regulatory sandbox
A time-limited test environment operating under regulatory supervision, where companies can trial new technologies under real-world conditions before launching them to the market.
In plain terms
Think of a driving school: you drive a real car, but in a closed area with an instructor in the passenger seat. If you make a mistake, you learn from it without causing an accident in traffic. A regulatory sandbox works the same way for new technologies, such as artificial intelligence.
The company operates under regulatory supervision, testing its development in a limited scope before a broad market launch. It learns whether it complies with rules and avoids discovering errors later in the form of fines. Meanwhile, the authority observes the new technology in real operation to inform future regulation.
What distinguishes it from simple internal testing is that the authority officially participates: it takes place according to a pre-agreed plan, with deadlines and protective measures. It is not a loophole; the rules still apply here. However, the company receives guidance, and good-faith errors are typically not fined if the agreed framework is followed.
How it works
- Application: the company presents to the authority what it wants to try, why it is novel, and what risks it sees.
- Test plan: the authority and the company jointly define the experiment's duration, the number of people affected, the data used, and the safeguards protecting participants.
- Supervised trial: the system operates in a limited scope while the authority continuously monitors, asks questions, and provides advice.
- Intervention if needed: if a serious risk is identified, the test is modified or halted.
- Exit: at the end, the company receives a written summary of the experiences. It can use this to prove compliance when it actually enters the market.
An everyday example
A Hungarian startup is developing an AI agent that prepares loan applications for bank customers. Before the system makes recommendations on the finances of thousands of people, the company tests it with a few hundred volunteer customers under regulatory supervision. If you participate in such a test, you are informed in advance and asked for your consent.
Who is affected
A common misconception
Limits and risks
- A successful sandbox test does not grant automatic permission or a quality seal; the standard compliance procedure must still be completed to enter the market.
- The number of spots and the capacity of authorities are limited, so not every applicant gets in, and the process can take months.
- A small-scale, short-term trial does not reveal every bug that may emerge during mass usage.
- What the sandbox offers may vary by country, so what is sufficient in one member state may not be in another.
Similar, but not the same
- sandbox in computing – This is technical isolation that prevents a program or agent from harming the machine; the regulatory sandbox is a legal and supervisory framework.
- testing under real-world conditions – Under the EU regulation, this can be conducted outside the sandbox under certain conditions and based on a plan approved by the authority, without the same level of close regulatory oversight.
- conformity assessment – This is the mandatory pre-market test for high-risk AI systems, while the sandbox is voluntary preparation that facilitates but does not replace it.
Related concepts
AI regulation, innovation policy, experimental regulation
AI regulatory sandbox, fintech sandbox, testing under real-world conditions
EU MI-rendelet, High-risk AI system, conformity assessment, market surveillance authority, risk assessment
A brief history
The first known regulatory sandbox was launched in 2016 by the UK's Financial Conduct Authority (FCA) for fintech companies, and the model has since been adopted by many countries. In Hungary, the National Bank of Hungary (Magyar Nemzeti Bank) has run one for financial innovation since 2018. The EU's Artificial Intelligence Act, adopted in 2024, also requires every member state to set up at least one in the field of AI, either alone or jointly with other member states.
Where things stand
It is a globally established tool in the financial sector, but is currently being rolled out in the AI field. The EU regulation originally set August 2, 2026, as the deadline for launching national AI sandboxes; the regulation's 2026 simplification amendment (digital omnibus) extended this to August 2027. Member states are progressing at different paces, and actual experience remains limited.
In the news
- Utah lets AI prescribe acne medication without direct physician oversightOct 5, 2026 · Health
- Manchester NHS and MHRA launch AI testing sandbox; MHRA commission publishes findingsSep 2, 2026 · Health
- Israel's regulatory sandbox tests autonomous AI in clinical careAug 24, 2026 · Health
- Poland to build AI regulatory sandbox with 35 million zlotyAug 11, 2026 · Jog
- Bermuda, BIS, Slovenia and Poland weigh in on AI rules for financial servicesAug 10, 2026 · Money
- London launches AI health regulatory sandbox for NHS innovation, MHRA seeks applicantsAug 3, 2026 · Health
Also known as
regulatory sandbox · regulatory sandbox · AI regulatory sandbox · regulatory sandbox · regulatory sandbox · AI regulatory sandbox · AI sandbox · innovation sandbox
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.