EU sets rules for AI regulatory sandboxes starting December 1

The sandbox framework

The European Commission has adopted common rules for AI regulatory sandboxes, which are controlled environments where companies can develop and test AI systems under official supervision. These rules, which take effect on December 1, aim to provide a structured way for providers to navigate the requirements of the EU AI Act. The framework is designed to help authorities understand how to oversee real-world testing while helping companies identify regulatory hurdles early.

Under these rules, a project is eligible if an AI system has not yet been released, is undergoing substantial modification, or has transitioned into a high-risk use case. Authorities will evaluate applications based on five criteria: the specific regulatory challenges faced, the potential for sector-wide learning, the contribution to public interest, the maturity of the proposal, and the system's innovativeness. This ensures that the most impactful and well-prepared projects receive the resources they need.

While the rules are now set, each EU Member State has until August 2, 2027, to have at least one national sandbox operational. The AI Office may also run a Union-level sandbox for certain systems, and the European Data Protection Supervisor can establish a sandbox for EU institutions. This multi-layered approach allows for both broad and specialized testing environments across the continent.

Evidence vs. clearance

A critical distinction in the new rules is what a company actually receives at the end of a sandbox project. When a project concludes, the supervising authority provides an exit report and written proof of completed activities. These documents must be delivered within 45 working days and cover the regulatory issues examined, tools used, and a log of any serious incidents. However, the regulation explicitly states that neither document carries the legal status of an EU declaration of conformity.

This means that even after successfully completing a sandbox program, companies must still undergo the standard conformity procedures before they can legally market their AI products. The AI Act does ask market surveillance authorities to take these documents into account positively, with the goal of speeding up the assessment process. The intent is to provide a record of work done rather than a legal shortcut to market entry.

Because the sandbox does not provide a presumption of conformity, some experts worry it may struggle to attract certain providers. Companies can choose to pursue standard assessment routes instead, which may offer more direct legal certainty. The gap between having evidence of testing and receiving official clearance remains a significant hurdle for firms looking to sell AI in Europe.

Liability and protections

One of the primary incentives for joining a sandbox is a specific protection against administrative fines. The AI Act bars these fines for infringements if a participant follows the agreed-upon sandbox plan and the authority's guidance in good faith. This 'shield' is tied directly to the plan, which can only be changed if both the authority and the company agree to the modifications during the project.

However, this protection is not absolute and does not cover liability for harm to third parties. If an AI system causes damage to a person or property, the company remains legally responsible regardless of its participation in a sandbox. Critics have noted that this full third-party liability might deter developers from joining the program, particularly for high-risk applications in sensitive fields like healthcare or employment.

To address these concerns, some have suggested that mandatory insurance for third-party harm could make sandboxes more attractive to developers. The current regulation, however, contains no such insurance requirement. For many companies, the decision to join will depend on whether the benefits of supervised testing and the fines shield outweigh the continued risk of third-party liability.

SME support and data

The rules place a heavy emphasis on supporting smaller players by offering priority access to SMEs and start-ups. These smaller entities receive free regulatory guidance, advice, and supervision, including support for real-world testing and conformity assessments. This is intended to level the playing field, as smaller firms may lack the extensive legal expertise required to navigate complex oversight from large corporations.

For larger companies, the rules allow authorities to charge participation fees, which must be published clearly when applications open. Public authorities and bodies are subject to the same pricing regime as SMEs. Furthermore, providers without a registered office in the EU must pay participation fees and appoint an authorized representative who can be held responsible for ensuring the conformity of the activities.

Data privacy is also strictly managed within these environments. The AI Act limits the reuse of personal data in sandboxes to five specific public-interest areas, such as public safety, health, and environmental protection. Even in these cases, companies must prove that anonymized or synthetic data is insufficient and must keep the data in a functionally isolated and protected environment that is deleted once the project ends.

Looking ahead

The rollout of these sandboxes will provide the first real-world test of how the EU intends to balance innovation with safety. As the first national calls for applications are published, the industry will see which sectors receive the most attention and what the actual costs of participation will be. These early projects will serve as a blueprint for how authorities handle complex issues like data poisoning, model poisoning, and adversarial examples.

Observers will be watching to see if the exit reports carry enough weight with market surveillance authorities to actually speed up the path to market. If the documents are treated as a significant record of good-faith effort, the sandbox model could become a standard first step for AI deployment. If they are ignored, the sandbox may be viewed more as a learning exercise than a practical tool for commercialization.

Finally, the diversity of national sandboxes will be a key factor to watch. While the common rules aim to prevent 'sandbox arbitrage'—where companies pick the most lenient jurisdiction—different interpretations of the AI Act by various Member States could still create a fragmented landscape. The success of the framework depends on whether it provides a consistent and useful experience for developers across all 27 countries.