AI regulatory sandbox

For a smaller provider, a sandbox converts regulatory uncertainty into a documented supervisory relationship: you leave with a written exit report describing what was tested and how, which is usable evidence in a later conformity assessment and a genuine credibility asset in enterprise procurement. The mitigating effect on fines matters too — good-faith participation and adherence to the sandbox plan is something authorities take into account when they exercise discretion over penalties.

A five-person startup building an AI tool for assessing training-programme outcomes is unsure whether it falls into Annex III education. Rather than guess, it applies to its national sandbox, agrees a plan with the authority, tests the system under supervision, and leaves with an exit report and a documented view on classification — an answer it could not have obtained from a law firm with the same authority behind it.

Treating a sandbox as a compliance holiday. Participation does not suspend the Act or shield participants from liability for harm caused; it provides supervision, guidance and mitigating weight on fines, while the underlying obligations continue to apply.

When must AI regulatory sandboxes be available?

Member States must ensure at least one sandbox is operational at national level by 2 August 2026, established alone, jointly with other member states, or by participating in an existing sandbox.

Who gets priority access to a sandbox?

SMEs including start-ups get priority access, provided they meet the eligibility conditions. Access is free of charge for them, without prejudice to exceptional costs authorities may recover.

Does sandbox participation protect against fines?

Not automatically. Participants remain liable under applicable law, but where they follow the sandbox plan and act in good faith, supervisory authorities take that into account when deciding whether and how to impose administrative fines.