The European Commission has set common rules for AI regulatory sandboxes: supervised schemes in which companies can develop and test AI systems, including in real-world conditions, before putting them on the EU market. Implementing Regulation (EU) 2026/2487 was published in the EU’s Official Journal on 11 November and, according to reports of its text, applies from 1 December.
Each EU country must have at least one working sandbox by 2 August 2027. Companies can also apply in a member state other than their own, and should get a decision within three months of a complete application. Start-ups and small firms get priority and free core supervision; larger firms can be charged.
The limits matter as much as the access. Taking part is not approval: the exit report a company receives, due within 45 working days according to the same reports, is not a declaration of conformity, though regulators must take it into account. Firms stay liable for any harm to others, and are shielded from AI Act fines only if they follow the agreed plan and the authority’s guidance in good faith.
Sandboxes are the AI Act’s main tool for helping smaller developers comply. Common rules should mean similar treatment across the bloc, while anyone harmed by a system under test can still hold its maker liable.
Source: Official Journal of the EU, Implementing Regulation (EU) 2026/2487