WHAT THE ARTICLE IS ABOUT
Safe spaces for AI innovation under regulatory supervision
Article 57 mandates the creation of AI regulatory sandboxes across the EU — controlled environments established and managed by national competent authorities where businesses can develop, train, test and validate AI systems before they go to market. The sandbox concept balances innovation with oversight, giving companies — especially smaller ones — access to regulatory guidance while working on AI systems that may be subject to the Act’s requirements.
WHAT IT SAYS
At least one per member state, prioritising SMEs and startups
- Each member state must establish at least one AI regulatory sandbox at national level by 2 August 2026 — before the Act’s main high-risk AI requirements apply
- The AI Office may establish a Union-level AI regulatory sandbox in addition to national sandboxes
- Sandboxes must be accessible to any provider or prospective provider of an AI system, with priority given to SMEs and startups
- Participation in a sandbox does not exempt participants from their obligations under other applicable EU or national law
- Participants benefit from direct engagement with national competent authorities and can receive guidance on compliance without that guidance being used against them in enforcement proceedings
- Sandbox participants must meet eligibility criteria and sign a sandbox plan setting out objectives, duration, and oversight arrangements
- Information exchanged in sandbox proceedings is subject to confidentiality protections
WHO IS AFFECTED
AI developers and startups — with SMEs given explicit priority
- SMEs and startups developing AI systems — explicitly prioritised for sandbox access
- Any provider or prospective provider of an AI system wanting pre-market regulatory guidance
- National competent authorities responsible for establishing and managing sandboxes
- The AI Office potentially managing a Union-level sandbox
- Research institutions and public bodies developing AI in regulated areas
WHAT IT MEANS FOR SMES
The most directly beneficial provision in the Act for small AI innovators
- If you are building an AI system that may be high-risk and you are concerned about compliance costs and regulatory uncertainty, sandboxes are your entry point — they let you develop under regulatory supervision without full pre-market compliance costs
- The priority access for SMEs and startups is written directly into the article — regulators are required to prioritise you, not treat you equally with large corporations
- Sandbox participation gives you access to regulatory guidance that is specifically designed not to be used against you in enforcement — you can ask hard questions without fear
- Monitor your national competent authority’s sandbox programme — as of August 2026 these must exist in every member state, and several countries have already established them ahead of the deadline
- Spain was an early mover with its national AI sandbox — check the AI Office website for an up-to-date map of available sandboxes across member states
Related Articles
- Article 58 — Sandbox functioning (the detailed rules for how sandboxes operate)
- Article 59 — Personal data processing in sandboxes (special data rules for sandbox participants)
- Article 60 — Real world testing outside sandboxes (the alternative testing pathway)
- Article 62 — Measures for SMEs (Article 57 is part of the broader SME support framework)
