A sandbox run under regulatory supervision produces a lot of paper. Test logs, monitoring output, authority feedback, and a plan that was agreed at the start. Teams will be tempted to treat the pile as the answer to one question: can we launch? The recitals of the new Commission rules on AI regulatory sandboxes describe that work as support before placement on the market. They do not, in the passages relied on here, assign the launch decision. This column’s position is that the provider should name an owner for that decision and record it.
Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 lays down rules for the establishment, development, implementation, operation, and supervision of AI regulatory sandboxes under the AI Act. The Official Journal text is here.
What the text says a sandbox is for
Its first recital describes a sandbox as a controlled framework for the development, training, validation, and testing of innovative AI systems, including in real-world conditions, under a sandbox plan and regulatory supervision, for a limited time before their placement on the market. Member States should have at least one sandbox at national level. The recitals say these sandboxes should be established in the Member States by 2 August 2027.
Recital 4 lists services that typically include regulatory guidance, advice, and supervision of AI system development, testing, and validation, and support for conformity assessment. That is a long list of help, and every item on it supports the participant. The quoted text describes support for conformity assessment. It does not treat that support as the participant’s launch decision.
Recital 19 says the sandbox plan should address the participants’ objectives, the relevant risk management safeguards, and the support the competent authorities provide. Recital 8 adds that the terms and conditions of participation should be agreed before entering the sandbox, and that these can be accepted with an electronic signature. Recital 13 provides for mutual recognition of the results. The plan, the signed terms, and the results are documents the regulation describes. The provider’s launch decision is a separate act this column recommends recording. The recitals cited here do not set its timing.
The launch decision is a separate act
Supervision that supports conformity assessment is a different job from the conformity assessment itself. Recital 1 puts the sandbox before placement on the market. Between the last day in the sandbox and the first day on the market, the company that takes the system to market still has a decision to make. The recitals relied on here support that sequence. They do not by themselves settle every legal question about exit documents or conformity procedures.
The decision sits alongside other exposure. Europe Withdrew Its AI Liability Directive, and the Exposure It Left Behind Is Larger covers how the surviving Product Liability Directive reaches software and AI. This column does not treat a sandbox file as a substitute for that analysis.
Responsible AI checks can be automated. AI Governance owns the launch.
Recital 4’s list includes human regulatory guidance, advice, and supervision, as well as support for conformity assessment. Test logs, monitoring, and conformity tooling may include automated checks. When those checks are machine checking machine, with no human review at the point of the check, they are Responsible AI. They are valuable, and a sandbox will produce plenty of records.
AI Governance asks a different question. Who inside the provider reads what the sandbox produced, and decides to accept, modify, or reject the launch? Under Checkpoint-Based Governance that person holds binding authority and can show the decision afterward. A pile of output does not show it. As Narrative Is Not Custody: Why AI Compliance Needs Traversable Rules argues, a compliance narrative is not a record someone can traverse and check.
What to do before the system leaves the sandbox
- Name one launch owner and one deputy. They sign the post-sandbox decision.
- List every finding, recommendation, and incident the sandbox produced. Give each an entry of accept, modify, or reject, with the reasons.
- Record what changed between the system that was tested and the system placed on the market.
- Where the system is high-risk, tie the launch signature to the conformity declaration the system needs.
- Keep the record where an auditor can open it without asking the people who wrote it.
Three measures show whether this works:
- Share of sandbox findings with a recorded disposition before launch. The target is every finding.
- Days from receiving the final sandbox documents to the signed launch decision. Track the trend.
- Number of differences between the sandboxed and the launched system that carry a signed review. The target is all of them.
The recitals describe sandbox supervision as support before placement on the market. The launch remains a separate, owned, recorded act: one launch owner, a disposition for every sandbox finding, and a record an auditor can open. This column is general information, not legal advice.
Sources
- European Commission. (2026, November 11). Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 laying down rules for the application of Regulation (EU) 2024/1689 of the European Parliament and of the Council as regards the establishment, development, implementation, operation and supervision of AI regulatory sandboxes. Official Journal of the European Union, L series. https://publications.europa.eu/resource/cellar/e501019f-dd70-11f1-8501-01aa75ed71a1.0006.03/DOC_1
- Puglisi, B. C. (2026, July 5). Europe withdrew its AI liability directive, and the exposure it left behind is larger. https://basilpuglisi.com/eu-ai-liability-product-directive/
- Puglisi, B. C. (2026, October 11). Narrative is not custody: Why AI compliance needs traversable rules. https://basilpuglisi.com/ai-compliance-traversable-rules/
FAQ
What is an AI regulatory sandbox under the new Commission rules?
The first recital describes a controlled framework for developing, training, validating, and testing innovative AI systems, including in real-world conditions, under a sandbox plan and regulatory supervision, for a limited time before placement on the market. Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 lays down the common rules for those sandboxes.
What does the authority do in a sandbox?
Recital 4 lists services that typically include regulatory guidance, advice, and supervision of AI system development, testing, and validation, and support for conformity assessment.
What belongs in a sandbox plan?
Recital 19 says the plan should address the participants’ objectives, the relevant risk management safeguards, and the support the competent authorities provide.
When should Member States have a sandbox?
Member States should have at least one sandbox at national level. The recitals say these sandboxes should be established in the Member States by 2 August 2027.
Does supervision in a sandbox complete the conformity assessment?
No, not on the recitals relied on here. Recital 4 describes support for conformity assessment. That support is not the same act as the assessment, and it is not the participant’s launch decision.
Who should decide the launch after a sandbox, and how would you know it worked?
The company that takes the system to market should name one launch owner and a deputy, record accept, modify, or reject for every sandbox finding, and keep that record where an auditor can open it. The test is whether every finding has a recorded disposition before launch, and whether differences between the tested system and the launched system also carry a signed review. If those measures stay empty, the sandbox file is still only a pile of paper.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply