Angst vor dem Risiko ist eine Bremse: Löse sie und fang einfach an | Umair Zaffar
Umair Zaffar November 15, 2026 0 Comments

EU AI Act Article 57: AI Regulatory Sandboxes, Implemented in Practice

EU AI Act Article 57 creates AI regulatory sandboxes: an authority accompanies the development of a system under an agreed plan, with an exit report and without fines as long as the plan is followed in good faith. We keep plan and report as registry records and run the same governance layer in the sandbox as in production, so the evidence carries over.

Many treat the sandbox as a topic for start-ups. That is too narrow. If you build a high-risk system yourself, this is where you get an authority as a counterpart before the obligations bite. Article 57 of the AI Regulation is addressed to the Member States and their authorities (paragraphs 1 to 4) and to the providers and prospective providers that take part. Article 58 hands the Commission the implementing acts, Article 59 governs when personal data may be further processed in the sandbox for a different purpose.

The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Chapter VI with Articles 57 to 59 has applied since 2 August 2026; the deadline by which every Member State must have a national sandbox operational stands at 2 August 2027 in the consolidated text of paragraph 1. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.

What Article 57 requires

  • Art. 57(1) to (4): Every Member State ensures at least one national AI regulatory sandbox. The European Data Protection Supervisor may set one up for the Union institutions (paragraph 3), the AI Office one at Union level for systems under Art. 75(1) (paragraph 3a).
  • Art. 57(5): The sandbox is a controlled environment in which an innovative system is developed, trained, validated and tested for a limited time, under a sandbox plan agreed between provider and authority which, under Art. 3(54), describes objectives, conditions, timeframe, methodology and requirements. Real-world testing may be part of it; the sandbox plan then incorporates the plan under Art. 60 and 60a.
  • Art. 57(6) and (7): The authority guides, supervises and supports on risks. It explains its expectations and issues written proof of the activities successfully carried out as well as an exit report. The provider may use both to demonstrate compliance; market surveillance and notified bodies take them positively into account to accelerate conformity assessment to a reasonable extent. That is not a formal presumption of conformity.
  • Art. 57(8) to (10): The Commission and the Board get access to the exit report with agreement. The objectives are legal certainty, regulatory learning and faster market access (paragraph 9). Where personal data are processed, the data protection authorities must be involved (paragraph 10). Paragraphs 13 to 17 govern coordination within the Board, notification to the AI Office, a public list and annual reports.
  • Art. 57(11) and (12): Supervisory and corrective powers are unaffected; significant risks that cannot be mitigated lead to suspension. Participants remain liable under Union and national law for damage to third parties. Whoever observes the plan and the terms of participation and follows the authority’s guidance in good faith receives no administrative fine for infringements of the Regulation.
  • Art. 58: The Commission sets eligibility, participation, monitoring and exit by implementing act (paragraph 1): a decision within three months, joint applications with deployers, no fees for SMEs, time-limited participation, mutual recognition, tools for testing and explainability (paragraph 2), safeguards for real-world testing (paragraph 4).
  • Art. 59(1): Personal data collected for other purposes may be further processed in the sandbox if all conditions are met together: substantial public interest in a listed area, such as health, environment, energy, critical infrastructure or public administration (a); anonymised or synthetic data are not enough (b); risk detection with a stop (c); an isolated, protected environment under the provider’s control (d); data created in the sandbox stay there (e); no decision affecting data subjects (f); deletion when participation ends (g); logs (h); a description of training and testing under Annex IV (i); a short public summary (j).
  • Art. 59(2) and (3): Law enforcement needs its own legal basis. Other legal bases and prohibitions on changing the purpose remain unaffected.

The technical example

A bank builds a credit pre-check model itself, Annex III point 5(b), and is a prospective provider. Before 2 December 2027 it wants to know whether its gates meet the authority’s expectations. The data flow is the same as later in production: application from the credit portal to the policy gateway, pseudonymisation, scope check, model call, back come score band, confidence and the features that carried the result. Art. 59 does not help the bank, because credit pre-checks are not an area under paragraph 1(a); its training data stay on the bank’s own GDPR basis, pseudonymised or synthetic. What it gets is the guidance under Art. 57(7) and, if the authority agrees, real-world testing in shadow mode: real applications, no effect on the decision. The fact that the German financial supervisor has publicly shown openness to sandboxes and such tests makes this route more realistic than many assume.

Second, a social benefits authority that wants to train a model to pre-sort benefit applications for fraud patterns, Annex III point 5(a), on case files that were collected to pay benefits. Here Art. 59 applies through paragraph 1(a)(v), efficiency and quality of public administration. The authority must show that synthetic data are not enough (b), isolate the data (d), and the model may not influence any decision about an applicant (f): it runs on historical cases, and its outputs reach no caseworker. If the purpose is criminal prosecution rather than administrative checking, paragraph 2 demands a specific legal basis; that line is drawn by the data protection authority, not by the project team.

How we implement it

At SIFAMO we treat the sandbox not as an exception from governance but as its first environment. The governance layer runs there in full. That sounds like a lot of effort for a prototype, but it is the reason the exit report is worth something later.

Sandbox plan and exit report as registry records. The model’s registry record gets the status “sandbox”, with the authority, the duration and the sandbox plan as a versioned document with a hash. The expiry is the end of participation; after that the model loses its route, unless the release gate has moved it to “production”. Every change to the plan is a plan amendment agreed with the authority and filed in the registry. Exit report and written proof under paragraph 7 hang on the same record and move into the technical documentation under Annex IV.

The same layer, so the evidence carries over. Registry, policy gateway, evidence log, tiered oversight and kill switch run in the sandbox as in production, and so do the quality gates: evaluation set owned by the business owner, regression tests, adversarial tests, bias check, rollback plan. Every run writes to the evidence log: input hash, model and version, policy version, output, signals. The exit report is generated from this log, not from memory. On the move to production, record, evaluation history and log stay; only status, route and oversight tier change.

Data zones for Article 59. For the benefits authority we put the further-processed data in a dedicated red zone, separated from the administration at network level: its own network segment, its own credentials, no outbound connections, an access list with names. That is our reading of “functionally separate, isolated and protected” under paragraph 1(d). The purpose is a field in the registry, and the gateway rejects every call that does not carry it. Pseudonymisation with IDs before the model, here too. Point (e) means for us: outputs, embeddings and intermediate data never leave the zone; point (f): no route to any case-handling system. Deletion under point (g) is a job on the expiry date with a deletion record in the registry. The logs under point (h) are the evidence log, the summary under point (j) comes from the purpose field.

The authority’s access. The Regulation does not say how the authority supervises. We agree it in the sandbox plan: a dedicated role with read access to the registry record, the pseudonymised evidence log and the monitoring view. The kill switch stays with the business owner, but the plan obliges them to execute a suspension under paragraph 11 within a defined time and to log it. I consider that cleaner than a switch for the authority in our system: they supervise, we operate.

The liability point. Paragraph 12 protects against fines, not against liability. That is why real-world testing gets the same oversight tier as a production decision about people: full review, no sampling. And the protection hangs on the plan. Whoever quietly deviates loses it. That is why the plan amendment is a mandatory step for us.

Typical mistakes

  • The sandbox as a rule-free zone. Paragraph 11 leaves all powers of the authorities intact, paragraph 12 the liability. What happens there is more visible to supervisors, not less.
  • Prototype without governance. Then the exit report describes a system that no longer exists in that form, and the evidence starts from zero at release.
  • Article 59 for business interests. Further processing applies only to the listed areas in the public interest. A bank needs its own legal basis.
  • Silent deviation from the plan. New model or new data source without agreement with the authority. The protection from fines under paragraph 12 hangs on exactly this plan.

EU AI Act Article 57 checklist

  • Sandbox plan agreed and filed in the registry as a versioned artefact with hash; expiry equals end of participation.
  • Governance layer fully active in the sandbox: registry, gateway, evidence log, tiered oversight, kill switch.
  • Plan amendment as a mandatory step for every change to model, data or scope; agreement with the authority logged.
  • For Art. 59: area under paragraph 1(a) shown, necessity justified, isolated zone with no route to case-handling systems, deletion job with record, public summary.
  • Read-only role for the authority set up; suspension deadline under paragraph 11 agreed.
  • Exit report generated from the evidence log and included with the written proof in the technical documentation under Annex IV.

Open questions

  • The implementing acts under Art. 58(1) have yet to define eligibility and exit. Until then it is open how uniform the procedures will be.
  • Which authority in Germany runs the sandbox and how the financial supervisor is involved is not finally settled. For a bank that decides who sits at the table.
  • What “accelerate to a reasonable extent” under paragraph 7 means in practice is undefined. That the exit report replaces parts of the internal control under Annex VI is a hope, not a settled reading.

Frequently asked questions

Does this also apply to us as a deployer?

The sandbox is aimed at providers and prospective providers. As a deployer you can apply jointly with the provider under Art. 58(2)(b); your obligations under Art. 26 remain. What you need anyway is under AI governance.

Does the sandbox replace the high-risk obligations?

No. Guidance, written proof and exit report must be taken positively into account in the conformity assessment, nothing more. You still have to meet Chapter III Section 2.

Do we have to join a sandbox?

No, participation is voluntary. It makes sense if you build a high-risk system yourself or, as a public body, want to further process data under Art. 59. The other articles in the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment