How can smaller firms use AI regulatory sandboxes?

What is an AI regulatory sandbox?

An AI regulatory sandbox is defined in Article 3 as a controlled framework set up by a competent authority which offers providers or prospective providers the possibility to develop, train, validate and test, where appropriate in real world conditions, an innovative AI system, pursuant to a sandbox plan for a limited time under regulatory supervision. The sandbox is therefore a supervised testing environment, not a market authorisation and not a certification. The competent authority provides guidance, supervision and, where relevant, involvement of other authorities including data protection authorities. The provider remains responsible for the system and for any damage caused, subject to the specific provisions on administrative fines in Article 57, which must be read in full. The sandbox plan sets out the system, the testing, the risks, the safeguards, the duration and the exit arrangements. A sandbox is established by the competent authority, and a provider cannot declare its own testing to be a sandbox. The distinction matters because testing outside a sandbox, under Article 60, has different conditions, different approval and different safeguards, and is dealt with in the codes and sandboxes cluster of this site.

What priority access do SMEs have?

Article 62(1)(a) requires Member States to provide SMEs including startups having a registered office or a branch in the Union with priority access to AI regulatory sandboxes, to the extent that they fulfil the eligibility conditions and selection criteria. Priority access is not a guarantee of a place and does not preclude other SMEs that also fulfil the criteria. The eligibility conditions and selection criteria are set by the competent authority under Article 58, taking account of the detailed arrangements in that Article, including the criteria for selection, the application procedure, the duration, the supervision and the exit arrangements. A small firm that wishes to use a sandbox should therefore check the sandbox notice in its Member State, the eligibility criteria, the application form, the timetable and the information required, and should prepare an application that addresses the innovation, the risk, the safeguards and the learning objectives. An application that describes only the commercial opportunity, without the risk and safeguard analysis, is unlikely to meet the criteria. The application should be prepared with the same care as a conformity file, because the sandbox authority will test the reasoning.

For the broader regulatory framework, see SMEs, startups and open source overview.

What does a sandbox plan contain?

A sandbox plan, under Articles 57 and 58, sets out the system to be tested, the intended purpose, the testing in real world conditions where applicable, the duration, the participants, the data, the risk assessment, the mitigation measures, the supervision arrangements, the monitoring and reporting, and the exit arrangements. The plan is agreed with the competent authority before testing begins, and testing must stay within the plan. Changes to the plan, to the system or to the testing conditions should be agreed in advance where the Article requires it, and should be recorded where they are made. The plan should state what will be learned from the sandbox, how the learning will be used, and what evidence will be produced for the exit report. It should also state the safeguards for affected persons, including information, consent where required for real world testing, and the arrangements for stopping the test if a risk materialises. A plan that is vague about risks, safeguards or exit arrangements will not give the authority the basis to supervise the test, and will not give the provider the basis to show that the test was conducted properly. The plan should be treated as a controlled document, with versions and approvals.

What happens at the end of a sandbox?

At the end of sandbox testing, the competent authority drafts an exit report, under Article 57, describing the testing, the outcomes, the compliance lessons and any further action. The provider should contribute the evidence and the analysis for the exit report, including test results, incidents, failures and the action taken. The exit report does not certify the system, does not grant a CE mark, and does not authorise placing on the market. It provides a supervised learning record that the provider can use in its conformity assessment, its risk management and its post market planning, and that the authority can use to improve supervision and guidance. A provider that treats an exit report as a market authorisation will misunderstand the sandbox. A provider that treats the sandbox as a way to avoid conformity assessment will also misunderstand it. The correct use of a sandbox is to test, to learn and to produce evidence under supervision, and then to use that evidence in the ordinary conformity and placing on the market process, with the exit report cited as part of the file. The annual reporting duties of competent authorities on sandboxes, under Article 57, should also be noted, because sandbox activity is reported and can inform future guidance.

What about personal data in a sandbox?

Article 59 provides for further processing of personal data for developing certain AI systems in the public interest in a regulatory sandbox, subject to safeguards and to Union data protection law. The provision is narrow and conditional, and it does not create a general permission to process personal data in a sandbox. Personal data processing in a sandbox must have a lawful basis under data protection law, must meet the purpose, minimisation, security and rights requirements, and must involve the data protection authority where Article 57 requires it. A sandbox application that involves personal data should include a data protection assessment, the data categories, the sources, the retention, the access controls and the arrangements for data subject rights. Sensitive data, data about children, and data about vulnerable persons need particular care. A small firm should involve its data protection adviser early, and should not assume that sandbox supervision replaces data protection compliance. The sandbox plan and the data protection assessment should be consistent, with the same data categories and the same safeguards described in both documents.

Continue with codes, standards and sandboxes overview.

What does sandbox participation not do?

Sandbox participation does not decide that a system is lawful, does not decide that it is high risk or not high risk, does not replace a conformity assessment, does not grant a CE mark, does not remove liability for damage, and does not authorise placing on the market or putting into service. It does not suspend the prohibited practices in Article 5, and it does not remove data protection duties. A provider that participates in a sandbox must still meet the duties that apply to its system outside the sandbox, subject to the specific fine provision in Article 57 which must be read in full and should not be overstated. A provider should therefore plan the sandbox as one stage in a longer process, with the conformity, registration and market stages planned separately. This guide does not state whether a sandbox is available in any Member State on a given date, and the position must be checked with the national competent authority. The date wording for sandbox operation, that Member States should have at least one sandbox operational by 2 August 2027, should be kept in its Service Desk form and not upgraded to a hard obligation without checking the consolidated Article 57 text.

Frequently asked questions

Continue through this AI Act cluster

• SMEs, startups and open source overview
• codes, standards and sandboxes overview
• regulatory sandboxes in Articles 57 to 59
• real world testing in Article 60
• proportionate duties and support
• providers