Inferensys

Glossary

Regulatory Sandbox Notification

The formal process of informing a National Competent Authority that an AI system is being tested under a controlled regulatory sandbox, often with modified registration requirements.
Developer building agentic RAG system, retrieval pipeline diagram on laptop, technical workspace with notes.
AI SYSTEM REGISTRATION

What is Regulatory Sandbox Notification?

The formal process of informing a National Competent Authority that an AI system is being tested under a controlled regulatory sandbox, often with modified registration requirements.

Regulatory Sandbox Notification is the formal administrative procedure by which a provider or prospective provider informs a National Competent Authority that a specific AI system will undergo development, testing, and validation within a supervised regulatory sandbox environment. This notification serves as a prerequisite for accessing the sandbox's controlled conditions, where certain standard registration and conformity assessment requirements may be temporarily modified or waived under the direct supervision of the competent authority.

The notification dossier typically includes a detailed Intended Purpose Declaration, a description of the testing plan, and a preliminary risk assessment. Unlike a full EU AI Act Database registration, the sandbox notification establishes a temporary legal safe harbor, allowing innovators to validate high-risk systems under real-world conditions without immediate full compliance. The competent authority uses this notification to define the sandbox plan's specific parameters, exit criteria, and the applicable supervisory modalities.

REGULATORY MECHANISM

Key Features of a Sandbox Notification

A formal declaration to a National Competent Authority enabling supervised testing of an AI system under a modified legal framework. The notification triggers specific derogations from standard registration requirements while maintaining a controlled environment for innovation.

01

Modified Registration Requirements

The sandbox notification activates a derogation from full pre-market registration. Instead of submitting the complete Technical Documentation File, providers file a sandbox-specific plan outlining:

  • The testing scope and duration
  • The specific regulatory provisions being derogated
  • The risk mitigation measures in place

This allows for iterative development without triggering the full conformity assessment process under the EU AI Act.

Art. 53-54
EU AI Act Legal Basis
02

Competent Authority Oversight

The notification is submitted to the National Competent Authority (NCA) of the member state where the sandbox operates. The NCA:

  • Reviews the sandbox plan for completeness
  • Issues written guidance on regulatory expectations
  • Monitors testing activities for compliance with the agreed scope

This creates a supervised learning loop where regulators provide real-time feedback before market entry.

03

Time-Bound Authorization

A sandbox notification is not an indefinite license. It specifies a fixed duration aligned with the testing phase. Key temporal constraints include:

  • A defined start and end date for the sandbox period
  • An exit plan detailing transition to full registration
  • A requirement to notify the NCA of any substantial modification during testing

Upon expiration, the provider must either pursue full registration or cease operations.

04

Liability and Safeguard Provisions

The notification does not exempt providers from liability. The sandbox framework requires:

  • Mandatory safeguards to protect fundamental rights
  • Clear informed consent from sandbox participants
  • A residual risk disclosure specific to the testing environment

Any serious incident occurring during the sandbox phase must be reported through the incident reporting linkage tied to the provisional registration ID.

05

Cross-Border Recognition

A sandbox notification approved in one EU member state can facilitate cross-border testing under mutual recognition principles. This requires:

  • Coordination between the home NCA and host member states
  • Adherence to the harmonized standard for sandbox operations
  • A unified Unique Registration ID for traceability across jurisdictions

This mechanism prevents regulatory fragmentation while allowing localized supervision.

06

Transition to Full Registration

The sandbox notification serves as a precursor to formal registration. The exit process mandates:

  • Compilation of a complete Technical Documentation File incorporating sandbox learnings
  • A finalized Declaration of Conformity based on validated performance data
  • Submission to the EU AI Act Database for a permanent Unique Registration ID

The sandbox phase effectively acts as a pre-market authorization proving ground.

REGULATORY SANDBOX NOTIFICATION

Frequently Asked Questions

Clarifying the formal process and legal implications of notifying National Competent Authorities when testing high-risk AI systems under controlled regulatory sandbox conditions.

A Regulatory Sandbox Notification is the formal legal process by which a prospective AI provider or authorized representative informs a National Competent Authority (NCA) of their intent to develop, test, and validate an innovative AI system within a controlled regulatory environment. This notification serves as a structured entry point that temporarily modifies standard Conformity Assessment and registration requirements, allowing supervised experimentation under a specific sandbox plan. The notification must include a detailed description of the planned testing activities, the expected duration, and the specific regulatory flexibilities being requested. Crucially, this process does not exempt the provider from liability for harm caused to third parties; it merely provides a supervised space to gather evidence for future compliance under the EU AI Act.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.