Regulatory Sandboxes for Medical Devices: The proposed MDR/IVDR framework and its interplay with the AI Act

Introduction 

Emerging healthcare technologies increasingly challenge regulatory frameworks built around relatively stable products, clearly defined intended purposes and evidence generated before market entry. AI-enabled medical device software may evolve through model updates, rely on complex data pipelines and perform functions that do not fit neatly within existing clinical or conformity-assessment pathways. Because such software sits at the intersection of clinical safety, data processing and cybersecurity, developers must navigate several regimes at once: the Medical Devices Regulation (MDR) or In Vitro Diagnostic Medical Devices Regulation (IVDR), the AI Act, data-protection law, cybersecurity requirements and product-liability rules. 

Regulatory sandboxes are intended to address part of this problem by providing a controlled and time-limited setting in which an innovative product or technology can be developed, tested or validated under the supervision of a competent authority - with specific requirements temporarily adapted or waived where the standard rules do not fit. The objective is twofold: to help the participant understand and address applicable legal requirements, and to allow regulators to learn how existing rules operate when applied to new technologies. 

The European Commission’s proposal to amend the MDR and IVDR (Simplification Proposal) published on 16 December 2025, introduces a wide range of changes to the medical devices framework - we have already discussed it our previous blogpost in relation to the classification of medical device software (MDSW). Among its other reforms, the Simplification Proposal introduces dedicated regulatory sandboxes at both Member State and Union level. The proposed national model would, in narrowly defined circumstances, allow specified MDR or IVDR requirements to be temporarily adapted or waived under a supervised sandbox plan. The proposed Union model would instead allow the Commission to test whether existing regulatory requirements remain appropriate for a category of devices or an emerging technology. The proposal remains pending legislation and may change during the legislative process. 

Regulatory Sandboxes: The General Concept 

While there is no universal definition, a regulatory sandbox is understood as a structured and time-limited framework in which innovators may develop and test new products, services, technologies or business models in a controlled, often real-world environment under the supervision of one or more competent authorities. Testing takes place under a specific plan and with safeguards designed to preserve overarching regulatory objectives, including safety, consumer protection and fundamental rights. Depending on the legal basis, authorities may provide tailored guidance, apply requirements flexibly or permit narrowly defined temporary derogations. A derogation is not essential, but competent-authority involvement is a defining feature. 

The principal purpose is to bridge the gap between innovation and regulation. For innovators, a sandbox can clarify applicable rules, support compliance-by-design and reduce uncertainty before market entry. For regulators, it generates operational evidence about emerging technologies, potentially informing guidance, supervisory practice or legislative reform. 

Sandboxes may be bottom-up, where an innovator identifies the regulatory barrier, or top-down, where an authority defines the issue and invites participants and hybrid models are also possible. They differ from testbeds, which focus primarily on technical evidence, and living labs, which emphasise co-creation and societal impact.

At the EU level, sandboxes are increasingly supported by complementary infrastructures. These include the EUSAiR project, which supports Member States in establishing AI sandboxes, Testing and Experimentation Facilities providing specialised technical environments, and European Digital Innovation Hubs offering regional expertise and testing support, particularly for SMEs and start-ups. 

A healthcare-specific example is TEF-Health, the European Testing and Experimentation Facility for AI in health. It combines physical and virtual facilities across hospitals, universities, research institutions and testing bodies to support the integration, testing, validation and demonstration of AI-based healthcare products in clinical and care settings. It is open to technology providers, including start-ups and SMEs, and can provide access to technical infrastructure, clinical expertise, testing services and regulatory-pathway support. TEF-Health is not itself a regulatory sandbox, but it may supply the technical and evidentiary environment needed to support sandbox participation 

Regulation of sandboxes in the EU 

Regulatory sandboxes form part of the European Commission’s broader strategy for digital regulation: supporting innovation and growth while maintaining the protection of fundamental rights, health and safety. Across the EU, sandboxes have developed through a mixture of sector-specific EU legislation, national laws, experimentation clauses, government regulations and initiatives led by regulatory authorities. 

The Council’s 2020 conclusions on regulatory sandboxes and experimentation clauses called for an EU-wide exchange of information and good practices, including an overview of national initiatives, analysis of their legal bases and evaluation of how lessons from national sandboxes could support evidence-based policymaking at EU level. This was followed by the Commission’s 2023 guidance on regulatory sandboxes, testbeds and living labs, which sought to build a common understanding of their design and operation. 

Sandboxes under the AI Act, as amended by the Digital Omnibus 

The AI Act itself defines an 'AI regulatory sandbox' as a controlled framework set up by a competent authority which offers providers or prospective providers of AI systems 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. Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI) amends the horizontally-mandated sandbox framework established by the AI Act. Under the amended Article 57(1), Member States must ensure their competent authorities operate at least one national AI regulatory sandbox by 2 August 2027. 

The amendments introduce: 

• Union-level sandbox (Article 57(3a)): The European AI Office may establish a centralized sandbox for general-purpose AI models and very large online platforms or search engines 
• National sandboxes: Mandatory Member State establishment with discretion over design • Multi-authority coordination (Article 57(10)): Data protection, market surveillance, and sector-specific authorities must participate in sandbox governance 
• Commission governance framework (Article 58): Implementing acts will standardize procedures across Member States 
• Real-world testing integration: Sandbox plans incorporate testing protocols (Article 57(5)) 
• Real-world testing outside sandboxes (Article 60): Testing outside a sandbox, previously limited to Annex III systems, now also covers high-risk AI systems related to Annex I Section A products - a category that, as of today, still includes medical devices and in vitro diagnostics. A new, parallel provision, Article 60a, creates the equivalent route for Section B products.

The AI sandbox model differs from predecessors by imposing mandatory nationwide establishment and explicitly mandating multi-authority oversight, reflecting AI systems' cross-sectoral impact. 

MDR/IVDR proposal for Union and national level sandboxes: key features and conditions for setting up 

Under the Simplification Proposal, the Commission would introduce regulatory sandboxes into both the MDR and the IVDR. The proposal defines a 'regulatory sandbox' as a controlled environment set up by a competent authority which offers manufacturers or prospective manufacturers the possibility to develop, test, validate and use, where appropriate in real-world conditions, an innovative product or technology potentially falling within the scope of the MDR/IVDR, pursuant to a sandbox plan for a limited time under regulatory supervision. The Commission proposal distinguishes between national sandboxes, which are designed around a specific device or project, and Union sandboxes, which are intended to test the adequacy of the regulatory framework for a category of devices or an emerging technology. 

Regulatory sandbox at national level 

A Member State could establish one or more sandboxes on its own initiative or following a substantiated request from a manufacturer or prospective manufacturer. Two or more Member States could also establish a joint sandbox. The responsible competent authority would supervise the project. 

The activities would have to take place under a specific sandbox plan identifying the MDR or IVDR requirements to be temporarily adapted or waived, explaining why their ordinary application is inappropriate and describing how the resulting risks will be controlled and mitigated. The plan would also identify the duration, participants and their respective roles. The Member State may request an expert panel to provide scientific, technical or regulatory advice on the draft sandbox plan before it is finalised. 

A national MDR and IVDR sandbox could provide targeted derogations from the application of the following provisions: 

The sandbox plan must identify the particular requirement being adapted or waived and justify why the proposed alternative safeguards remain adequate. Establishing the sandbox would be subject to two cumulative conditions: 

• the device must be expected to address an unmet medical need or provide a significant clinical benefit to patients or to the health system compared with similar existing alternatives or the state of the art; and 
• the application of the relevant MDR or IVDR requirements would have to impede or significantly delay both the development of the device and access to it by healthcare professionals or lay users. Novelty by itself would therefore not be enough. The applicant would need to link the technology to a concrete clinical or health-system benefit and a specific regulatory obstacle. 

As noted above, the foundation of the entire mechanism is the sandbox plan. Beyond the minimum elements the plan must legally contain, a robust sandbox plan for an AI-enabled medical device software should, in our view, also address: 

• the intended purpose, target population and clinical role of the device; 
• the innovative or emerging technological feature and the regulatory uncertainty it creates; 
• the precise legal requirements to be tested, adapted or waived; 
• the testing methodology, endpoints and success criteria; 
• clinical, technical, data-protection and cybersecurity risks; 
• model versioning, dataset provenance, representativeness, bias and human oversight; 
• incident reporting, stopping rules and corrective measures; and 
• the exit strategy and use of the evidence in a later conformity assessment. 

Regulatory sandbox at Union level 

The Commission could establish a Union regulatory sandbox, by means of an implementing act, for a limited time and pursuant to a specific plan, on its own initiative or following a substantiated request from a Member State. 

The purpose would be to determine whether existing requirements appropriately regulate a particular type of device with specific characteristics or an emerging technology. A Union sandbox could be justified where the existing rules risk impeding or significantly delaying development and access, or where they may not adequately protect patients, users, other persons or public health. 

The Commission would be required to obtain expert-panel advice on the design of the Union sandbox. Importantly, a Union sandbox could not involve placing on the market or putting into service a device that does not comply with the MDR or IVDR. It is therefore primarily a regulatory-learning instrument rather than an alternative market-access pathway. 

Notably, the Commission may in future amend these provisions by delegated act to reflect experience with regulatory sandboxes in other sectors entirely which is an explicit acknowledgment that the MDR/IVDR sandbox model is not meant to develop in isolation from sandbox practice in other domains in the EU law. 

Interplay with AI Act as amended by the Digital Omnibus on AI 

The Simplification Proposal would move both MDR and IVDR from Section A to Section B of Annex I of the AI Act. That shift is intended to avoid a duplicative layer of AI Act obligations for AI-enabled medical devices and to make the sectoral medical-device framework the principal route for compliance. 

Under Article 2(2) of the AI Act, as amended by the Digital Omnibus on AI, only a limited set of AI Act provisions applies directly to high-risk AI systems related to products governed by Section B legislation. These include Article 6(1), new Article 60a and Articles 102 to 112. Articles 57 to 59 on AI regulatory sandboxes apply only insofar as the relevant high-risk AI requirements have been integrated into the applicable sectoral legislation in Section B, Annex I. The Digital Omnibus also inserts a parallel, converse mechanism for high-risk AI systems classified under Article 6(1) - that is, systems that are a safety component of, or are themselves, a product covered by Section A legislation (new Article 2(13)): the Commission may adopt delegated acts, by 2 August 2027, limiting specific AI Act obligations (Articles 9–15 and 17–25) for such systems, where the relevant Section A sectoral legislation already provides an equivalent or higher level of protection of health, safety or fundamental rights, and the limitation would not reduce the overall level of protection under the AI Act. 

In the context of a medical device software that falls within the scope of a high-risk AI system under Article 6(1) of the AI Act, the Commission would introduce specific requirements for such high-risk AI systems taking into account the specific requirements set out in the AI Act, by using its implementing and delegated powers. 

New Article 60a, inserted by the Digital Omnibus, lets Member States set up a real-world testing framework for AI systems covered by Section B legislation, with a view to assessing and verifying conformity with the AI Act's high-risk requirements. Because MDR and IVDR are not yet Section B instruments, Article 60a does not currently apply to medical device software. The regime that applies today is the amended Article 60, which the Digital Omnibus extended from Annex III systems to also cover high-risk AI systems related to Section A products - the category medical device AI systems currently sit in. Article 60a will only become the relevant real-world-testing route for medical devices if the Simplification Proposal is adopted and amends Annex I of the AI Act by moving both MDR and IVDR out of Section A and into Section B. 

Based on current guidelines published by the MDCG on interplay between the AI Act and the MDR/IVDR, real-world testing constitutes a clinical investigation or performance study of a high-risk medical AI system. Under the proposed regulatory sandbox framework real world testing is a constituent element which must be included in a sandbox plan, which is not mandatory in nature and subject to targeted derogations from the MDR and IVDR as deemed necessary. Such derogations would include requirements for clinical investigation or performance study. 

Pending adoption of the Simplification Proposal, a developer of medical device software that qualifies as a high-risk AI system should bear in mind that any real-world testing outside a sandbox conducted today already falls under the amended Article 60 rather than Article 60a. Where a developer instead conducts real-world testing within an AI Act sandbox, the amended Article 57(5) requires the sandbox plan to incorporate the Article 60 real-world testing plan, so the same procedural framework applies in either setting. Critically, however, such testing is not yet mandatory: the substantive Chapter III requirements it would demonstrate compliance with do not become binding for Article 6(1)/Annex I systems - including medical devices and in vitro diagnostics - until 2 August 2028. Until then, real-world testing under Article 60, whether conducted inside or outside a sandbox, is a voluntary, preparatory step, not a compliance obligation. Further clarification may still be issued by the Commission or the MDCG. 

Prior example of sandboxes in the context of healthcare 

Existing healthcare initiatives demonstrate that a sandbox can provide substantial value even without formal derogations. In 2021, the French data-protection authority, the Commission nationale de l’informatique et des libertés (CNIL), launched a sandbox for innovative digital health projects. Its focus was compliance with privacy and data-protection law, including lawful processing, consent, data minimisation, cloud infrastructure, data flows and data-protection impact assessments. For instance one of the projects targeted software supporting clinical decision support making with recommendations to define the purpose of data collection, verification of data sources, data minimisation, data retention period, verification of informed consent compatible with reuse of data and similar transparency measures. 

More recently, a German AI regulatory sandbox pilot project led by the Bundesnetzagentur, the Federal Commissioner for Data Protection and Freedom of Information and the Hessian Ministry for Digitalisation and Innovation tested the practical implementation of the AI Act sandbox model using two medical-device use cases. The pilot found that clear selection criteria, confidentiality, professional communication, informative documentation and an appropriate level of project maturity are essential. It also showed that companies primarily expect regulatory guidance, while authorities need mechanisms to convert project-specific experience into broader regulatory learning. 

The German pilot further emphasised that a sandbox should not function as a general legal-advice service. The legal uncertainty should be substantial, relevant beyond the individual project and capable of producing generalisable lessons. Participating products should also be mature enough to support meaningful analysis but still capable of being adjusted in response to regulatory feedback. 

Overlap between the AI Act, MDR and the GDPR is also identified as a central issue. These practical lessons are directly relevant to the future operation of MDR and IVDR sandboxes. 

How does sandbox plan impact software classification? 

The proposed sandbox framework may be particularly relevant to medical device software because Annex VIII classification rules are among the requirements that could potentially be adapted or waived in a national MDR sandbox plan. However, sandbox participation does not remove the need to resolve two distinct questions of qualification as a medical device software and its classification which was examined in more detail in our earlier blog post. The CJEU’s recent judgment in C-427/24, Zentrale v Diagramm Halbach, reinforces the importance of the intended purpose of the medical device. The Court held that account must be taken not only of information supplied by the manufacturer, but also of whether the product objectively performs one or more of the functions listed in Article 2(1) MDR. Use in a healthcare environment or the handling of patient-related information is not, by itself, sufficient to establish medical-device status. 

For software, this means that a carefully drafted intended-purpose statement remains essential, but it must correspond to the product’s objective functionality. Software that merely stores, organises or transmits information should be distinguished from software that analyses patient-specific data for diagnosis, prediction, prognosis, monitoring or treatment. 

The Simplification Proposal also introduces a separate route for obtaining an expert panel opinion on regulatory status. A competent authority, notified body, manufacturer, developer or the Commission could submit a substantiated request asking whether a product falls within the medical-device definition and, where the requester considers that it does, specifying the proposed classification. The opinion of the expert panel would inform the analysis, while the ordinary MDR classification and conformity-assessment framework would remain relevant. 

A sandbox may therefore provide a structured environment in which classification assumptions are tested against the product’s evolving functions and evidence and subject to targeted derogations. 

WHAT SANDBOX PARTICIPATION DOES AND DOES NOT PROVIDE 

Sandbox participation offers genuine benefits, but it is important to understand its limits. 

What participation provides: 

• Regulatory dialogue. Ongoing engagement with a competent authority; the opportunity to ask questions, test interpretations, and receive feedback without waiting months for written guidance. 
• Documented regulatory position. The sandbox plan, exit report, and authority guidance constitute a record of the regulator's position on specific compliance questions as applied to the participant's system. This record can inform the participant's conformity assessment and risk management strategy. 
• Potential for derogations. Where the authority agrees that standard requirements are inappropriate, limited and temporary relief may be negotiated, with alternative safeguards substituted. 
• Structured testing environment. The sandbox provides a formal, supervised setting for testing in real-world conditions, with monitored risks and agreed stopping rules, reducing legal uncertainty compared to informal testing.
• Reduced administrative fines during participation. Article 57(11) states that no administrative fines shall be imposed for infringements during the sandbox period, provided the participant observes the sandbox plan and follows the authority's guidance in good faith. This protection can extend to infringements of other Union or national law too - but only where the authority responsible for that other law was actively and directly involved in supervising the sandbox and gave its own guidance.

What participation does NOT provide: 

• Compliance approval. Sandbox participation is not certification. The authority's guidance and the sandbox exit report do not constitute conformity assessment, CE marking, or a legal guarantee of compliance. When the participant exits the sandbox and places the system on the market, full conformity assessment under applicable law remains required. 
• Liability protection. Civil liability to third parties remains regardless of sandbox compliance. If the AI system/medical device causes harm, the participant may be liable to injured parties even if the sandbox plan was scrupulously followed. 
• Exemption from other law. Sandbox participation does not exempt the system from GDPR, MDR, sectoral cybersecurity law, or other applicable frameworks. The sandbox operates within the context of those laws, not instead of them. 
• Portability across jurisdictions. Findings and guidance from a sandbox in one Member State do not automatically apply in another. A manufacturer participating in a German sandbox who later markets the system in France may face different interpretations and requirements. 
• Permanent regulatory relief. The sandbox is time-limited any adaptation or waiver of a requirement is bounded to the sandbox period, not a permanent exemption. At the end of the period, the system must comply with all applicable requirements unless the requirements themselves have been formally changed through legislative amendment or delegated/implementing acts. The authority can only temporarily adapt a requirement, not permanently waive it. 

SHOULD YOU CONSIDER A SANDBOX? PRACTICAL GUIDANCE FOR DECISION-MAKING 

The AI Act sandbox framework is adopted law: Member States must operate at least one national AI regulatory sandbox by 2 August 2027, and the amended Article 57(5) already requires any sandbox plan that includes real-world testing to incorporate the Article 60 testing plan. By contrast, the MDR and IVDR sandbox mechanism exists only in the Commission's Simplification Proposal and will become available only if and when that proposal is adopted. The practical guidance below is therefore immediately relevant for developers considering an AI Act sandbox, and prospectively relevant for those planning around the proposed MDR/IVDR route. Where both regimes may eventually apply to the same product, the threshold questions and the commitment they require are broadly comparable.Before approaching a competent authority, assess these questions honestly: 

• Does your system address an unmet need or provide material clinical or operational benefit that existing solutions do not deliver? 
• Have you identified specific regulatory requirements or interpretations that are genuinely unclear, not merely difficult or burdensome? 
• Is your product mature enough to support meaningful testing and capable of being adjusted in response to regulatory feedback? 
• Can you commit 6–24 months of regulatory engagement and document your compliance decisions for authority scrutiny? 

If you cannot answer 'yes' to most of these, a sandbox may delay your market entry more than help it. Instead, consider direct early engagement with the authority, legal counsel to map the conformity pathway, or a regulatory gap analysis against existing standards.

If you decide to proceed: 

• Develop a concrete sandbox plan. Identify the precise regulatory question you are testing, the testing methodology, success criteria, and what alternative safeguards you propose. Vague applications will not succeed. 
• Engage the authority early before submitting a formal application. Understand their published criteria, current workload, and whether they view your issue as within scope. Informal dialogue prevents wasted effort. 
• Align internal stakeholders (product, compliance, legal, business) on the commitment. Sandbox participation affects development timelines, data governance, and market strategy. Ensure alignment before you begin. 
• Prepare technical documentation in draft form: risk assessments, testing protocols, data governance plans, incident-reporting procedures. This demonstrates maturity and reduces authority assessment time. Document everything: your submissions, the authority's feedback, your responses, and the reasoning behind your compliance decisions. This documentation feeds into conformity assessment and protects you in enforcement scenarios. Plan from day one how sandbox evidence will reduce the notified body's conformity assessment burden and what additional compliance work remains. 

Funding acknowledgement and disclaimer 

The authors are involved in EU-funded projects relevant to digital health, health data and AI-enabled healthcare technologies.

• Nayana Murali participates in the PATH project. PATH is supported by the Innovative Health Initiative Joint Undertaking (IHI JU) under Grant Agreement No. 101253520. The IHI JU receives support from the European Union’s Horizon Europe research and innovation programme and COCIR, EFPIA, EuropaBio, MedTech Europe and Vaccines Europe.
•  Marta Wilinska participates in the AI4LUNGS project. AI4LUNGS is funded by the European Union under the Horizon Europe Programme, Grant Agreement No. 101080756. The Health and Digital Executive Agency (HaDEA) is the granting authority. 
• Vaisakh Shaji participates in VOLABIOS project. VOLABIOS is funded by the European Union under the Horizon Europe Programme, Grant Agreement No. 101156162. The Health and Digital Executive Agency (HaDEA) is the granting authority. 

Views and opinions expressed in this article are those of the authors only and do not necessarily reflect those of the European Union, the Innovative Health Initiative Joint Undertaking (IHI JU), the Health and Digital Executive Agency (HaDEA), or any other relevant granting authority. Neither the European Union, IHI JU, HaDEA nor any other relevant granting authority can be held responsible for them.