In short
An AI regulatory sandbox (in German KI-Reallabor) is a controlled test environment operated by an authority, in which companies can develop, train, test and validate AI systems under supervision before they come onto the market. The legal basis is Article 57 of the AI Act, Regulation (EU) 2024/1689.
The core of the offer: anyone who sticks to the plan agreed with the authority gets legal certainty, and is not fined for infringements of the regulation that occur in the sandbox.
That sounds like a tool for AI corporations. It is expressly intended otherwise: Article 57 grants small and medium-sized enterprises and start-ups priority access. At the moment the bottleneck is not legal but practical: in most Member States, there is simply no operational sandbox yet.
What Article 57 of the AI Act prescribes
The obligation falls on the Member States, not on companies. Each Member State must have at least one AI regulatory sandbox operational at national level. The original deadline in the text of the regulation was 2 August 2026; the Digital Omnibus on AI has postponed it to 2 August 2027. The sandboxes are set up by the competent national authorities. Article 57(1) expressly permits two variants: joint establishment with other Member States and participation in an existing sandbox, provided national coverage remains equivalent. In addition, the AI Office can set up a sandbox at Union level.
The purpose is set out in paragraph 5: a controlled environment that fosters innovation and supports the development, training, testing and validation of AI systems before they are placed on the market. Paragraph 9 names the objectives more concretely: legal certainty for those involved, exchange of best practices, promotion of innovation, regulatory learning on the part of the authorities and accelerated market access, "in particular when provided by SMEs".
Protection from fines and its limits
The economically most interesting paragraph is paragraph 12. It makes two things clear. First: civil liability remains unaffected. Anyone who causes damage in the sandbox is liable under applicable law as they would be otherwise. Second: as long as the prospective providers respect the specific plan and follow the authority's guidance, the authorities will not impose fines for infringements of the AI Act.
This is not a blank cheque but narrowly tailored protection. It applies to the AI Act, not to the GDPR, not to product safety law, not to competition law. And it only applies within the agreed plan. Anyone who does something other than what was agreed in the sandbox is in the same position as any other company.
The exit report as proof of conformity
The second practical value lies in paragraph 7: on completion, the authority produces an exit report setting out in detail the activities carried out, the results and the lessons learned. These documents can be used to demonstrate conformity; they are therefore not just documentation for the file but usable material in the later conformity assessment procedure. For a high-risk AI system with considerable evidence obligations, this is a real advantage.
Paragraph 16 also obliges the authorities to publish annual reports starting one year after the sandbox has been established. These reports are interesting for companies because they show which questions were actually dealt with, and thus where supervision is setting its priorities.
How a sandbox works in practice
Article 58 of the AI Act governs the details of implementation. The typical process can be summarised in five steps:
- Application and admission. The company describes the AI system, the planned use case and the open regulatory questions. The authority checks whether the case is suitable for a sandbox.
- The specific plan. The authority and the company jointly determine what will be tested, with which data, under which safeguards and with which termination criteria. This plan is the basis for the protection from fines under paragraph 12.
- Operation under supervision. The tests run with ongoing supervision. If risks to health, safety or fundamental rights arise, the supervisory authority can intervene or end participation.
- Involvement of other authorities. Under paragraph 10, data protection authorities and other competent bodies must be involved to the full extent of their tasks and powers. Anyone processing personal data in the sandbox therefore has the data protection supervisory authority at the table, which is intentional and in practice the greatest added value.
- Completion and report. The result is the exit report under paragraph 7.
An important special case is set out in Article 59: under strict conditions, personal data originally collected for other purposes may be further processed in the sandbox for the development of certain AI systems in the public interest. This is one of the few places where the AI Act enables something in data protection terms rather than requiring something, and it is correspondingly narrowly defined.
Distinction: sandbox, testing in real-world conditions, pilot project
| Format | Legal basis | Supervision | Typical question |
|---|---|---|---|
| AI regulatory sandbox | Art. 57–59 AI Act | The authority provides continuous supervision | "Is our system compliant the way we are building it?" |
| Testing in real-world conditions | Art. 60 AI Act | Notification and protection obligations, no continuous supervision | "Does the system work outside the lab?" |
| Internal pilot project | No special basis | None | "Does the technology do anything for us at all?" |
Confusing the sandbox with testing in real-world conditions is the most common mistake. Article 60 permits testing outside regulatory sandboxes and makes it subject to its own conditions, including a testing plan and information obligations towards the persons concerned. The protection from fines in Article 57(12) does not apply there.
Real-world example: the situation in Germany
Since the end of July 2026, Germany has had a competent supervisory structure with the AI Market Surveillance and Innovation Promotion Act; market surveillance lies with the Federal Network Agency (Bundesnetzagentur). However, this does not yet mean there is a regular, permanently open AI regulatory sandbox under Article 57. What there is, is preparatory work: a completed pilot project and participation in a European initiative. This is not a German peculiarity: implementation is proceeding slowly across Europe, which is why the deadline was postponed in the Digital Omnibus in the first place.
For a company planning an AI product today, this means: the sandbox is an option for 2027 and beyond, not for the current project. Anyone who needs clarity now will get it from the Federal Network Agency's guidance and deadline overview, the Commission's guidelines and a sound risk classification of their own, not from a sandbox that is not yet open. The AI Act Explorer's overview of Member State sandbox approaches tracks how the Member States are tackling implementation; the wording of Article 57 of the AI Act (German version) is freely available.
When a sandbox is worthwhile for mid-sized companies
Not every AI project belongs in a sandbox. Participation takes time, requires documentation and ties up people who would otherwise be working on the product. It typically makes sense when three things come together:
- The classification is unclear. The system is on the borderline of the high-risk area, for example in human resources, in creditworthiness assessment or in controlling access to education.
- Money depends on it. A later misclassification would devalue a product development, not just change an internal process.
- Data protection is part of the problem. If the training data is the real question, the involvement of the data protection supervisory authority under paragraph 10 is the main prize.
Anyone using a standard model for internal text work, on the other hand, does not need a sandbox but a usage policy, training and a look at the transparency obligations.
And there is a fourth, often overlooked reason: anyone selling as a provider into a regulated market will be asked by their customers for evidence long before an authority asks. An exit report from a sandbox is a document that carries weight in the purchasing process, more robust than any white paper and considerably cheaper than a lost place in a tender. This aspect should be included in the cost-benefit calculation, instead of regarding the sandbox merely as compliance effort.
Why the EU offers regulatory sandboxes at all
The AI Act is a product safety law. It describes obligations for systems, some of which did not even exist when the legislation was drafted, and it has to be applied to technology that changes faster than any set of rules. This tension creates a familiar problem: companies do not invest because they do not know how the rules will be interpreted, and authorities do not interpret because they lack cases.
Regulatory sandboxes are an attempt to break this circle, in both directions. Paragraph 9 expressly names "regulatory learning" as a goal. The supervisory authority sees real systems, real data flows and real conflicts of objectives before it writes guidelines. Companies learn how an authority assesses their classification before they invest two years of development in an assumption. Paragraph 14 ensures that this knowledge does not seep away in a single authority: the national bodies, the European Data Protection Supervisor and the AI Office coordinate within the AI Board.
The fact that this mechanism was not ready by the date originally set is the most honest description of the state of implementation of the regulation as a whole: the obligations for companies are starting on schedule, while the accompanying infrastructure lags behind. For planning purposes, this means not relying on relief that does not yet exist. Anyone planning a high-risk system should build up the documentation now; a sandbox can confirm it later, but it does not replace it.
Frequently asked questions about AI regulatory sandboxes
By when does Germany have to provide an AI regulatory sandbox?
Under the amended Article 57(1) of the AI Act, by 2 August 2027. The regulation originally provided for 2 August 2026; the Digital Omnibus on AI postponed the deadline by one year.
Does participation cost anything?
Article 57 does not regulate fees, but it obliges the Member States to provide sufficient resources. How the national sandboxes handle this will be decided with each implementation; reliable statements will only be possible once a sandbox has actually opened.
Does a sandbox protect against every penalty?
No. The protection under paragraph 12 relates to fines for infringements of the AI Act and only applies as long as the plan and the authority's guidance are followed. Civil liability and obligations from other areas of law, in particular data protection law, remain.
Is a sandbox the same as testing in real-world conditions?
No. Testing in real-world conditions is governed by Article 60 and expressly takes place outside regulatory sandboxes. It has its own requirements and does not offer the protection from fines under Article 57.
Do SMEs have an advantage?
Yes. Article 57 provides for priority access for SMEs and start-ups, and Article 62 obliges the Member States to take supporting measures for smaller providers. The political will is clear; the problem is availability, not priority.