AI Regulatory Sandboxes and Real-World Testing

The legal framework

Chapter VI of the AI Act, Regulation (EU) 2024/1689, is called "Measures in support of innovation". It has three parts:

• AI regulatory sandboxes (Articles 57 and 58), with a special rule on personal data (Article 59);
• testing of high-risk AI in real-world conditions outside sandboxes (Articles 60, 60a and 61);
• measures for SMEs and derogations for specific operators (Articles 62 and 63).

Chapter VI applies from 2 August 2026, the general date of application in Article 113. The Digital Omnibus on AI, Regulation (EU) 2026/1744, changed Articles 57, 58, 60 and 63 and added Article 60a, with effect from 27 July 2026. The consolidated AI Act of 27 July 2026 shows the current text.

Sandboxes (Articles 57 and 58)

A sandbox is "a controlled environment that fosters innovation and facilitates the development, training, testing and validation of innovative AI systems for a limited time before their being placed on the market", run under a sandbox plan agreed between the provider and the competent authority (Article 57(5)).

• Who sets them up. Each Member State must ensure at least one national sandbox is operational by 2 August 2027 (Article 57(1), as amended; the original date was 2 August 2026). Member States may set one up jointly, or join an existing one with equivalent national coverage. Regional and local sandboxes are allowed (Article 57(2)). The European Data Protection Supervisor may run one for EU institutions (Article 57(3)).
• An EU-level sandbox. The AI Office may set up a sandbox at Union level for AI systems that it supervises under Article 75(1), with priority access for SMEs, start-ups and small mid-caps (Article 57(3a)).
• What you get. Guidance on regulatory expectations (Article 57(7)). On request, written proof of the activities carried out and an exit report. Market surveillance authorities and notified bodies must take that proof and report positively into account in conformity assessment and market surveillance.
• What you risk. You stay liable under EU and national liability law for damage to third parties caused by the experiments. But if you follow the plan and the terms of participation and follow the authority's guidance in good faith, no administrative fines are imposed for infringements of the AI Act (Article 57(12)). The authority may suspend testing if significant risks cannot be mitigated (Article 57(11)).
• Data protection authorities. Where personal data are involved, the data protection authority is associated with the sandbox and supervises those aspects (Article 57(10)).
• Common rules. The Commission adopts implementing acts on eligibility, application, participation and exit, terms for participants and governance (Article 58(1)). These must ensure that applicants hear within three months (Article 58(2)(a)), that access is free for SMEs and start-ups, except exceptional costs (Article 58(2)(d)), and that participation in one Member State's sandbox is recognised across the EU (Article 58(2)(g)).

Personal data in the sandbox (Article 59)

Article 59 allows personal data lawfully collected for another purpose to be used in the sandbox to develop, train and test AI systems that safeguard a substantial public interest. The areas are listed: public safety and health, the environment, energy sustainability, transport and critical infrastructure, and public administration (Article 59(1)(a)). Ten conditions apply, among them: the data must be necessary because anonymised or synthetic data will not do, they stay in a separate and protected environment, they are deleted at the end, and the processing may not lead to decisions about the people concerned.

Real-world testing outside a sandbox (Articles 60 and 61)

Providers of high-risk AI listed in Annex III, and since 2026 also of high-risk AI under Section A of Annex I, may test in real-world conditions before placing the system on the market, alone or with deployers (Article 60(1) and (2)). Article 60(4) sets eleven conditions, including:

• a real-world testing plan, submitted to and approved by the market surveillance authority; silence for 30 days counts as approval, unless national law requires an express authorisation (points (a) and (b));
• registration of the test in the EU database with a unique identification number (point (c)), visible only to authorities unless you agree otherwise (Article 71(4));
• establishment in the EU or a legal representative in the EU (point (d));
• a maximum of six months, extendable once by six months after notifying the authority (point (f));
• an agreement with each participating deployer on roles and responsibilities (point (h));
• informed consent of the test subjects (point (i)), and outcomes that can be reversed or disregarded (point (k)).

Article 61 lists what test subjects must be told: the nature and objectives of the test, its conditions and duration, their right to withdraw at any time, how to have outcomes reversed and the test's identification number. Consent must be dated and documented, and subjects receive a copy (Article 61(2)). Serious incidents during a test are reported under Article 73 (Article 60(7)). The provider remains liable for damage caused during the test (Article 60(9)).

Article 60a, added in 2026, lets Member States allow real-world testing for AI-enabled products under Section B of Annex I, such as vehicles or aircraft, under a national framework notified to the Commission.

SMEs (Articles 62 and 63)

Member States must give SMEs and start-ups with a registered office or branch in the EU priority access to sandboxes, organise training, provide dedicated channels for questions and help them join standardisation (Article 62(1)). Conformity assessment fees must reflect their size (Article 62(2)). The AI Office provides templates and a single information platform (Article 62(3)). SMEs without partner or linked enterprises may meet certain quality management elements in a simplified way, following Commission guidelines (Article 63(1), as amended). That does not exempt them from Articles 9 to 15, 72 and 73 (Article 63(2)).

In practice

A Liège medtech start-up develops an AI triage tool for emergency departments. It applies to its national sandbox with a hospital as partner (Article 58(2)(b)). Health is a listed public interest, so under Article 59 it can use hospital data collected for care, inside a protected environment, if anonymised data will not do. At the end it asks for an exit report and uses it in its conformity assessment. The tool also falls under the medical devices rules, so the competent authorities for that law should be involved too (Article 57(12) protects against fines under that law only where those authorities were actively involved).

A Ghent HR-tech company wants to test a candidate-ranking tool, an Annex III system, with two employers before launch. It does not need a sandbox. It files a testing plan with the market surveillance authority, registers the test, signs an agreement with each employer, obtains informed consent from candidates and ensures a recruiter can override every ranking. The test may last six months, plus six more after notice.

A Mechelen software firm of 40 people with no partner or linked enterprises uses Article 63 to set up a simpler quality management system for its high-risk product, once the Commission's guidelines are out. It still needs full risk management, data governance and technical documentation.

Practical steps

1. Decide whether you need a sandbox. A sandbox gives guidance and protection from fines. Real-world testing outside a sandbox is quicker to arrange but carries more formal conditions.
2. Prepare the plan. Describe objectives, duration, participants, safeguards and how outcomes can be reversed. Our AI Act register keeps the plan, the registration and the consent records together.
3. Assess data protection first. Article 59 and Article 60 do not replace the GDPR. Run a data protection impact assessment before using personal data.
4. Put the partnership in writing. Article 60(4)(h) requires an agreement with each deployer. Cover roles, instructions for use, incident reporting, data, liability and insurance.
5. Ask for the exit report. It counts in your conformity assessment.
6. Look at funding. Sandboxes and testing often fit research and development grants. See our R&D grants and subsidies service.

Frequently asked questions

When must Belgium have a sandbox?

Every Member State must have at least one national sandbox operational by 2 August 2027 (Article 57(1)). Check the AI Office's public list of planned and existing sandboxes (Article 57(15)).

Does a sandbox protect me from all liability?

No. Liability for damage to third parties stays (Article 57(12)). What you avoid, if you follow the plan and guidance in good faith, is administrative fines under the AI Act.

Can I test a high-risk system on real users without a sandbox?

Yes, under Article 60, if you meet all conditions in Article 60(4), including an approved plan, registration and informed consent.

Is sandbox access free?

For SMEs and start-ups it must be, apart from exceptional costs that may be recovered fairly and proportionately (Article 58(2)(d)).

Can I keep testing after six months?

Once, for six more months, after notifying the market surveillance authority with an explanation (Article 60(4)(f)).

How this connects

This article is part of our series on the EU digital rulebook. See the EU digital rulebook in 2026 for the full map, the Digital Omnibus on AI for the 2026 changes and general-purpose AI: codes, guidelines and standards for model providers. For the conformity assessment that a sandbox prepares, read conformity assessments for high-risk AI.

Our AI Act compliance team prepares sandbox applications and testing plans.

Provisions cited. AI Act, Articles 57(1), (2), (3), (3a), (5), (7), (10), (11), (12) and (15), 58(1) and (2), 59(1), 60(1), (2), (4), (7) and (9), 60a, 61, 62, 63, 71(4), 73 and 113.

Key dates. 27 July 2026: Omnibus amendments in force. 2 August 2026: Chapter VI applies. 2 August 2027: national sandboxes operational.

Sources. AI Act, Regulation (EU) 2024/1689, its consolidated text of 27 July 2026 and Regulation (EU) 2026/1744.