The AI Act is not only about prohibitions and obligations. One of its more practical instruments is the regulatory sandbox: a controlled framework in which companies can test innovative AI systems under the supervision of the competent authority before putting them on the market.
Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, moved the date by which each Member State must have at least one national sandbox operational to 2 August 2027. This article explains what sandboxes are, how they work, who can take part, and what they do and do not change.
What a regulatory sandbox is
A controlled framework, set up and supervised by a national competent authority, in which providers and prospective providers can develop, train, validate and test AI systems for a limited time, on the basis of an agreed sandbox plan, before placing them on the market or putting them into service.
The idea is not new — financial supervisors have run fintech sandboxes for years. The AI Act brings the model to artificial intelligence in Articles 57 to 62. The essential point: a sandbox does not suspend the law. It creates a space in which testing takes place under conditions, with safeguards for fundamental rights and the safety of the people affected.
How it works
- 1Application
The company applies to the competent authority, describing the AI system, the purpose and duration of testing, and how it intends to meet the requirements.
- 2Assessment and the sandbox plan
The authority assesses the application and agrees with the participant a sandbox plan: objectives, conditions, timeframe, methods, the data that may be used and the safeguards.
- 3Supervised testing
The system is tested under the authority’s supervision. The authority provides guidance, may require adjustments, and may suspend testing where significant risks are identified and cannot be mitigated.
- 4Exit
At the end, the authority provides written proof of the activities successfully carried out and an exit report. Providers can use them to demonstrate compliance in the conformity assessment, and market surveillance authorities take them into account.
Who can take part
Providers and prospective providers of AI systems, of any size. The Regulation gives SMEs, including start-ups, priority access, and access is free of charge for them, without prejudice to exceptional costs the authority may recover fairly and proportionately.
Sandboxes are most useful for:
- AI systems that may be high-risk under Annex III: recruitment, credit scoring, biometrics, education, critical infrastructure
- innovative systems whose classification under the risk categories is uncertain
- systems that need to be tested with real data or in real conditions before launch
Outside sandboxes, Article 60 also allows providers of Annex III high-risk systems to test them in real-world conditions, on the basis of a real-world testing plan approved by the market surveillance authority and with the informed consent of the people taking part. For a first view of where a system falls, a free AI risk classifier is available in Romanian.
What participation does — and does not — change
Compliance problems are identified and fixed during testing rather than after investing in a product that turns out not to meet the requirements.
Where the participant follows the sandbox plan and the authority’s guidance in good faith, no administrative fines are imposed for infringements of the AI Act. Other rules continue to apply.
Participants remain liable under applicable EU and national liability law for damage caused to third parties as a result of the testing.
Guidance from the supervisor on how the requirements apply to the specific system — interpretation that is otherwise available only through formal procedures.
Personal data in the sandbox — the overlap with the GDPR
Article 59 allows personal data lawfully collected for other purposes to be further processed in the sandbox, but narrowly. It applies only to AI systems developed to safeguard a substantial public interest — public safety and health, protection of the environment, energy sustainability, transport and mobility, critical infrastructure, the efficiency and quality of public administration and services.
The conditions are cumulative: the data is necessary and cannot be replaced by anonymised or synthetic data; it is processed in a functionally separate, isolated and protected environment; it is not transmitted to third parties; it is deleted when participation ends or the retention period expires; and logs are kept. Outside those cases, training an AI system on personal data needs an ordinary lawful basis under the GDPR.
Timetable and the situation in Romania
In Romania, a government memorandum of 12 March 2026 designated ANCOM as market surveillance authority and single point of contact for the AI Act. The national law that would give the authorities their powers, and organise the sandbox, had not been adopted according to the most recent public information.
Practical recommendation
Do not wait for the sandbox to start preparing. Inventorying AI systems, classifying them under the AI Act and identifying compliance gaps are steps for now; a sandbox is a tool for validating the result, not a substitute for internal preparation. For the systems most likely to benefit, see the EU guidelines on high-risk AI systems.
Frequently asked questions
What is an AI regulatory sandbox?
A controlled framework set up and supervised by a national competent authority, in which providers and prospective providers can develop, train, validate and test innovative AI systems for a limited time before placing them on the market or putting them into service. The legal framework is in Articles 57 to 62 of the AI Act.
Who can take part?
Providers and prospective providers of AI systems, of any size. The Regulation gives SMEs, including start-ups, priority access, and access is free of charge for them, without prejudice to exceptional costs the authority may recover.
When will sandboxes be available in Romania?
Regulation (EU) 2026/1744, the Digital Omnibus on AI, set 2 August 2027 as the date by which each Member State must have at least one national AI regulatory sandbox operational. In Romania, the institutional framework for the AI Act depends on a national law that had not been adopted according to the most recent public information.
Does taking part in a sandbox exempt us from the AI Act?
No. Participants remain liable under applicable liability law for damage caused to third parties during testing. What the Regulation offers is that no administrative fines are imposed for AI Act infringements where the participant follows the sandbox plan and the authority’s guidance in good faith.
Can personal data be processed in a sandbox?
Article 59 allows personal data lawfully collected for other purposes to be further processed in the sandbox only for developing AI systems that safeguard a substantial public interest — public safety and health, the environment, energy, transport, critical infrastructure, public services — and under strict conditions: necessity, an isolated environment, no transfer to third parties, deletion at the end, and logs kept. It is not a general legal basis for training AI on personal data.