A company that tests an unapproved medicine on patients in Europe does so under an authorised protocol, with an ethics committee and a regulator watching. When the trial ends, the company holds data and a regulator's familiarity with its product. It does not hold a marketing authorisation. The Clinical Trials Regulation (EU) No 536/2014 keeps those two steps apart. The Commission's sandbox regulation of 10 November 2026 keeps a comparable separation for artificial intelligence, and the analogy is ours: the two regimes differ in almost every procedural detail.
Commission Implementing Regulation (EU) 2026/2487, published in the Official Journal on 11 November and in force from 1 December 2026, sets the common rules for every AI regulatory sandbox established under Article 57 of the AI Act. A sandbox is a supervised setting where a provider develops, trains, tests and validates an AI system under a plan agreed with a competent authority, for a limited time, before the system reaches the market. Early coverage, including a detailed account in Tech Times on 12 November, summarised the outcome as no compliance clearance and continuing liability, and also set out the conditional fines protection, the fee rules and the publication rules. This post reads the instrument itself against three questions a company weighing an application will ask: what leaving a sandbox is worth, what protection participation gives, and what it costs.
Practical takeaway. A sandbox exit report counts as recognised documentation in every Member State and should speed conformity assessment; it never replaces it. The AI Act fines shield depends on keeping to the agreed sandbox plan in good faith, which makes the drafting of that plan the most consequential step in the process.
What Implementing Regulation 2026/2487 regulates, and the dates that apply
The regulation has ten articles. They cover access and fees (Article 2), application and selection (Article 3), participation (Article 4), the sandbox plan (Article 5), the written proof and exit report (Article 6), voluntary follow-up (Article 7), extension and suspension (Article 8) and annual reporting (Article 9). Article 1 makes the rules apply to all sandboxes under Article 57, including the Union-level sandbox that the AI Office may now run for systems covered by Article 75(1), the case where an AI system is built on a general-purpose AI model from the same provider. That Union-level option, and the later national deadline, came from the Digital Omnibus on AI, Regulation (EU) 2026/1744, adopted on 8 July and published on 24 July 2026. Article 57(1) now requires each Member State to have at least one national sandbox operational by 2 August 2027; the original date was 2 August 2026.
The timing matters for anyone building a high-risk system. The Omnibus moved the core high-risk obligations to 2 December 2027 for Annex III systems and to 2 August 2028 for AI embedded in products covered by Annex I. A company entering a national sandbox in late 2027 would test against duties that apply weeks later.
What the written proof and exit report under Article 6 are worth in conformity assessment
When participation ends, the supervising authority evaluates the project against the objectives in the sandbox plan. Under Article 57(7) of the AI Act, the authority gives the provider a written proof of the activities successfully carried out, on request, and an exit report. Article 6 of the new regulation sets their minimum content. The written proof lists the activities completed, the legal requirements evaluated and whether the project was finished. The exit report describes the key regulatory issues examined and how they were resolved, any serious incidents, and lessons that can be scaled to similar cases. The authority must deliver the exit report within 45 working days after participation ends.
Article 6(2) and 6(5) then state that neither document has the status or legal effect of a declaration of conformity under Article 47 of the AI Act. Recital 20 adds that leaving a sandbox is no approval to place the system on the market and grants no presumption of conformity. The value lies elsewhere. Article 6(9) obliges market surveillance authorities and notified bodies in every Member State to recognise both documents as documentation of the activities, results and learning outcomes, and recital 20 asks them to take the documents positively into account with a view to accelerating conformity assessment to a reasonable extent. Article 57(7) of the AI Act lets providers use them to demonstrate compliance in that process.
The presumption of conformity still comes from harmonised standards. Quentir's reading of EN 18286 and the quality-management duties of Article 17 showed how a cited standard does that work. A sandbox and a standard therefore answer different questions: the standard tells a provider what conformity looks like, and the sandbox lets a regulator watch the provider try to reach it.
When Article 57(12) of the AI Act removes administrative fines, and which liability remains
Article 57(12) keeps providers liable under Union and national liability law for damage to third parties caused by experimentation in the sandbox. A patient, worker or applicant harmed during a real-world test keeps every route to compensation, which is the right outcome for the people on whom systems are tested. The same paragraph then provides a protection. Where a participant observes the specific plan and the terms of participation and follows the authority's guidance in good faith, no administrative fines shall be imposed for infringements of the AI Act. Where authorities responsible for other Union or national law were actively involved in supervising the system and gave compliance guidance, no fines are imposed under that law either.
The regulation also states, in Article 2(6), that participants remain responsible for complying with applicable Union law. Read together, the two provisions mean that the duty to comply continues while the fines exposure for good-faith work inside the plan is removed. That combination is unusual in EU product law, and it gives a regulator a reason to see difficult cases early.
Who pays under Article 2: free core services for SMEs, full cost for others, an EU representative for non-EU providers
Article 2(3) makes participation free of charge for SMEs, including start-ups, for regulatory guidance, advice, supervision of real-world testing, support for conformity assessment and reliance on the regulatory privileges of Articles 57(12) and 59. Public authorities get the same regime. Authorities may still recover exceptional costs for other services agreed in the plan, provided the participant is told in advance and receives an itemised invoice. Recital 16 ties SME priority access to a registered office or branch in the Union.
Other companies may be charged a fee equal to the full cost of the services, and the authority must publish its fee list when it opens applications (Article 2(4)). A provider with no registered office or branch in the Union must pay a participation fee and designate an EU authorised representative by written mandate, who can be held responsible and liable as a guarantee for the provider's activities in the sandbox (Article 2(5)). Applicants may apply in any Member State, and authorities must answer within three months of a complete application or the call deadline, with a right of appeal (Article 3).
Why the sandbox plan under Article 5 decides how far the protection reaches
The fines shield follows the plan, so the plan's contents carry legal weight. Article 5(2) requires it to state the timeline and estimated end date, the objectives and scope of activities, the roles of the provider and any deployers, the services the authority provides, the AI Act requirements and other laws that will be examined, whether personal data is processed and the data protection authority must be involved, risk-management safeguards with a serious-incident procedure, and any plan for testing in real-world conditions. The plan can change only if both sides agree (Article 5(4)). An extension must be agreed at least 30 calendar days before the planned end date, and an authority may suspend testing or participation when the plan is breached, when risks to health, safety or fundamental rights cannot be mitigated, or at the participant's request (Article 8).
Personal data adds a further boundary. Article 59 of the AI Act allows personal data collected for other purposes to be reused in a sandbox only for systems serving a substantial public interest, such as public health, the environment, energy sustainability, transport and critical infrastructure, or public administration. Every condition in Article 59(1) must be met. The data must be necessary to comply with one or more of the high-risk requirements in Chapter III, Section 2, where anonymised, synthetic or other non-personal data cannot effectively meet them. The data must sit in a functionally separate, isolated and protected environment under the provider's control. Personal data created in the sandbox cannot be shared outside it. The processing may not lead to measures or decisions affecting the people concerned, and the data must be deleted once participation ends or its retention period expires. Recital 25 states that the European Data Protection Supervisor gave its opinion on the draft on 6 March 2026.
How Quentir Reads It
The regulation makes the sandbox plan the main document of a project, both for the fines protection and for later publication. An exit report reaches the Commission's single information platform only if both the participant and the authority explicitly agree (Article 6(7)). Without that consent, Article 6(8) still lets the authority publish key findings and lessons containing only non-confidential information, and it presumes information disclosable unless it was marked confidential in the plan or by the participant before the project ends, and an authority may publish key findings without the participant's consent. A provider therefore decides at the drafting stage which duties it wants examined, which data is confidential, and which lessons it is prepared to see published. Providers of systems that generate content may list the transparency duties of Article 50 among the obligations examined. Signature Brief 2026.3, AI Act Article 50 Disclosure Readiness, covers what Article 50 requires to be disclosed, the deepfake and public-interest-text triggers and the assisted-versus-generated distinction, with an internal-use license for the whole organization.
Common selection criteria in Article 3 and the Union-wide recognition rule in Article 6(9) may reduce the incentive to apply to the most lenient national authority, since applicants can choose any Member State. Whether they do, and how much weight market surveillance authorities and notified bodies give exit reports in practice, will show only after the first cohorts finish. Authorities that charge companies other than SMEs must publish their fee lists when they open applications, and those lists will show whether larger companies and non-EU providers find the price worth paying.
Sources: European Commission, Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 laying down rules for the application of Regulation (EU) 2024/1689 as regards AI regulatory sandboxes, OJ L, 11 November 2026 (recitals 16, 20 and 25; Articles 1 to 10). Article 57 and Article 59 of Regulation (EU) 2024/1689 (AI Act), as amended, via the Commission's AI Act Service Desk. European Parliament and Council, Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), OJ L, 24 July 2026. Clark Johnson, "EU AI Regulatory Sandbox Rules Start December 1: Tests Earn No Compliance Clearance, Liability Stays", Tech Times, 12 November 2026. Regulation (EU) No 536/2014 on clinical trials on medicinal products for human use. Sources checked 13 November 2026.
Published intelligence, built to inform your own decisions. Published: November 13, 2026.