Regulatory sandbox (AI Act)

What is a regulatory sandbox under the AI Act?

A regulatory sandbox under the AI Act is an arrangement with a supervisor, not a piece of software. A national authority lets you develop and test an AI system for a set period under its guidance. You apply, you agree a sandbox plan setting the objectives, conditions, timeframe and method, you build and test while the authority watches, and you leave with a written record. Your engineers work on the real system. What gets fenced off is the regulatory risk.

This is worth saying up front, because engineers use the same word for something else. An agent sandbox is a container or a virtual machine that stops an AI agent from touching files, networks or credentials it should not reach. Same word, different field. This entry is about the legal instrument in Chapter VI of the AI Act, Articles 57 to 63.

What the AI Act asks member states to do

Article 57(1) tells every member state to have at least one AI regulatory sandbox running at national level. The original deadline was 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since July 2026, moved it to 2 August 2027. Countries may add regional sandboxes or set one up jointly with another member state.

Article 58 gives the Commission the job of writing implementing rules, so a sandbox in Ghent and one in Lisbon run on the same principles for eligibility, application, exit and reporting. It also makes access free of charge for SMEs including start-ups, apart from exceptional costs an authority may recover fairly, and Article 62 adds priority access for SMEs and start-ups based in the Union.

What participating actually gets you

Guidance and supervision. Under Article 57(6) the authority helps you identify risks to health, safety and fundamental rights and test whether your mitigations hold. The value is timing: you hear what your supervisor thinks while you can still change the design.

An exit report. Article 57(7) gives you, on request, written proof of what you completed plus an exit report on the results. Market surveillance authorities and notified bodies have to take that report positively into account with a view to accelerating conformity assessment to a reasonable extent. It speeds a procedure up, it does not replace one. There is no sandbox CE mark.

Cover on fines, not on liability. Article 57(12) says that if you stick to the plan and follow the authority's guidance in good faith, no administrative fines are imposed for infringements of the regulation. Harm someone during the experiment and you still answer for the damage under ordinary law, and the authority keeps its supervisory and corrective powers.

A narrow route for personal data. Article 59 allows further processing of data collected for another purpose, but only for AI serving a listed public interest such as public health, energy or transport, and only in an isolated environment nobody exports data from. Outside that list, the GDPR basis you had is the basis you still have.

Testing in real world conditions is a separate mechanism

Article 60 covers what people most often mix up with the sandbox: testing a high-risk system in real world conditions, outside any sandbox. A provider or prospective provider may do it alone or with a deployer. You draw up a real-world testing plan covering the objectives, the method and the geographical, population and temporal scope, submit it to the market surveillance authority and register the test in the EU database. It can run six months at most, extendable once by another six.

The safeguards sit around the people in the test. Article 61 requires informed consent, anyone can withdraw at any moment without a reason and ask for their data to be deleted, qualified staff oversee the test, and serious incidents go to the authority under Article 73.

To keep the two apart, ask who supervises and where the system runs. In a sandbox the authority supervises directly and the work sits inside an arrangement you negotiated with it, and Article 57(5) even allows real-world testing inside a sandbox. Outside one you supervise your own test against a plan the authority approved, and the system runs in the real setting on real people. The first puts a regulator in the room. The second gives you evidence from the field, with the burden of proof on you.

Where Belgium stands in September 2026

There is nothing to apply to yet. Belgium has no operational national AI regulatory sandbox, and after the Digital Omnibus the country has until 2 August 2027 to open one. The federal government agreement of 31 January 2025 named BIPT as the central AI Act authority, but the law giving it those powers has not passed, and in July 2026 FOD Economie confirmed to VRT NWS that the reading of a legal vacuum is correct. The Flemish government published a reflection paper on what a sandbox in Flanders could look like. Brussels has FARI, the test and experience centre for AI run by the VUB and the ULB, which is a place to experiment but not a sandbox in the sense of Article 57.

What we could not confirm: which authority will host the Belgian sandbox, when it opens, and how the federal and regional levels divide it. If a vendor says they can get you into the Belgian sandbox today, ask for the name of the authority and the application form.

Who this is for, and where everyone else should go

The kind of project that belongs in one: a Belgian HR software company of thirty people adds a ranking model to its recruitment product, so the tool sorts candidates for a vacancy. That makes it the provider of a high-risk system under Annex III, with a conformity assessment, technical documentation, risk management and human oversight to get right before those rules apply on 2 December 2027. Nobody there has done a conformity assessment before, and the two questions they cannot settle alone are whether the bias testing is good enough and whether the logging meets the standard.

Most companies are nowhere near that. You bought an assistant, you run a chatbot, your accounting package grew an AI feature: you are a deployer, and a sandbox is not the tool for your question. The Commission opened the AI Act Service Desk and Single Information Platform on 8 October 2025, where a compliance checker tells you whether the obligations apply and a form puts your question to a team working with the AI Office. Article 62 also tells member states to set up communication channels for smaller companies, the door to knock on once the Belgian supervisor is in place.