ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI compliance
EU AI Regulatory Sandbox Rules Take Effect on 1 December: What You Can Test Inside — and What You Leave With

EU AI Regulatory Sandbox Rules Take Effect on 1 December: What You Can Test Inside — and What You Leave With

AI compliance • Admin • • 13 views

The rules for the EU's AI regulatory sandboxes are set: the European Commission adopted Implementing Regulation (EU) 2026/2487 on 10 November 2026, published it in the Official Journal on 11 November, and it enters into force on 1 December 2026. The regulation implements Article 57 of the AI Act and harmonises the rules for sandboxes across the Union: companies can develop and test innovative AI systems before placing them on the market, under supervision and including testing in real-world conditions. Companies offering AI systems in the EU are first in line — start-ups above all.

Who can apply, and what it costs

Applications go to the competent authority of any Member State, with no restriction to the country of registration. Only three kinds of systems qualify: those not yet placed on the market or put into service; those already on the market but due for substantial modification; and those not high-risk by intended purpose but which become high-risk through the way they are used once on the market. Providers and prospective providers can apply, alone or together with deployers and other third parties. Authorities select participants against five criteria: the regulatory challenge, learning potential, public interest, maturity of the proposal, and innovativeness. Once a complete application is submitted, the outcome must be communicated within three months, and decisions can be appealed.

Fees are the most concrete part: regulatory guidance, advice and supervision are free for SMEs and start-ups; only exceptional costs beyond the usual services are charged, and they must be flagged in advance with an itemised bill. Other companies pay the actual costs incurred, and the fee list must be published when applications open. Non-EU providers with no registered office or branch in the Union must pay a fee and, in addition, designate in writing an authorised representative established in the EU, who is liable for their activities in the sandbox. Where personal data is processed, the competent data protection authority must be involved in supervision.

What gets tested inside, and what you leave with

After selection, the company and the authority agree a sandbox plan setting out the timeline, objectives, the legal requirements under supervision, risk management and the procedure for reporting serious incidents; testing in real-world conditions is written in too, and the plan can be updated if both sides agree. When participation ends, the authority issues two documents: written proof listing the development and testing completed, and an exit report explaining how the regulatory questions were resolved, which practices can be replicated, and any serious incidents — the report must reach the company within 45 working days. Both documents are recognised across the Union, and notified bodies and market surveillance authorities must take them into account in conformity assessment.

A sandbox is not a free pass

The easiest mistake is to treat entry into a sandbox as a licence to place a product on the market. The regulation is blunt: participation and exit are neither an authorisation to place on the market nor a presumption of conformity, and a conformity assessment is still required after leaving the sandbox; liability for harm to third parties remains, as does the duty to comply with EU law. There is only one route to relief from fines: participants who observe the sandbox plan and follow the authority's guidance in good faith are exempt from administrative fines under the AI Act. If the plan is breached, or if significant risks to health, safety or fundamental rights emerge that cannot be mitigated, the authority can suspend testing or participation and must inform the AI Office.

The regulation takes effect on 1 December, but national sandboxes need not open that day: Member States must have at least one in place by 2 August 2027, and the AI Office may set up one at Union level too. Companies can start now by checking their systems against the three eligibility categories and preparing the elements of a sandbox plan, above all personal-data processing and serious-incident reporting. There is one more date in the same batch: EU AI Act Adds Two New Bans on 2 December: These AI-Generated Systems Face a Direct Sales Ban, when the two bans and the marking transition period expire on the same day.

Recommended Tools

More