Accelerating AI Innovation: EU Regulatory AI Development Sandboxes

23 June 2025

The rapid advancement of Artificial Intelligence presents both immense opportunities and complex regulatory challenges.

Recognising the need to foster innovation while ensuring safety and fundamental rights, the European Union’s AI Act introduces a crucial mechanism: AI regulatory sandboxes.

These sandboxes are designed to provide a controlled environment for the development and testing of innovative AI systems, aiming to accelerate their efficient integration into the market while upholding the EU’s human-centric approach to AI.

Purpose and Objectives of AI Regulatory Sandboxes

AI regulatory sandboxes are established by competent authorities to create a controlled environment for experimentation and testing of innovative AI systems before they are released to the market or put into service.

This framework is intended for a limited time and operates under regulatory supervision. The establishment of these sandboxes aims to achieve several key objectives:

Enhance legal certainty: They help innovators understand and comply with the AI Act and other relevant laws.
Support best practice sharing: They foster cooperation among authorities involved in the sandbox.
Foster innovation and competitiveness: By providing a safe space for experimentation, sandboxes facilitate the development of a thriving AI ecosystem.
Contribute to evidence-based regulatory learning: They enable authorities and undertakings to learn from practical application, which can inform future adaptations of the legal framework.
Facilitate and accelerate access to the Union market: This is particularly beneficial for AI systems provided by Small and Medium-sized Enterprises (SMEs), including start-ups, by removing potential barriers.

Member States are mandated to establish at least one AI regulatory sandbox at the national level, which must be operational by 2 August 2026.

The Commission can provide technical support, advice, and tools for their establishment and operation.

Key Features and Operational Aspects

The functioning of AI regulatory sandboxes is structured to balance innovation with oversight:

Controlled Environment: Sandboxes offer a supervised setting for the development, training, testing, and validation of AI systems. This can include testing in real-world conditions under supervision within the sandbox.
Guidance and Supervision: Competent authorities provide guidance, supervision, and support to participants. This includes identifying risks, testing, implementing mitigation measures, and assessing their effectiveness in relation to the AI Act’s obligations. They also guide providers on regulatory expectations.
Documentation and Accountability: Upon successful completion of activities, the competent authority provides a written proof and an exit report detailing the activities, results, and learning outcomes. This documentation can positively contribute to demonstrating compliance during conformity assessment procedures.
Personal Data Processing: Crucially, personal data lawfully collected for other purposes may be processed within the sandbox. This is permissible solely for the development, training, and testing of specific AI systems that serve a substantial public interest. Strict conditions apply, including necessity, effective monitoring mechanisms, isolated processing environments, and the deletion of data after use.
Protection from Administrative Fines: Providers and prospective providers participating in the sandbox are generally not subject to administrative fines for infringements of the AI Act, provided they observe the specific plan and terms of participation, and follow the guidance given by the national competent authority in good faith.
Cross-Border Cooperation: Sandboxes are designed to facilitate cross-border cooperation between national competent authorities, with coordination activities taking place within the European Artificial Intelligence Board.

Accessibility and Support for SMEs

Recognising the particular needs of SMEs and start-ups, the AI Act includes specific provisions to facilitate their participation:

Priority Access: SMEs, including start-ups with a registered office or branch in the Union, are given priority access to AI regulatory sandboxes, provided they meet eligibility and selection criteria.
Cost Considerations: Access to the sandboxes is free of charge for SMEs, although national competent authorities may recover exceptional costs in a fair and proportionate manner.
Tailored Support: Member States are encouraged to organise specific awareness-raising and training activities tailored to the needs of SMEs. Dedicated channels for communication are to be utilised or established to provide advice and respond to queries about the implementation of the Regulation.
Streamlined Processes: The detailed arrangements for sandboxes ensure that procedures for application, selection, participation, and exit are simple, easily intelligible, and clearly communicated to facilitate participation by SMEs with limited legal and administrative capacities.
Referral to Support Services: Prospective providers, especially SMEs and start-ups, are to be directed to pre-deployment services, such as guidance on the AI Act, assistance with standardisation and certification, and access to testing and experimentation facilities, European Digital Innovation Hubs, and centres of excellence.

Role of Competent Authorities and Related Measures

Member States must ensure their competent authorities have adequate technical, financial, and human resources to effectively fulfil their tasks.

National data protection authorities and other relevant authorities should be involved in the operation of the sandbox, particularly when personal data processing or other areas within their jurisdiction are involved.

Importantly, the establishment of AI regulatory sandboxes does not affect the supervisory or corrective powers of these competent authorities.

They retain the power to temporarily or permanently suspend testing or participation if significant risks are identified and cannot be effectively mitigated.

Annual reports on the progress and results of sandboxes, including best practices and lessons learned, are to be submitted to the AI Office and the Board. The Commission, through the AI Office, will maintain a public list of planned and existing sandboxes to encourage interaction and cross-border cooperation.

Testing in Real-World Conditions (Outside Sandboxes)

Beyond the structured environment of regulatory sandboxes, the AI Act also provides a specific regime for testing high-risk AI systems in real-world conditions outside sandboxes. This applies to high-risk AI systems listed in Annex III.

Such testing requires a real-world testing plan, which must be submitted to and approved by the relevant market surveillance authority.

Key conditions for such testing include:

The provider or prospective provider must be established in the Union or have an authorised legal representative.
Data collected for testing must be transferred to third countries only with appropriate safeguards.
Testing duration is limited to six months, extendable once for another six months.
Special protection for vulnerable groups like children or persons with disabilities.
Informed consent from human subjects involved in the testing is generally required, after they receive clear information about the testing’s nature, objectives, risks, and their rights. An exception exists for law enforcement where seeking consent would prevent testing, provided the testing does not negatively affect subjects and their data is deleted afterwards.
The predictions, recommendations, or decisions of the AI system must be effectively reversible and disregardable.
Any serious incident during testing must be reported to the national market surveillance authority, requiring immediate mitigation or suspension of testing.

The Path to Trustworthy and Efficient AI

The EU AI Act’s provisions for regulatory sandboxes and real-world testing demonstrate a proactive approach to enabling responsible AI innovation. By providing structured environments for experimentation, offering tailored support for SMEs, and integrating rigorous oversight, the EU seeks to ensure that AI development is efficient, ethical, and trustworthy.

This framework aims to foster competitive advantages for businesses by cultivating public trust and mitigating risks, ultimately contributing to the healthy growth of the AI ecosystem within the Union.

Businesses, particularly small and medium-sized enterprises (SMEs), are encouraged to leverage these support mechanisms to navigate the new regulatory landscape and bring human-centric AI innovations to market.

Looking ahead, the sandbox model may set a precedent for other jurisdictions exploring how to balance regulatory certainty with flexibility.

As regulatory convergence becomes increasingly important for global AI deployment, the EU’s sandbox experience could inform international standards and best practices.

The lessons learned from sandbox operations are likely to influence future iterations of the AI Act, as well as inspire sector-specific frameworks that extend beyond the current scope.

For AI developers, engaging with sandboxes not only facilitates compliance but also opens channels for constructive dialogue with regulators, helping to shape the trajectory of AI governance in a meaningful way.