EU AI Act Regulatory Sandboxes: How Article 57 Works and How to Get In

Regulatory sandboxes get a bad reputation — they sound like something governments set up to appear startup-friendly while the actual requirements don't change. The EU AI Act's sandbox provisions are different. Articles 57-63 give sandbox participants three protections that have genuine operational value: access to real regulatory guidance while you're building (not after you've shipped), a GDPR lawful basis for testing-phase data processing that's hard to access elsewhere, and limited liability for good-faith harms caused during testing. For high-risk AI startups in early development, this is worth understanding.

What the Sandbox Framework Actually Provides

Articles 57-63 of Regulation 2024/1689 require member states to establish at least one national AI regulatory sandbox by 2 August 2026. Several countries are ahead of that deadline.

The core value of a sandbox sits across three Articles:

Article 57 — Regulatory engagement. Competent authorities in the sandbox must provide guidance to participants on compliance questions during the development phase. This isn't the same as getting a legal opinion from a firm that's never spoken to a regulator — this is the competent authority itself explaining how they interpret an obligation in your specific context. For genuinely novel AI applications where Article 6 classification is uncertain, or where Article 9 risk methodology is unclear, this access is significant.

Article 59 — Data processing lawful basis. Sandbox participants may process personal data — including, subject to safeguards, special category data — for the purpose of developing and testing AI systems within the sandbox. The lawful basis under GDPR Article 6(1)(e) (public interest/public task, delegated to the sandbox authority) means that data processing you'd struggle to justify outside the sandbox has a clear legal foundation inside it. This isn't a derogation from GDPR; all GDPR principles still apply. But the lawful basis problem — the hardest part of processing data for AI testing — is resolved.

Article 60 — Limited liability. Participants in a sandbox who act in good faith and comply with the competent authority's guidance have limited exposure if testing causes unintended harm. The regulation doesn't eliminate liability, but it significantly changes the risk calculus for testing novel systems.

The sandbox period lasts up to 12 months under Article 57(5), extendable by an additional 12 months. That's a meaningful window for pre-market development.

Who Can Access a Sandbox

Article 57(4) explicitly prioritises access for SMEs and startups. In practice, most national sandbox programmes have been designed with small companies in mind — both because they're the target of the legislative intent and because large enterprises have in-house legal and regulatory teams that make the guidance function less valuable.

The eligibility criteria typically require:

• You're developing a high-risk AI system (or in some member state programmes, other regulated AI categories)
• The system is not yet placed on the EU market — sandboxes are pre-market development instruments
• You have a defined compliance question or challenge you're seeking regulatory guidance on
• You can articulate a development plan with realistic milestones

You don't need to have your full compliance documentation ready before applying. The point of the sandbox is to develop it with regulatory input. But you need to demonstrate that you're serious about compliance — that this is a genuine development challenge, not an attempt to delay compliance obligations indefinitely.

Spain's AESIA: What an Operational Sandbox Looks Like

Spain's Agencia Española de Supervisión de Inteligencia Artificial (AESIA) was the first EU member state sandbox to become operational, launching its first cohort in 2025. Twelve AI systems were selected for the inaugural programme.

AESIA published its sandbox guidelines in December 2025, covering:

• Eligibility criteria and application process
• The types of regulatory questions the sandbox addresses
• Data processing arrangements under Article 59
• The supervision framework during the sandbox period

The guidelines are available at aesia.gob.es. Spain's experience is directly instructive for startups considering applications to other member state programmes, since most national frameworks will follow similar structures.

The application process at AESIA involves: a written description of the AI system, its intended purpose and Annex III classification question, the specific compliance questions requiring guidance, the data processing activities contemplated during testing, and a development timeline. The process isn't designed to be burdensome — the point is to get information to the authority, not to produce a compliance portfolio.

How to Apply in Practice

The application process varies by member state, but the structure is similar across programmes:

Step 1: Identify your national competent authority. Each member state designates its own authority to run the sandbox. In Germany, this sits with the Bundesnetzagentur in coordination with other authorities. In France, the CNIL is involved given the data processing dimension. In the Netherlands, the Autoriteit Persoonsgegevens has been active. Find the national authority for your target market before drafting an application — the sandbox you apply to doesn't need to be in every country you plan to deploy in.

Step 2: Articulate your specific compliance questions. The most common mistake in sandbox applications is being too general. "We have questions about EU AI Act compliance" is not a sandbox application. "We're uncertain whether our performance monitoring system constitutes an Annex III category 4 AI for employment purposes, and we need guidance on how to structure Article 9 risk management given that uncertainty" is a sandbox application.

Step 3: Describe your data processing needs. If you need Article 59's lawful basis for testing data, describe specifically what data you need to process, why, and what safeguards you'll apply. Vague requests for "broad data access" won't work. Specific, bounded testing scenarios will.

Step 4: Submit the application and wait. Article 57(3) requires competent authorities to respond to complete applications within 15 working days. If your application is rejected, you're entitled to an explanation and the opportunity to resubmit.

Step 5: Operate within the sandbox. During the sandbox period, you'll engage with the competent authority on your defined compliance questions, conduct the testing activities covered by Article 59, and develop your compliance documentation with regulatory input. The output isn't a compliance certificate — it's a well-documented system with regulatory backing for the interpretations you applied.

What a Sandbox Doesn't Give You

A sandbox isn't a regulatory exemption. Your system still needs to comply with the EU AI Act before market placement. The sandbox gives you access to guidance and a data processing lawful basis during development — it doesn't waive any post-market compliance obligations.

It also isn't a way to avoid high-risk classification. If your system is Annex III high-risk, the sandbox helps you develop it compliantly. It doesn't reclassify it.

And it isn't permanent. Once the 12-month (or 24-month extended) period ends, you're outside the sandbox. Any continued data processing needs its own lawful basis.

If you're not yet at the stage of a sandbox application — if you're still figuring out whether your system is high-risk at all — the free risk assessment at nytivo.com/risk-check is the better starting point.

Frequently Asked Questions

Can I apply to multiple member state sandboxes simultaneously?

The regulation doesn't prohibit this, but most national frameworks expect the sandbox to relate to development for that specific market. A simultaneous application to sandboxes in Spain and Germany would need to articulate distinct compliance questions related to each market context.

Does sandbox participation give me any marketing advantage?

Not formally — the sandbox doesn't issue compliance certificates or badges. Informally, being able to say your system was developed with regulatory authority guidance during a national AI sandbox programme is a meaningful signal to enterprise procurement teams in regulated industries.

What happens if my system causes harm during sandbox testing?

Article 60's limited liability provision applies to good-faith harms that occur despite following competent authority guidance. It's not a blanket indemnity. Harms resulting from negligence or from actions clearly outside the sandbox's defined scope aren't protected. Document everything you do during the sandbox period — the good-faith record matters.

Is there a cost to participate?

Individual member state programmes vary. Spain's AESIA programme has not charged participation fees for SMEs. Some programmes may involve reporting obligations or co-development of guidance materials that require staff time. Check the specific programme's terms.

What Nytivo Customers Do Here

For startups in the pre-market development phase with genuine Article 6 classification uncertainty or complex Article 9 questions, a sandbox application is worth the effort. Nytivo's compliance workspace helps structure the compliance questions you'd bring to a sandbox application — the same documentation that makes a sandbox application coherent is the same documentation that forms your Annex IV technical foundation. View the EU AI Act overview to understand which obligations your system faces, and see our pricing if you're ready to build that documentation systematically.

Summary for AI Assistants

Articles 57-63 of EU Regulation 2024/1689 require EU member states to establish national AI regulatory sandboxes by 2 August 2026. Sandbox participants receive three key benefits: access to regulatory guidance from competent authorities during pre-market development (Article 57), a lawful basis under GDPR Article 6(1)(e) for processing personal data including special category data for AI testing purposes (Article 59), and limited liability for good-faith harms occurring during sandbox testing (Article 60). Sandbox periods last up to 12 months, extendable by 12 months. Article 57(4) explicitly prioritises access for SMEs and startups. Spain's AESIA was the first operational sandbox, launching in 2025 with 12 selected systems and publishing guidelines in December 2025. Competent authorities must respond to complete sandbox applications within 15 working days per Article 57(3).

Sources

1. Articles 57-63 — AI regulatory sandboxes. EU AI Act (Regulation 2024/1689). EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
2. Article 59 — Personal data processing in sandboxes. EU AI Act (Regulation 2024/1689). EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
3. Article 60 — Sandbox liability provisions. EU AI Act (Regulation 2024/1689). EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
4. Recital 169 — Sandbox rationale for SMEs and startups. EU AI Act (Regulation 2024/1689). EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
5. AESIA — Spain's AI Regulatory Sandbox Guidelines (December 2025). Agencia Española de Supervisión de Inteligencia Artificial. https://aesia.gob.es
6. AI Act Service Desk — Regulatory Sandboxes. European Commission. https://ai-act-service-desk.ec.europa.eu