- All news
- AI Act: common rules for AI regulatory sandboxes
AI Act: common rules for AI regulatory sandboxes
The AI Act's regulatory sandboxes now have common operating rules. Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026, published in the Official Journal L series on 11 November 2026, is adopted under Article 58(1) of Regulation (EU) 2024/1689 and enters into force on the twentieth day after publication, 1 December 2026. It applies to every sandbox set up under Article 57: national, joint, the one the European Data Protection Supervisor may run, and the Union-level sandbox the AI Office may establish. For a provider developing an AI system, it settles who may apply, what participation costs, what goes into the sandbox plan, and what the exit documents are worth in a later conformity assessment.
In short:
- Eligibility is limited to three cases: an AI system not yet placed on the market or put into service, one already on the market that will be substantially modified, or one that becomes high-risk under Article 25(1)(c).
- Free for SMEs, including start-ups, for regulatory guidance, advice and supervision; other undertakings may be charged full cost, and the fee list must be published when applications open.
- Non-EU providers without a registered office or branch in the Union pay a participation fee and must name an authorised representative by written mandate.
- Selection decision within three months of a complete application, with a right of appeal.
- Exit report within 45 working days after participation ends. Market surveillance authorities and notified bodies in every Member State must recognise the exit report and written proof, but neither is a declaration of conformity.
- National sandboxes are due by 2 August 2027, the date set by Article 57(1) as amended by Regulation (EU) 2026/1744.
What the regulation covers
Section titled “What the regulation covers”The regulation has 26 recitals and 10 articles in four chapters: general provisions, participation, operation of a sandbox project, and the final provision on entry into force. Article 1(2) applies it "fully" to sandboxes established under Article 57(1), (2), (3) and (3a) of the AI Act, and recital 5 explains why: uniform application and mutual recognition of results across all sandboxes.
The sandboxes themselves come from Article 57. As amended by Regulation (EU) 2026/1744, Article 57(1) requires each Member State to ensure that its competent authorities establish at least one AI regulatory sandbox at national level, "operational by 2 August 2027", possibly jointly with other Member States. Article 57(3a), added by the same regulation, lets the AI Office establish a sandbox at Union level for AI systems covered by Article 75(1), with priority access for SMEs, including start-ups, and small mid-cap enterprises. Recital 6 of the implementing regulation repeats the 2 August 2027 date and encourages pilot projects before competent authorities are formally designated.
Who can apply
Section titled “Who can apply”Under Article 3(1), any provider or prospective provider may apply, using a form the competent authority makes available. Article 3(2) sets two cumulative eligibility criteria:
- The applicant is the provider or prospective provider, or, for a provider with no registered office or branch in the Union, its authorised representative jointly with it. Applications may be made in partnership with deployers or other third parties.
- The AI system is one of the following: (1) not placed on the market or put into service; (2) already placed on the market or put into service, but to be subject to a substantial modification within the meaning of Article 3(23); (3) not classified as high-risk according to its intended purpose, but then placed on the market or put into service in such a way that it becomes high-risk under Article 25(1)(c).
A provider is not restricted to the sandbox of the Member State where it is established. Article 3(3) lists the minimum selection criteria: the need to participate given specific regulatory or practical challenges, the potential for regulatory learning and its scalability, the contribution to public interest, societal benefit and Union policy, the maturity of the proposal, and the innovative character of the AI system. Authorities may periodically prioritise sectors (Article 3(4)) and must inform the applicant of the outcome "no later than three months" after the complete application or after the deadline of a call (Article 3(7)).
What participation costs
Section titled “What participation costs”Article 2 sets the pricing regime:
| Participant | Rule |
|---|---|
| SMEs, including start-ups, and public bodies | Free of charge for regulatory guidance, advice and supervision, including supervision of real-world testing and support for conformity assessment. Exceptional costs for other agreed services may be recovered, with prior notice and a detailed invoice. |
| Other undertakings | The authority may charge a fee equal to the full costs of the services; it must publish the fee list when it opens the application phase. |
| Providers with no registered office or branch in the Union | Pay a participation fee and give details of an authorised representative established in the Union, designated by written mandate, who can be held responsible and liable as a guarantee for the provider's activities in the sandbox. |
Article 2(6) adds that participation does not lift the participant's responsibility for complying with applicable Union legislation.
The sandbox plan
Section titled “The sandbox plan”After selection, the provider and the competent authority agree a sandbox plan "without undue delay" (Article 5(1)). Article 5(2) lists its minimum content: the applicant and the AI system; an estimated timeline with an end date; objectives, scope, roles and the services provided, including those subject to fees or exceptional costs; the AI Act requirements and other Union or national rules supervised in the project; whether personal data is processed and the data protection authority needs to be involved; risk management safeguards and a serious incident procedure where appropriate; and, where applicable, a plan for testing in real-world conditions.
Confidentiality is the provider's job to flag. Article 5(3) asks the provider to identify, "to the best of their ability", which information in the plan may be confidential under Article 78. This matters because Article 6(8) presumes information available for disclosure unless it was stated as confidential in the plan or explicitly indicated as such before the end of the project.
The project can be extended, where both sides agree and the extension is justified, until 30 calendar days before the planned end date (Article 8(1)). The authority may suspend testing or participation where no mitigation is possible, for example after a breach of the plan or in the presence of significant risks to health, safety or fundamental rights that cannot be effectively mitigated (Article 8(2)).
Exit report and written proof
Section titled “Exit report and written proof”Article 57(7) of the AI Act already provided for two documents: a written proof of the activities successfully carried out in the sandbox, issued on request of the provider, and an exit report on the activities, results and learning outcomes. It lets providers use them "to demonstrate their compliance with this Regulation through the conformity assessment process or relevant market surveillance activities", and requires market surveillance authorities and notified bodies to take them "positively into account" with a view to accelerating conformity assessment "to a reasonable extent".
Article 6 of the implementing regulation fixes their content and weight:
- Written proof (Article 6(2)): the types of activities successfully carried out, the legal requirements evaluated, and whether the project was completed. It "shall serve as documentation evidence informing notified bodies and relevant market surveillance authorities", but does not have the status or legal effect of a declaration of conformity under Article 47 of the AI Act or any other legislation supervised in the sandbox.
- Exit report (Article 6(3) and (5)): the key regulatory issues examined and how they were resolved, any serious incidents, and lessons learned. It is drafted and signed by the competent authority, has the same limit on legal effect, and must reach the participant "no later than 45 working days after completing participation".
- Recognition (Article 6(9)): market surveillance authorities and notified bodies "from any Member State" shall recognise both documents as documentation of the activities carried out, their results and learning outcomes.
Article 57(12) of the AI Act sets the other benefit of participation: participants remain liable for damage to third parties, but no administrative fines are imposed for infringements of the AI Act where the prospective provider observes the plan and the terms of participation and follows the authority's guidance in good faith.
Products under Section B of Annex I
Section titled “Products under Section B of Annex I”Article 5(6) has a specific rule for AI systems covered by the Union harmonisation legislation in Section B of Annex I: sandbox projects take account of the specificities of those products, and Member States may enable their testing in real-world conditions, including by giving access to physical public infrastructure, to verify conformity with Articles 8 to 15. That rule is expressly "subject to Article 2(2)" of the AI Act, which, as amended by Regulation (EU) 2026/1744, applies Articles 57, 58 and 59 to those products "only in so far as" the AI Act requirements have been integrated into the sectoral legislation. Since the same amending regulation moved the Machinery Regulation (EU) 2023/1230 to Section B, point 21, this applies to AI in machinery too.
What to do
Section titled “What to do”- Check eligibility first. A sandbox is open to a system not yet on the market, one heading for a substantial modification, or one that becomes high-risk under Article 25(1)(c). A system already on the market and unchanged does not qualify.
- Watch national calls. Each Member State must have a sandbox operational by 2 August 2027; the application form, the selection rounds and, for non-SMEs, the fee list are published by the competent authority when the application phase opens.
- Prepare the non-EU set-up early. A provider with no registered office or branch in the Union needs an authorised representative with a written mandate before it can apply.
- Mark confidential information in the sandbox plan. Anything not flagged as confidential before the end of the project is presumed available for disclosure.
- Ask for the written proof and file both documents. Request the written proof under Article 57(7), keep it and the exit report with the technical documentation, and present them to the notified body or market surveillance authority as evidence, not as a substitute for the declaration of conformity.
Going further
Section titled “Going further”- AI Act and products: AI embedded in an already regulated product, requirements and conformity assessment routes
- AI Act: applicable, but high-risk duties are deferred: the Article 113 timeline and the move of machinery to Section B after Regulation (EU) 2026/1744
- AI Act: EN 18286 is the first harmonised standard cited: the quality management system standard and its presumption
- Substantial modification: when to re-certify in the EU: when a change to a product already on the market restarts conformity assessment
Sources & references
- Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 laying down rules for the application of Regulation (EU) 2024/1689 as regards the establishment, development, implementation, operation and supervision of AI regulatory sandboxes (OJ L, 11 November 2026), Publications Office of the European Union data.europa.eu/eli/reg_impl/2026/2487/oj
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence, EUR-Lex eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230, EUR-Lex eur-lex.europa.eu/eli/reg/2026/1744/oj