h));
- the sandbox helps develop tools for testing and assessing systems, for instance on accuracy, robustness and cybersecurity (point (i)).
Paragraph 3 says that prospective providers, in particular SMEs and start-ups, are directed to help such as guidance on the implementation of the Regulation, help with standardisation documents and certification, and testing facilities. Paragraph 4 concerns testing in real world conditions within the sandbox: the authority specifically agrees the terms and conditions and the safeguards with the participants.
The Implementing Regulation at a glance
| Article of Implementing Regulation (EU) 2026/2487 | Subject |
|---|---|
| 1 | Scope: all sandboxes under Article 57(1), (2), (3) and (3a) of the AI Act |
| 2 | General terms and conditions, priority and costs |
| 3 | Application and selection, decision within three months, appeal |
| 4 | Participation: records kept by the authority, duration, confidentiality |
| 5 | The sandbox plan |
| 6 | Written proof and exit report |
| 7 | Voluntary follow-up activities after participation |
| 8 | Extension and suspension |
| 9 | Annual reports and evaluation |
| 10 | Entry into force |
The Regulation is addressed mainly to the competent authorities. They must apply the templates and guidance material adopted by the Board under Article 66(d) of the AI Act, for instance for the application form, the sandbox plan and the exit report. We have not been able to establish which templates already exist. So ask the authority.
The Implementing Regulation has no separate chapter on governance. The involvement of the data protection authority runs through the sandbox plan, see below.
Who can apply, and with which system?
Any provider or prospective provider may apply, using a form the authority makes available (Article 3(1) of Implementing Regulation (EU) 2026/2487). A proposal is eligible for consideration if it meets two criteria at the same time (Article 3(2)).
The first concerns the applicant. That is the provider or prospective provider. It may submit the application in partnership with deployers or other third parties needed for the participation. In our reading a deployer therefore cannot apply on its own. A provider without a registered office or a branch in the Union applies jointly with an authorised representative in the Union.
The second concerns the system. It must be one of three situations:
- the AI system has not been placed on the market or put into service;
- it has already been placed on the market or put into service, but will be subject to substantial modification within the meaning of Article 3(23) of the AI Act (the Dutch text of the implementing regulation uses a different word here, "ingrijpend", but refers to the same definition);
- according to its intended purpose it was not classified as high-risk, but it has been placed on the market or put into service in such a way that it becomes a high-risk AI system (Article 25(1)(c) of the AI Act).
So the sandbox is not meant for a system that is already running and does not change.
You do not have to apply in the Member State where you are established. A Dutch provider may also turn to a sandbox in another Member State.
How does the authority select?
The authority takes at least five selection criteria into consideration (Article 3(3) of Implementing Regulation (EU) 2026/2487):
- the need to participate, due to specific regulatory and practical challenges in relation to the AI Act and other applicable law;
- the potential for and scalability of regulatory learning, including the expected benefits for the sector;
- the contribution to objectives of public interest or societal benefit and Union policy;
- the maturity of the proposal;
- the innovative character of the AI system. The Dutch language version reads "intended character" here; the English version and recital 14 are about the innovative character.
According to recital 14, innovation can also lie in existing technology applied to a new domain, or in new measures to comply with the law or to better protect fundamental rights. An authority may add criteria of its own if they are reasonable, proportionate and duly justified.
The authority may also periodically decide to give more priority to specific sectors in industrial or public areas of particular strategic importance, and to areas of notable regulatory implementation challenges, while still respecting broad and equal access (Article 3(4)). SMEs, including start-ups, having a registered office or a branch in the Union get priority access, to the extent that they fulfil the eligibility conditions and selection criteria (Article 62(1)(a) of the AI Act). In a sandbox of the AI Office the priority also applies to small mid-cap enterprises (Article 57(3a) of the AI Act and Article 2(3) of the implementing regulation).
The applicant is informed of the outcome no later than three months after submitting the complete application, or three months after the submission deadline if the authority works with a call for applications. The applicant can appeal the decision under the national rules for administrative decisions (Article 3(7)). According to recital 15 this includes the right to be heard, access to one's own file and a reasoned decision. How this will run in the Netherlands depends on the authority that will operate the sandbox and on national rules.
What does it cost?
| Who | What the authority may charge |
|---|---|
| SME, including a start-up | Nothing for regulatory guidance, advice and supervision concerning the AI Act and other law supervised in the sandbox. That includes supervision of testing in real world conditions, support for conformity assessment, reliance on the regulatory privileges of Article 57(12) and Article 59, redirection to pre-deployment services and supervision of the testing and validating of AI systems. Exceptional costs for other, agreed services may be recovered |
| Public authority, body or agency | The same pricing regime as for SMEs |
| Other undertaking | The authority may charge the full costs of the services and then publishes the list of fees when opening the application phase |
| Provider without a registered office or branch in the Union | A participation fee, and it designates an authorised representative in the Union |
This is laid down in Article 2(3), (4) and (5) of Implementing Regulation (EU) 2026/2487. Exceptional costs are costs for services other than the free ones, for which the authority requires additional financial capacity. Those services are agreed with the participant and set out in the sandbox plan. The authority recovers them in a fair, transparent and proportionate manner, announces the possibility as early as possible and in any case before participation starts, and sends an invoice that shows which costs occurred and how they were calculated.
That public authorities also fall under the pricing regime for SMEs is not in Article 58 of the AI Act. The Implementing Regulation adds it (Article 2(3), last subparagraph, and recital 9). It does not define what a public authority, body or agency is.
The sandbox plan
After selection the participant and the authority agree a sandbox plan without undue delay. It specifies at least (Article 5(2) of Implementing Regulation (EU) 2026/2487):
- who the applicant is and which AI system is concerned;
- the estimated timeline, including an estimated end date;
- the objectives, the activities, the roles of the provider and of any deployers and other third parties, and the services the authority provides, including those for which costs are charged;
- the requirements and obligations of the AI Act and other law that are supervised in the project;
- whether personal data are processed and whether the data protection authority needs to be involved;
- risk management safeguards and a procedure for monitoring, managing and reporting serious incidents, where necessary and appropriate given the risks;
- where applicable: a plan on testing in real world conditions under supervision.
The participant identifies, to the best of its ability, which information in the plan is confidential. That matters later: what you do not mark as confidential in the plan, or explicitly before the end of the project, is presumed to be available for disclosure (Article 6(8)). In our reading this matters mainly for the key findings and lessons learned from the exit report that the authority may publish even without your agreement; those may contain non-confidential information only. The plan may be updated during the project if both sides agree.
How long, extending and stopping
The duration must be proportionate to the capacity of the authority, to whether the project covers part or all of the conformity assessment of Article 43, and to the complexity and scale of the project (Article 4(4) of Implementing Regulation (EU) 2026/2487). The Regulation names no fixed maximum duration.
An extension is possible until 30 calendar days before the planned end date, if both sides agree and the extension is justified (Article 8(1)).
The authority may temporarily or permanently suspend the testing process or the participation if no mitigation is possible. It may consider that to be the case where the conditions of the sandbox plan have been breached, where there are significant risks to health, safety or fundamental rights that cannot be effectively mitigated (Article 57(11) of the AI Act), or at the motivated request of the participant. It informs the AI Office of the decision and the reasons (Article 8(2)).
What you receive at the end, and what it is worth
After completion the authority evaluates the project against the objectives in the plan. There are two documents (Article 57(7) of the AI Act and Article 6 of Implementing Regulation (EU) 2026/2487).
The written proof is provided on request. It lists the activities successfully carried out, the legal requirements and obligations evaluated, and proof of whether the project was completed. It may be drafted in collaboration with other authorities involved and signed by each of them.
The exit report goes to every participant, within a reasonable period and no later than 45 working days after completing participation. It describes the key regulatory issues, how they were resolved, any serious incidents and the lessons others can use.
What are those documents worth? Three things are certain.
First, they are not a declaration of conformity. They do not have the same status or legal effect as the EU declaration of conformity of Article 47 (Article 6(2) and (5)). Recital 20 of the Implementing Regulation says that participation and exit are not an approval to place the system on the market or put it into service, and that exiting provides no presumption of conformity.
Second, they do count. Market surveillance authorities and notified bodies take the exit report and the written proof positively into account, with a view to accelerating conformity assessment procedures to a reasonable extent (Article 57(7) of the AI Act).
Third, they hold across the Union. Market surveillance authorities and notified bodies from any Member State recognise both documents as documentation of what was done and learned (Article 6(9)). In our reading, a project in another Member State is therefore worth no less.
The exit report may be published on the Commission's information platform if the participant and the authority explicitly agree. If the participant does not agree, the authority may still publish the key findings and lessons learned, without confidential information (Article 6(7) and (8)).
Fines and damage during the test
Protection during the test is in Article 57(12) of the AI Act, not in Article 58. If a prospective provider observes the plan and the terms and conditions and follows the guidance of the authority in good faith, the authorities impose no administrative fines for infringements of the AI Act. For other law, such as the GDPR, that only holds if the authority responsible for that law was actively involved in the supervision in the sandbox and provided guidance for compliance. On fines the text names prospective providers only. In our reading it is therefore not certain that the same protection covers a provider testing a system that is already on the market.
Damage to third parties is not covered. Whoever causes damage during the experiment must compensate it under the ordinary rules. The Implementing Regulation repeats that participants remain responsible for ensuring compliance with Union law (Article 2(6)).
In our reading the sandbox does not lift the prohibitions of Article 5 either. The supervisory and corrective powers of the authorities remain fully in place (Article 57(11)).
Personal data: the GDPR keeps applying
If the system processes personal data, the authority ensures that the data protection authority is associated with the sandbox and supervises its own field (Article 57(10) of the AI Act and recital 11 of the Implementing Regulation). In the Netherlands that is the Autoriteit Persoonsgegevens. The sandbox plan states whether personal data are processed.
A sandbox changes nothing about your own obligations under the GDPR. You need a legal basis, you inform data subjects and where required you carry out a data protection impact assessment (Article 35 GDPR).
The AI Act adds one thing. Article 59 allows personal data lawfully collected for other purposes to be processed in the sandbox for developing, training and testing certain AI systems in the public interest, for instance for the efficiency and quality of public administration and public services. Ten conditions apply (paragraph 1), including these. The data are necessary for complying with the requirements for high-risk AI systems and that cannot be done with anonymised or synthetic data. There are monitoring mechanisms to identify high risks to the rights and freedoms of the data subjects as referred to in Article 35 GDPR, and the processing can then be stopped at once. The data sit in a functionally separate, isolated and protected environment under the control of the prospective provider. Personal data created in the sandbox are not shared outside the sandbox. The processing does not lead to measures or decisions affecting the data subjects. The data are deleted afterwards. And the authority publishes a short summary of the project on its website. According to recital 140 of the AI Act all other obligations of the controller and all rights of data subjects remain applicable.
Main example: the municipality and the supplier
Back to Waterveen. The system helps assess whether a resident is entitled to special assistance benefits. In our reading that falls in principle under Annex III, point 5(a) (evaluating the eligibility of natural persons for essential public assistance benefits), unless the exception of Article 6(3) applies.
Who applies? In our reading the software company, as prospective provider, where the system is supplied under its name or trademark. The municipality takes part as deployer, and its role is set out in the sandbox plan. If the municipality has the system developed itself and puts it into service under its own name, it is itself the provider and can apply itself (Article 3(3)).
What does it cost? The software company is an SME, provided it stays under the thresholds together with its partner enterprises and linked enterprises. Then guidance, advice and supervision are free of charge and it has priority. The same pricing regime applies to the municipality.
Does the project stand a chance? It concerns a public service and a question many municipalities struggle with: how do you organise human oversight for decisions on assistance benefits? That fits the criteria of public interest and learning for the sector. Whether it is selected also depends on the capacity of the authority.
And the data? If the company wants to test with old benefit files, those are personal data collected for another purpose. That is possible under the conditions of Article 59. Outside that route the GDPR applies in full: a legal basis is needed and the new purpose must in principle be compatible with the purpose for which the data were collected (Article 6(4) GDPR). If the files contain health data, Article 9 GDPR applies as well. The municipality remains the controller for its files and checks whether a data protection impact assessment is required (Article 35 GDPR). Under the route of Article 59 the processing may not lead to measures or decisions affecting real applicants (paragraph 1, point (f)).
What use is the exit report to the municipality? It shows which requirements were evaluated and what the authority thought of them. That helps with procurement and with its own fundamental rights impact assessment (Article 27). It does not replace that assessment, nor registration in the EU database (Article 49(3) and Article 26(8)). The rules for high-risk AI systems under Annex III apply from 2 December 2027. For a system placed on the market or put into service before that date, the rules apply only after a significant change in its design; if it is intended to be used by public authorities, it must comply by 2 August 2030 at the latest (Article 111(2)).
Example 2: the bank
A bank builds its own model that evaluates the creditworthiness of private individuals. In our reading that is in principle a high-risk AI system (Annex III, point 5(b)).
The bank is not an SME. It has no priority, and the authority may charge the full costs. The list of fees should be known when the application phase opens.
For high-risk AI systems that financial institutions use in providing their financial services, the financial supervisor is the market surveillance authority (Article 74(6)). The protection against fines under other law only covers the law whose authority was actively involved and provided guidance. So ask in advance which authorities will be at the table and record that in the plan. On the GDPR side, Article 22 GDPR on automated decisions and Article 35 GDPR are the first points of attention. The AI Act itself does not provide a legal basis for a fully automated decision within the meaning of Article 22(2)(b) GDPR (recital 140 of the AI Act). If the bank uses its own model, it is also the deployer and performs a fundamental rights impact assessment before putting the system into use (Article 27(1)).
Example 3: the university of applied sciences
A university of applied sciences develops its own system that marks open exam questions. In our reading that is in principle a high-risk AI system (Annex III, point 3(b), evaluating learning outcomes). The institution is then a prospective provider and can apply itself. If it later uses the system itself, it may in our reading, as an entity providing a public service, also have to perform a fundamental rights impact assessment (Article 27(1)).
Whether it falls under the pricing regime for public authorities, the Implementing Regulation does not say in so many words. Ask before you start; the costs must be known before participation begins.
If the institution also wants to measure how tense students are during an exam, the sandbox is of no help. In our reading that is emotion recognition as soon as the tension is inferred from biometric data, such as facial expression or voice (Article 3(39)). Emotion recognition in education institutions is prohibited, except for medical or safety reasons (Article 5(1)(f)).
Example 4: the start-up with a selection tool
A start-up with eight employees builds a tool that filters job applications (Annex III, point 4(a), in our reading in principle high-risk). It is an SME, has a registered office in the Union and so gets priority and free guidance if it meets the criteria. It can apply together with a first customer. That customer is the deployer, and its role goes into the plan.
The start-up does need a real question. The need to participate is the first selection criterion. A general wish "to become compliant" is in our reading too thin; a concrete question, for instance about testing for bias with little data, stands a better chance.
And the Netherlands?
The Netherlands must have a national sandbox operational by 2 August 2027 (Article 57(1)). The Autoriteit Persoonsgegevens and the Rijksinspectie Digitale Infrastructuur published a design proposal on 25 March 2025 for one sandbox for all sectors. Under the draft AI Act Implementation Act (Uitvoeringswet AI-verordening), which was open for consultation from 20 April to 1 June 2026, the market surveillance authorities are jointly responsible for the sandbox.
The Dutch set-up must now fit within the Implementing Regulation. It applies directly and does not need to be transposed into Dutch legislation. According to recital 6, pilot projects are encouraged, also before the Member State has formally designated the competent authorities. Whether and when the Dutch sandbox accepts applications is for the AP and the RDI to announce.
From when?
| Rule | Date |
|---|---|
| Articles 57 and 58 of the AI Act (Chapter VI) | applicable since 2 August 2026 |
| Amendment of Article 58(1) by Regulation (EU) 2026/1744 | in force since 27 July 2026; Article 58 itself has applied since 2 August 2026 (Article 113) |
| Implementing Regulation (EU) 2026/2487 | adopted 10 November 2026, published 11 November 2026, in force on the twentieth day following publication (by our calculation 1 December 2026) |
| National sandbox operational | by 2 August 2027 |
| Requirements for high-risk AI systems under Annex III | 2 December 2027 |
| Requirements for high-risk AI systems in products under Annex I | 2 August 2028 |
What to do now
If you are a provider or develop a system yourself:
- Formulate the legal question you cannot resolve. Without a concrete question the chance of selection is small.
- Check whether your system is eligible: not yet on the market, a substantial modification, or a modified intended purpose that makes it a high-risk AI system.
- Establish whether you are an SME or a public authority, because that determines the costs and the priority.
- Consider which deployer you want to apply with and what each role will be.
- Map whether you need personal data and on which legal basis.
- List what is confidential, so you can identify it in the sandbox plan.
If you are a deployer, such as a municipality that procures:
- Ask your supplier whether it is considering a sandbox project and whether you can take part.
- For a system that has been through a sandbox, ask for the exit report and the written proof, and read which requirements were and were not evaluated.
- Treat those documents as supporting evidence, not as a quality mark. Your own obligations stay the same.
Want to know which obligations apply to your own situation? The case check lists them on the basis of your answers.
Frequently asked questions about the rules of the AI regulatory sandbox
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