1. Eligibility criteria

Chapter VI of the AIA outlines the operational rules for regulatory sandboxes but does not explicitly define eligibility criteria, that is, the parameters used to assess whether applicants qualify for sandbox entry. Article 58(1)(a) delegates this task to the EU Commission, which is expected to specify the criteria through implementing acts. Until those acts are adopted, it would be premature to propose specific eligibility rules or amendments.Footnote ⁸ That said, the AI Act already contains principles and constraints that will shape whatever criteria the Commission adopts. We therefore focus more on “meta-criteria”: the high-level profiles and considerations that can be inferred from the Act itself.

Generally, the AIA intends for regulatory sandboxes to be inclusive rather than restrictive. Rather than limiting entry to a niche group of “cutting-edge” companies, the AIA opens sandbox participation broadly to any provider or prospective provider of AI systems. This is explicitly referenced in Recitals 139 and 141 and Article 3. In keeping with this aim, the AIA intends for eligibility to depend on objective criteria set by the Commission’s implementing act and potentially further specifications by the relevant national sandbox authority. Of course, these criteria may differ to some extent between jurisdictions or industries (within the limits of the Commission’s implementing act), but all must be transparent and publicly available to ensure that applicants can assess their likelihood of qualifying before applying.

That the AIA intends for AI regulatory sandboxes to be inclusive in design does not, however, mean that there will be no prioritisation of sandbox applicants. Instead, the AIA stresses the importance of supporting small and medium-sized enterprises (SMEs) (Recital 143 and Article 62(1)). This priority is operationalised in two key ways: (a) if two applicants meet the baseline eligibility criteria and score similarly in other respects, SMEs may be given precedence; (b) Article 58(2)(d) stipulates that participation in the sandbox should be free of charge for SMEs, except in exceptional circumstances where authorities may recover costs.

To anticipate how the eligibility criteria may look under the AIA, we can gain lessons from what happens in other sandboxes. In particular, Seferi highlights a set of recurring eligibility criteriaFootnote ⁹, with three primary criteria emerging as the most significant:

1. (1) The degree of innovativeness is the most frequently used eligibility criterion (with 74 occurrences) and assesses how novel and groundbreaking a proposed project or solution is. Specifically, it evaluates whether a project addresses clear market gaps or introduces entirely new functionalities, often through emerging technologies or novel business models. Both the Italian Financial Services Regulatory Sandbox and the UK Financial Conduct Authority’s (FCA) Regulatory Sandbox use this standard to assess applicants.

   The FCA’s Regulatory Sandbox specifically requires “genuine innovation” (Financial Conduct Authority 2015, 7) as an eligibility criterion. This explicitly requires cutting-edge financial innovations rather than incremental product enhancements. A genuinely innovative solution under these criteria might be a novel approach to delivering financial services to underserved groups, whereas minor improvements to existing services would not qualify.Footnote ¹⁰

   This binary distinction poses challenges when applied to AI, as AI’s transformative impact often lies in incremental optimisations rather than radical reinvention. Many AI applications, especially those based on machine learning, significantly refine existing processes such as automating customer service, enhancing predictive analytics or streamlining supply chains. However, these incremental improvements do not necessarily align with traditional definitions of “genuine innovation,” which focus on radically disruptive breakthroughs. For example, an AI system improving retail demand forecasting by 15% offers meaningful efficiency gains but may not be viewed as redefining the sector. To qualify as genuinely innovative, an AI-driven project may likely need to go beyond automation and create new market opportunities or solve market gaps in new ways. For example, an AI-powered credit scoring system that leverages alternative data sources to expand financial access for the unbanked would likely meet this threshold – particularly if it provides a significantly better alternative to traditional credit evaluation methods in ways that have not yet entered the market.

   The AI Act (AIA) implicitly addresses this challenge by adopting a broader definition of innovativeness. Although placed within Chapter VI (“Measures in Support of Innovation”), suggesting innovativeness will remain essential, the AIA is unlikely to restrict eligibility solely to radical disruption. Two main rationales support this broader interpretation:

   The technological rationale, as anticipated, is that AI systems often optimise existing processes rather than create entirely new ones. Machine learning is frequently deployed to enhance efficiency, as, qui for instance, in supply chain management or customer service. Improvements can be significant, but they typically do not represent a complete departure from existing solutions, either technologically or procedurally.

   The hermeneutic rationale is that the AIA does not explicitly require applicants to demonstrate “genuine novelty” or disruptive potential, as is common in other sandbox frameworks (or IP rights). Instead, the Act repeatedly prioritises legal uncertainty as a key trigger for sandbox participation (Recital 139). This suggests that the primary rationale for inclusion is the need for regulatory clarity or guidance, regardless of whether the system is “disruptively” novel. The AIA’s acceptance of non-disruptive innovations finds further justification in the pragmatic realities of non-EU companies assessing market entry strategies. Many commercially operational AI systems – particularly those developed in jurisdictions with less stringent regulations – may not qualify as innovative by conventional sandbox standards, yet their operators (and EU regulators) face genuine uncertainty about EU compliance pathways and the feasibility of EU market entry.

   In other words, innovativeness matters, but primarily to the extent that the project raises unusual or emergent technical or legal questions or provides an opportunity for providers and regulators to “learn by doing” in areas of uncertain compliance. Consequently, unlike some sandbox regimes (e.g., fintech), the AIA would not explicitly exclude incremental innovations from eligibility. The focus would likely be on whether the project presents non-trivial questions or challenges related to compliance with the Act’s obligations.

2. (2) Public interest or societal benefit is a less common but still significant eligibility criterion, appearing in sixty-seven cases. It assesses the broader societal impact of a project, including its contributions to public welfare and market benefits. More specifically, this means that a project or solution “advances the public interest or […] generates positive social impacts, such as improving accessibility, promoting sustainability, or enhancing public and private services”Footnote ¹¹.

   A positive indicator of eligibility may be, for instance, the project’s ability to expand access to essential services for underserved and marginalised communities or to support environmental conservation. For instance, the UK’s “AI Airlock” for AI as a medical device requires that applicants demonstrate that their product or prototype will provide a benefit to patients and public health. Conversely, a negative indicator may be that the project exclusively benefits privileged groups or contributes to environmental strain.

   It is, however, important to recognise that criteria related to societal benefit are, by their nature, subjective and often qualitative. The EU’s understanding of the value of innovation has evolved subsequent to the passing of the AIA, with the publication of the Draghi Report in September 2024, and the subsequent acceptance of its overarching points by European Commission President Ursula von der Leyen. The Report argues that increased productivity through innovation is essential to preserving the EU values of equity and social inclusion. As such, the scope of societal benefits expands beyond its traditional boundaries to cover innovation that supports the EU’s global competitiveness and economic growth. In this context, commercial innovations that support the EU’s competitiveness with the United States and China, such as developing alternative LLMs or reducing the EU’s reliance on US cloud infrastructure, would meet the public interest criterion. In any case, such value-subjectivity may make it more challenging to reach agreement as to whether a particular technology meets the public benefit criterion.

   The AIA does not explicitly list public reason or societal benefit as a condition for accessing regulatory sandboxes. This omission may appear unexpected, given that its risk-based categorisation and obligations for AI deployers are firmly grounded in the EU’s cultural and legal values, which prioritise societal welfare. The AIA’s prohibition of harmful practices, such as biometric categorisation, and its focus on safety and trustworthiness, reflect a clear principle: AI development should not be driven solely by profit or private interests but should aim to serve the public good.

   That said, claiming that public interest plays no role in AIA regulatory sandboxes would be inaccurate. The AIA does address it, notably in Article 59(1), which allows for “further processing of personal data for developing certain AI systems in the public interest.” This exception grants data-sharing flexibilities within the sandbox for AI systems addressing substantial public-interest domains, such as public health (e.g., disease detection), environmental protection and energy sustainability. Projects targeting recognised public interest areas may benefit from enhanced data-use permissions or enhanced engagement with regulators, resulting in expedited regulatory guidance. Conversely, projects without a clear societal benefit are not excluded but may not access the same advantages.

   Looking ahead, public interest in the regulatory sandbox mechanism could still be integrated into the AIA’s implementing acts more explicitly. To align with the Act’s overarching framework, such eligibility criteria would need to reflect its specific conception of public interest, which spans from safeguarding fundamental rights, ensuring safety, promoting transparency and fostering innovation. Challenges remain in balancing these values, as the AIA lacks detailed guidance on prioritising or reconciling them in practice, nor, one could argue, is it necessarily intended to do so. Such judgements may be more appropriately addressed on a case-by-case basis, for instance, through judicial interpretation or administrative decision-making. In this respect, a degree of legal ambiguity surrounding eligibility may be both inevitable and, to some extent, desirable.

3. (3) Level of maturity is a similarly significant criterion, appearing sixty-three times. It assesses whether a project is sufficiently advanced for meaningful experimentation within a sandbox and considers s several factors, including technical capability, financial sustainability and scalability.Footnote ¹² However, technological readiness for sandbox candidates should not be equated with that of fully developed and tested technologies; it should be evaluated based on their potential for structured experimentation. The level of maturity is closely tied to timing. Applicants must demonstrate that their innovations are at a stage where they are ready to be tested so that the sandbox serves as a platform for actionable experimentation rather than premature trials.Footnote ¹³ Required levels of maturity will differ between types of sandboxes (e.g., early-stage versus late-stage).

   A positive indicator of maturity – especially for a medium- or late-stage sandbox – may be having a functional and testable working prototype, secured funding, a clear business model and a demonstrated understanding of relevant regulatory requirements. For example, in the Hong Kong Monetary Authority’s (HKMA) Fintech sandbox, participants must demonstrate measures to protect customers’ interests, such as by providing timely and fair compensation for any financial losses caused by trial failures.Footnote ¹⁴ In jurisdictions that apply this criterion – such as Slovakia’s NBS Financial Services Regulatory Sandbox – project maturity is also assessed based on the presence of a comprehensive risk management strategy, including contingency planning for premature termination and threat identification. Similarly, the Zurich AI sandbox requires candidates to demonstrate the necessary technological expertise as a mandatory eligibility criterion. Conversely, a project at the concept stage or one that completely disregards legal and compliance requirements is unlikely to meet the eligibility threshold for such a sandbox type.

   When it comes to AI, we can predict that, for instance, an AI company proposing a self-learning regulatory compliance system without any developed code or proof of feasibility would likely fail to qualify. Similarly, an AI solution that fails to acknowledge GDPR requirements for data quality and privacy (even when submitted for the AIA sandbox) would struggle to meet the necessary maturity criteria.

   Among the three criteria analysed, technological maturity or readiness appears to align most closely with what the AIA already mandates. From the outset, the AIA indicates that a certain degree of maturity is expected. Article 3(54–55) defines a sandbox plan as an agreement between the applicant and the competent authority, requiring clear objectives, testing methods and a defined timeframe (further detailed in Article 58(2)(b)). These requirements inherently assume that the applicant has a workable model or at least a feasible approach – otherwise, the authority would have nothing to oversee or test.

   Further support for this eligibility criterion can be found in Article 57(5), which states that the sandbox applies to AI systems “for a limited time before their being placed on the market or put into service.” This suggests that applicants must already be developing a real solution–not yet commercially deployed, but advanced enough for meaningful testing.

   The importance of the level of maturity (or technological readiness) can also be inferred from how risks arising from the sandbox experimentation are handled. Article 57(6)–(7) and Recital 139 require sandbox participants to cooperate with authorities to identify and mitigate significant risks to fundamental rights or health and safety. If any “significant risks to health and safety and fundamental rights” arise and cannot be mitigated, the authority may suspend or terminate sandbox participation (Article 57(11)). By implication, the system must have some basic risk controls or mitigation strategies in place.

   For all these reasons, while we cannot yet read any reference to “maturity,” the AIA implies that completely unformed or purely conceptual ideas are unlikely to gain acceptance since you need (1) a credible plan, (2) the ability to run meaningful tests and (3) minimal risk-mitigation controls.

Other, somewhat secondary, eligibility criteria were also identifiedFootnote ¹⁵. One relates to the applicant’s ability to identify and mitigate potential risks (4), which is closely linked to the maturity level of both the project and the organisation behind it. The ability to identify and mitigate potential risks aligns well with the AIA’s risk-based nature, which requires deployers, including prospective ones, to establish internal risk management systems.

Eligibility also depends on whether the project falls within the supervisory jurisdiction of the relevant authority, meaning that sector-specific legislation overseen by that authority governs it (5), and whether the project demonstrates a clear need for testing or experimentation (6).

(5) ensures that the project is subject to specific legislation that is explicitly within the authority’s regulatory mandate. While some sandboxes – such as the Zurich AI Sandbox and to some extent the AIA itself – are sector-agnostic, covering domains such as autonomous systems, sustainability and healthcareFootnote ¹⁶, their flexibility remains bounded by applicable regulations. Accordingly, a national competent authority overseeing AIA-aligned sandboxes would reject projects involving technologies excluded under the AIA, such as autonomous weapons.

(6) focuses on whether the project can clearly justify the added value of testing in a controlled environment, showing that sandbox participation would deliver meaningful benefits, particularly through direct regulatory guidance provided by the authority. The AIA strongly implies the relevance of this criterion. In fact, applicants will likely need to show why regulatory supervision is necessary, as per the sandbox plan (Article 3(54) and Article 58(2)(b)), which requires detailing objectives, testing methods and expected outcomes. Thus, an entity applying for sandbox participation should demonstrate a genuine need to clarify how to comply with the AIA (and possibly other Union or national rules) or to test risk-mitigation measures in a controlled environment.

Finally, some additional and less frequently cited eligibility criteria include the existence of clear exit strategies, the potential of the applicant’s project to increase legal certainty by clarifying ambiguous or grey areas and the submission of a transparent and complete application that ensures all necessary details are provided for a streamlined selection process. These criteria also seem to be implied by the AIA, as already pointed out. So, for instance, Article 57(7) mandates an exit report, signalling that projects should have a defined scope and testing lifecycle, and Article 57(9) highlights legal certainty as a core sandbox objective, favouring projects that help clarify ambiguous regulatory areas.

2. The operational procedure: the novelties in the AI Act

Typically, the operational procedures of regulatory sandboxes progress through the following phases: application, preparation, testing, validation and exit.Footnote ¹⁷ Below, each phase is described comprehensively as it pertains specifically to the regulatory sandboxes under the AIA.

1. The role of the commission and EU-level coordination

To maintain clear accountability despite complexity, the national competent authority has primary responsibility for overseeing the day-to-day operations of regulatory sandboxes. It maintains direct contact with participants and can exercise immediate enforcement powers. Member States must ensure these authorities have sufficient financial and human resources.

Where needed – especially for data protection or other specialised areas – the lead sandbox authority may involve other national regulators to co-supervise matters within their remit. In these cases, data protection authorities or other relevant bodies must be formally associated with sandbox operations (Article 57(10)) to ensure compliance with the GDPR and applicable sectoral laws.

However, national authorities are not the only bodies involved. As outlined in Recital 143, the European Commission plays a critical role in providing the implementing acts to establish common principles for sandboxes across all Member States (e.g., eligibility criteria, terms and conditions and standardised templates to ensure broad alignment). The implementing acts are expected to function as a “procedural harmoniser”: they can reduce divergence by standardising key steps in sandbox governance (application, monitoring and exit/termination) and by promoting mutual recognition of sandbox participation and its effects across the internal market, thereby reducing incentives for regulatory shopping and easing cross-border scaling.

Additionally, the Commission may analyse annual reports submitted by national sandbox authorities. Although these reports are not directly addressed to the Commission, it can access them indirectly through the AI Board and the AI Office. The Commission uses these reports to inform its broader responsibilities under the AIA, especially for one of the most salient tasks, that is, enabling it to propose regulatory adjustments based on insights gained from sandbox testing.Footnote ³¹ Where confidentiality safeguards allow, access to exit reports and aggregated findings can further strengthen this feedback loop by giving the Commission a structured evidence base for refining implementing measures, guidance or templates. For instance, if sandbox testing reveals an unanticipated high risk related to the accuracy of facial recognition algorithms, the Commission may propose amendments to delegated or implementing acts, such as introducing explicit accuracy thresholds.

Apart from national authorities and the Commission, various AI ecosystem actors also have advisory and technical roles. These include testing facilities, research labs, innovation hubs and civil society organisations (Recital 139; Articles 57–8). Their contributions focus on testing, data provision, risk mitigation best practices and standardisation input.

Sandbox coordination also involves the European AI Board and the AI Office. The AI Board can offer recommendations for regulatory updates based on sandbox reports. The AI Office manages a registry of sandboxes (Article 57(11)) and supports cross-border collaboration by sharing best practices, incidents and recommendations. It may also assist Member States by providing templates for exit reports, test methodologies or cooperation models for supervisory authorities.

1. Pre-testing phase challenges and strategies

The pre-testing phase involves preparatory challenges before sandbox entry, including defining clear eligibility criteria, determining optimal pipeline placement, managing multi-agency coordination, preventing regulatory arbitrage, attracting suitable companies, integrating sandboxes with existing EU ecosystems and avoiding unintended expansions of scope (scope creep). These early choices have a lasting impact and require structured, proactive management to ensure clarity, fairness and coherence.

1. a) Pipeline Placement. Pipeline placement within regulatory sandboxes refers to determining the optimal stage – early conceptual design, mid-stage development or late-stage pre-market deployment – at which an AI system should enter the sandbox. It also refers to the possibility for AI providers to revisit earlier development stages for updates or rechecks as needed. Under the AIA, the timing of sandbox entry significantly influences the supervisory approach and governance mechanisms applied to the sandboxed AI systems:

   • • Early-stage sandbox entrants in the initial stages of development require a guidance-oriented and collaborative approach to supervision. Authorities at this stage should proactively support compliance by assisting applicants in interpreting and integrating regulatory requirements into their system designs, providing tailored advice (Article 57(6)).

   • • Mid-stage sandbox entrants, for example those currently engaged in model refinement, require more structured, more continuous supervision, from multiple relevant authorities. These authorities should coordinate and actively monitor adherence to predefined sandbox plans, ensuring rigorous compliance with risk mitigation measures and data protection standards (Article 59). Non-compliance or failures at this stage should trigger targeted interventions or even suspension of the AI system’s progress within the sandbox.

   • • Late-stage sandbox entrants, for example, those in the throes of pre-market validation, require stringent supervision characterised by rigorous compliance reviews and conformity assessments. At this stage, the sandbox functions as a critical compliance checkpoint, filtering out systems that fail to meet essential regulatory requirements before they reach the market (Article 57(7–8)).

   Strategy. Sandbox eligibility criteria and guidelines should explicitly delineate the different priorities, expectations and goals of different pipeline entry points.Footnote ³² Guidelines for early- and mid-stage entrants should emphasise the importance of proactive compliance support, make clear that entry at this point is (a) designed to help both regulators and innovators address regulatory uncertainties early in the AI system’s lifecycle and (b) most suited to AI systems with uncertain risks. Conversely, guidelines for late-stage entrants should stress that entry at this point in the pipeline is designed to (a) help innovators and regulators manage critical risks and ensure robust market-readiness and (b) is most suited to systems with clearly defined high risks (Recital 139 & Article 57).

2. b) Multi-agency collaboration and overlapping oversight. We have seen that the regulatory sandbox framework under the AIA involves multiple authorities. National competent authorities must establish at least one regulatory sandbox, potentially involving regional or joint arrangements (Article 57(1)–(2)). The European Data Protection Supervisor (EDPS) may set up a separate sandbox for EU institutions (Article 57(3)). National data protection authorities (DPAs) must oversee data compliance, especially given the frequent use of personal data (Articles 57(10), 59(1)). Other regulators – covering fundamental rights, safety, cybersecurity or sectors such as transport and energy – may also participate where relevant (Articles 57(4), 57(10)).

   The involvement of multiple authorities necessitates careful coordination and governance models, as it can lead to overlapping jurisdictions and potential conflicts. For example, AI systems placed in the sandbox might simultaneously fall under the oversight of general AI regulatory authorities and specialised sectoral regulators (healthcare, transport, energy and environment). As a result, regulatory sandboxes become convergence points where oversight from various legal frameworks intersects, complicating the clear delineation of regulatory responsibilities and accountability.

   Regarding the EDPS sandbox and national sandboxes, direct regulatory overlap should generally remain limited due to their distinct primary focuses – the EDPS sandbox predominantly targets EU institutions and agencies. However, indirect overlaps could still occur, especially when AI systems involve cross-border data transfers or testing in multiple jurisdictions. For instance, if an EU agency develops an AI system that requires practical testing involving personal data collected nationally or implementation within Member States, coordination between the EDPS and relevant national authorities would become essential.

   Strategy. A practical control strategy for mitigating these risks revolves around structured information and resource sharing. For instance, in the UK, the Information Commissioner’s Office (ICO) and the FCA successfully implemented collaborative mechanisms by embedding staff within each other’s sandbox initiatives. Such an arrangement facilitated alignment on overlapping issues like anti-money laundering (AML) and anonymisation and enabled joint participation in innovation-focused activities, such as jointly-run hackathons.

   At the same time, we cannot assume that this approach will automatically succeed in the EU context. Concrete efforts are needed to mitigate the heightened risks of confusion and tension arising from the EU’s complex regulatory and political landscape. Although the AI Board and the Commission’s implementing acts strive for uniformity, the existence of separate sandboxes may nonetheless yield divergent interpretations regarding, for instance, compliance requirements and risk mitigation strategies. Projects involving both EU institutions and national entities may raise questions of jurisdiction and primary responsibility, potentially causing confusion among participants regarding which sandbox or supervisory framework to engage with. Finally, differences in funding, expertise and infrastructure between national authorities and the EDPS may further compound these challenges. In essence, unlike the ICO and FCA coordinated approach, EU regulators have to face the challenge of systemic information sharing, underscoring the need for clear guidance and strong collaboration to manage multi-agency oversight. Cross-border cooperation and extensive sharing and exchange of use cases and testbeds will be crucial to facilitate the harmonisation in the interpretation of AIA concerning regulatory questions on innovative AI systems.

3. c) Regulatory arbitrage and race to the bottom. Regulatory arbitrage refers to strategies firms employ to exploit differences or gaps between regulatory regimes to achieve similar economic benefits while avoiding stricter regulations (Pollman 2019). Exploiting the weaknesses of regulatory sandboxes, arbitrage can be a cascade effect of scope creep. Such arbitrage is particularly amplified within regulatory sandboxes, which inherently adopt a lighter-touch regulatory environment. Historical precedents such as Arizona’s fintech sandbox, which weakened consumer protections,Footnote ³³ illustrate how regulatory arbitrage can encourage jurisdictions to progressively lower regulatory standards, known as a race to the bottom. Larger or more permissive sandboxes may exacerbate this trend because they attract more companies, increasing competitive pressures among jurisdictions. To remain competitive, jurisdictions may feel compelled to further relax oversight, potentially sacrificing consumer protection and regulatory rigor in exchange for economic benefits such as increased investments or job creation.Footnote ³⁴

   Race to the bottom can also be a direct side effect of close collaboration – aka pressure or lobbying – between regulators and firms, favouring undue private and very partial pressure (a phenomenon also known in the literature as regulatory capture). Increased engagement – and thus an elevated risk of regulatory capture – is common in sandbox environments due to frequent regulator-firm interactions. Therefore, transparency is essential to ensure waivers or tailored rules remain justified and not preferentia. Cases like Italy’s sandbox, where no public information has been provided about the measures authorising approved experiments, demonstrate how a lack of transparency can lead to opaque decision-making and potential unfair advantages for participants (Ranchordas and Vinci 2024). Interestingly, broader eligibility scopes can heighten the risk of regulatory capture by incentivising firms to lobby aggressively for favourable regulatory adjustments.

   Under the AIA, Article 58(1)(b) delegates the definition of eligibility criteria to the European Commission’s implementing acts, but Member States retain considerable discretion regarding their application (Article 57). Differences in interpretation and implementation – such as varying entry criteria (defining innovation levels), operational rules, supervisory intensities and intervention thresholds – may encourage providers to engage in “sandbox shopping,” selecting jurisdictions that offer less stringent compliance requirements or more relaxed oversight. Additionally, a particular concern arises from SME prioritisation under Article 62(1), which waives participation fees and simplifies quality-management obligations. While beneficial for innovation, this provision could incentivise Member States to relax oversight to attract startups competitively. Jurisdictions with stricter standards may then face pressure to lower requirements to retain competitiveness.

   Strategy. Although this is a significant challenge, the AIA already incorporates specific safeguards against excessive regulatory arbitrage (e.g., harmonisation and standardisation duties). Indeed, some degree of jurisdictional competition, or sandbox shopping, might be both inevitable and even beneficial, provided it occurs transparently and maintains regulatory balance. Essential safeguards outlined by the AIA include harmonisation efforts through the AI Board and AI Office to ensure minimal EU-wide standards, particularly in safeguarding fundamental rights, alongside periodic reporting, sharing of best practices and comprehensive annual reports. Yet, as harmonisation consistency mechanisms to address such issues, something more could be done, like limiting regulatory derogation through ex-post evaluation sunset clauses, which concretely means that any waivers, exemptions or tailored regulatory conditions (e.g., on data governance) granted within sandboxes would automatically expire after a defined period unless their renewal is explicitly justified through a rigorous post-trial evaluation. The post-trial evaluation should show that the relaxation was necessary for experimentation, did not materially increase risks (especially to fundamental rights) and that adequate safeguards were in place. This might help keep regulatory sandboxes experimental and evidence-based rather than becoming default practices that incentivise regulatory arbitrage or pressure jurisdictions into competitive deregulation.

4. d) Attract “good” companies. A key challenge in regulatory sandbox implementation is ensuring appeal to high-potential companies – those with strong innovation and scalability prospects. Certain sandbox features, such as compliance burdens and time limitations (e.g., participation caps tied to project complexity under Article 58(2)(h) of the AIA), may deter ambitious firms seeking rapid growth or large-scale testing. Additionally, the close regulatory oversight within sandboxes can be perceived as overly restrictive, discouraging participation. This issue has been observed in FinTech sandboxes, where criticism centred on excessive bureaucracy and limited pathways to commercialisation. Many successful startups bypassed formal sandboxes altogether, opting instead for direct market entry or informal testing frameworks. This risk is especially pertinent in the context of the EU’s focus on competition with other economic blocs, as in many cases, organisations will be considering whether to base product development and related trials in the EU or elsewhere.

   Strategy. While the AIA recognises this risk – encouraging simplified procedures for SMEs and startups (Art. 58(2)(g)) – further measures are needed to attract high-growth firms. These could include: (1) demonstrating viable, commercial-scale pathways post-sandbox (e.g., fast-track licensing, market access support) and (2) aligning sandbox incentives with those of successful accelerator programs (e.g., funding opportunities, investor exposure). In the first case, for instance, a structured transition plan agreed at entry could operationalise “fast-track” outcomes not by waiving AI Act requirements, but by reducing duplication – reusing sandbox-generated evidence for conformity assessment, pre-validating testing protocols, and coordinating early with notified bodies or relevant regulators so that the move from experimentation to compliance is smoother. In the second case, the AI Act might interact productively with these incentives by making compliance a market signal: sandbox graduation could culminate in a standardised “compliance readiness” package (documentation, test results, incident reporting plan) that investors and enterprise buyers can evaluate, reducing uncertainty and due diligence friction.

   Further, a proven way to attract strong companies is through structured stakeholder engagement. Although the AI Act references strategic plans (e.g., exit strategies), it does not yet explicitly require a stakeholder engagement plan. Early and systematic dialogue with industry players can clarify the commercial advantages (e.g., investor networks), mitigate concerns about restrictive oversight or intellectual property risks and help sandbox designers align with market demands, track innovation trends and streamline procedures. Crucially, such engagement also equips local regulators with the necessary domain-specific insights to exercise informed discretion in tailoring sandbox operations to their local environments and economies. A virtuous example is, as pointed out by a report of the World Bank Organization, the Bank of Sierra Leone, which revised its sandbox framework after industry feedback to simplify procedures.Footnote ³⁵

5. e) Integration with the existing EU ecosystem. To support the growth of innovative SMEs and startups, it’s essential to connect them with existing EU frameworks. In fact, by providing structured environments that accompany them in the innovation cycle that leads to regulatory sandboxes, different infrastructures and initiatives in the EU ecosystem can offer support that simplifies compliance requirements and allows these innovators to focus on developing their technologies.

   Strategy. AI Factories provide the necessary resources to develop advanced AI models and offer access to cutting-edge technology, significantly reducing the time and cost associated with bringing a product to market. Concurrently, European Digital Innovation Hubs (EDIHs) are the major point of entry and information for users and providers of AI systems, facilitating digital transformation across various sectors by offering training, networking opportunities and access to funding. Finally, Testing Experimentation Facilities (TEFs), leveraging their AI expertise, can provide tailored support for establishing sandbox environments.Footnote ^{36,Footnote 37} To integrate regulatory sandboxes with the EU innovation ecosystem, the different infrastructures should function as a single “innovation-to-sandbox” pathway with clear handovers and shared deliverables. EDIHs would act as the main entry point, conducting a standardised intake and pre-screening to assess the use case, likely AI Act risk profile, and compliance readiness and then referring SMEs to the appropriate support. AI Factories would provide compute and engineering resources to develop and document systems, producing a technical package (e.g., system/model documentation and baseline evaluations) usable for subsequent assessment. TEFs would then generate a conformity-oriented evidence pack through structured testing (e.g., robustness and other relevant performance and risk metrics), which can be directly used in sandbox applications and later compliance processes. (National and supranational) Competent authorities would manage the sandbox itself, defining the legal conditions, safeguards, monitoring and exit criteria based on the technical package and TEF evidence. This division of labour enables TEFs and AI Factories to concentrate on technical development and validation, while sandboxes focus on regulatory supervision, thereby collectively reducing the burden on SMEs and increasing the value of sandbox participation.

6. f) Scope creep. In regulatory sandboxes, the gradual and unplanned expansion of scope is a debated, but mostly theoretical issue.Footnote ³⁸ Borrowed from project management,Footnote ³⁹ the concept of scope creep here describes the uncontrolled broadening of eligibility criteria or deviation from a sandbox’s original objectives. Additionally, regulators themselves might inadvertently or deliberately misuse sandboxes for purposes beyond safe experimentation, undermining compliance standards.

   Under the AIA, scope creep is particularly pertinent due to the Act’s explicit intent to foster inclusivity by welcoming all providers or prospective providers. Although promoting broad participation is beneficial, this openness might inadvertently allow numerous incremental, minimally innovative projects into the sandbox.

   Moreover, scope creep concerns not only entry to the sandbox but also access to specific privileges within it. Particularly problematic is the flexible “public interest” criterion under Article 59(1), which participants may opportunistically leverage to access special sandbox benefits, such as enhanced personal data-use permissions, by framing their projects as fulfilling public interest objectives.

   Strategy. More granular eligibility requirements may be of some help in this context. Just think about the criteria for the level of maturity (or technological readiness). Rather than relying on a single, undifferentiated “maturity” metric, regulators can specify which dimension of readiness they are measuring – manufacturing, commercial or both – tailor the required threshold to the domain so that a life-critical area like healthcare faces stricter expectations than, say, logistics, and evaluate separate layers such as data readiness, operational fit and governance within a transparent multi-dimensional framework.Footnote ⁴⁰ A clear, disclosed methodology of this kind yields more balanced decisions and markedly reduces opportunities for regulatory arbitrage.

   Having said that, a balanced strategy should be taken as an excessive focus on preventing scope creep should not lead to the opposite problem: overly restrictive eligibility criteria could result in cohorts of sandbox participants that are disappointly few in number, and/or non-representative, undermining the generalisability of evidence generated in the sandbox – e.g., about the cost or duration of regulatory adaptation for sandbox’s participants – with other cohorts.Footnote ⁴¹

2. Testing phase challenges and strategies

The testing phase covers governance challenges encountered during active experimentation within the sandbox, specifically regarding integrating risk assessment frameworks, maintaining a realistic testing environment and facilitating effective regulatory learning from sandbox outcomes.

1. a) Integrating the AIA risk assessment model. One key issue involves integrating the AIA’s approach to risk assessment into regulatory sandbox procedures. This integration itself poses challenges due to a fundamental tension: sandboxes are intended precisely for testing and refining AI systems whose risks are uncertain or unknown, yet formal compliance under the AIA demands that risks be systematically identified, classified and mitigated from the outset. As previously noted (Section VI), this implies that providers and sandbox authorities will establish a provisional risk categorisation of the AI system in the mandatory sandbox plan prior to sandbox entry and subsequently adjust this classification based on outcomes observed during sandbox testing, documented explicitly in the exit report.

   Strategy. For this approach to be effective, sandbox authorities must be trained in the AIA’s risk assessment methodologies and capable of applying its classification criteria consistently. Effective governance thus demands coordination among national authorities and with the AI Office, particularly regarding complex or cross-cutting risk factors (such as fundamental rights or cybersecurity). This may be especially challenging in the early stages of the AIA going into force, as local regulators seek to apply the AIA with no previous experience; this challenge should then diminish over time as those regulators gain experience and further guidance and precedents are shared. In this context, the need for standardisation and speed is even greater than in regular (i.e., post-development) risk assessments, as sandbox procedures must accommodate legal oversight within a compressed timeframe.

   It must also be noted that sandbox-specific risk assessments may differ significantly from real-world conditions, requiring potentially extensive adjustments post-sandbox. As a result, after a product exits the sandbox and is placed on the market, a comprehensive reassessment of risk may be necessary, taking into account real-world use and conditions not fully captured within the controlled sandbox environment.

2. b) Realistic testing environment. Poorly designed experiments have low internal validity (uncertainty about whether outcomes result from the intervention or external factors) and external validity (limited generalisability), undermining their policy relevance.Footnote ⁴² To maximise benefits for participants – such as reduced regulatory uncertainty, guidance, market credibility and investor confidence – the internal functioning of a regulatory sandbox must strike a careful balance: it should provide meaningful concessions to foster innovation while still maintaining a realistic regulatory environment where possible. Participants must be able to understand the difference between the sandbox environment and the real world, so they can make realistic plans for full deployment. Otherwise, participants could face severe compliance challenges upon exit, such as GDPR obligations, data localisation requirements or strict consent management rules. Without proper preparation, these regulatory shifts could render their innovations commercially impractical.

   Strategy. To create a realistic regulatory environment, sandboxes could start with minimal, well-calibrated regulatory concessions – enough to encourage innovation – and then, in some cases, gradually reintroduce standard requirements over the testing period. In many cases, the sandbox participant is looking to experiment with one particular proposition or product feature. For example, if a participant seeks only to gauge customer reactions to a specific product feature – such as the rate of adoption of a biometric recognition technology perceived as either convenient or intrusive – the reintroduction of additional regulatory requirements (e.g., compliance with the AIA or GDPR around biometrics) might offer little to no value. In such cases, a brief and focused sandbox trial, tailored precisely to the specific experimental goals, could be more beneficial.

   Equally important is replicating the multi-agency oversight participants will encounter after exiting the sandbox. Access to realistic data is another critical factor. High-quality, real-world datasets – or synthetic alternatives that closely mirror actual conditions – enable accurate customer feedback, user testing and compliance stress-testing within the sandbox. For example, incorporating simulated audits by data protection regulators during sandbox testing can help participants prepare for real compliance assessments.

   To make regulatory sandbox outcomes more realistic and reliable, some scholars have proposed using control groups, as in scientific experiments. This approach compares sandbox participants against firms operating under standard regulatory conditions, providing a real-time benchmark.Footnote ⁴³ Hence, control groups improve internal validity by clarifying whether observed outcomes result from regulatory intervention or external factors. They also strengthen external validity by helping regulators assess whether sandbox-tested innovations can scale under real-world conditions. Yet, implementing control groups remains challenging, especially in emerging fields like AI, where few comparable firms operate under full regulatory compliance.

   Currently, the AIA does not prioritise “experimental realism” in regulatory sandboxes, with a focus on short-term innovation support, particularly for SMEs and startups. This is understandable as an initial goal for a new sandbox regime. A medium-term goal could be to develop more complex and rigorous testbeds that simulate the full regulatory environment. This could improve the long-term scalability and reliability of lessons drawn from sandbox experiments.

3. c) Regulatory learning. Reviewing and integrating insights from regulatory sandboxes constitutes a critical challenge. Regulatory sandboxes frequently fail to realise their full learning potential due to inherent design flaws, such as their casuistic approach and non-representative participant cohorts.Footnote ⁴⁴ Consequently, policymakers – including regulators and other key stakeholders – should prioritise establishing a robust, standardised methodology applicable across different national competent authorities.Footnote ⁴⁵

   Strategy. Effective regulatory learning presupposes the existence of reliable, standardised success and failure metrics, along with clear procedures for translating sandbox outcomes into regulatory action – such as rule modifications or updated guidance. Without such mechanisms, sandboxes risk becoming isolated experiments with no systemic impact.

   In this regard, the AIA provides only limited operational clarity. While Articles 57(7), (8) and (16) require exit reports and annual reports documenting sandbox results and lessons learned, the Act does not establish a harmonised mechanism for integrating these insights into the regulatory framework. Furthermore, the role of institutional bodies is constrained: the AI Office and the Board may access exit reports only with explicit consent (Article 57(8)), and there is no provision for a centralised, automatic review or synthesis of sandbox findings across Member States. This lack of coordination inhibits horizontal learning and cross-jurisdictional regulatory coherence.

   Additionally, although Article 57(16) states that sandbox insights should inform potential revisions to the AIA text (and its implementing acts), the Act lacks a structured mandate for how findings are to be channelled into formal legislative updates, standardisation processes or the development of codes of conduct (Article 58(2)(e)). Likewise, the AIA does not establish a formal mechanism for addressing conflicts or tensions revealed in sandboxes between AI-specific rules and sectoral or fundamental rights regimes (e.g., data protection). No cross-sectoral trigger exists to prompt regulatory coordination or legal reform in response to such conflicts, limiting the sandbox’s capacity to drive systemic regulatory improvement.

   Therefore, the legislator must establish a binding, harmonised mechanism, with specific metrics, to translate sandbox insights into regulatory updates and coordinated cross-sectoral reforms. In the startup phase of sandboxes, sharing use cases and testbeds among national competent authorities will be crucial to support the development of these metrics, especially since technical thresholds are still undefined.

3. Post-testing phase challenges and strategies

The post-testing phase involves three distinct categories of challenges: post-sandbox compliance, which covers regulatory decisions about market readiness and the enforcement of rules after exiting the sandbox; transition challenges, which concern the practical hurdles companies face in moving from sandbox participation to full market deployment and success evaluation, which assesses whether the sandbox itself has effectively balanced innovation, market competition and regulatory learning objectives.

1. a) Post-sandbox compliance. After testing, regulators must decide if the firm can operate under standard or modified rules. This depends on whether the firm met its objectives – e.g., technical feasibility, user uptake or functional performance – and whether the product shows real value at scale. Regulators also assess whether the sandbox trial ran safely, with effective safeguards and no major breaches (e.g., fraud, consumer harm or compliance failures).Footnote ⁴⁶ Another key consideration is whether the firm is prepared for full regulatory compliance post-sandbox (Rathnam 2024), a challenging transition that some navigate by using test results to meet licensing requirements, as seen in the FCA Regulatory Sandbox. Operational viability is also scrutinised, including whether the company has the necessary resources, technology and business model to succeed commercially. Additionally, regulators review the firm’s exit strategy, since sandbox programs, like Singapore’s MAS guidelines, often require participants to outline a clear transition plan upfront. Finally, the decision is informed by detailed reporting from the firm throughout the testing period, covering performance metrics, risk management and compliance. These reports help regulators gauge factors such as consumer adoption rates, adherence to sandbox rules and the innovation’s broader market potential before granting approval for wider deployment.

   Strategy. For AI-specific sandboxes, these foundational criteria should be supplemented with additional requirements. Participants may need to conduct algorithmic audits to detect biased outcomes and demonstrate concrete measures to prevent discrimination alongside standard consumer protection assessments.

   The AIA implicitly accommodates these criteria, though with varying degrees of explicitness. Article 57(7) directly emphasises regulatory compliance readiness, while Article 57(5) implies requirements for operational viability through sandbox plan evaluations. However, effective implementation demands that Member States and EU institutions provide sandbox authorities with sufficient resources – including financial support, technical expertise and specialised training in AI governance. Moreover, to ensure consistency, the Commission should establish clearly measurable graduation criteria and standardised guidance – e.g., via implementing acts – for post-sandbox compliance.

2. b) Success evaluation. Evaluating whether a regulatory sandbox has been successful is crucial for all stakeholders (also for the abovementioned regulatory learning reasons). Regulators must determine if the sandbox effectively balances innovation and control or if alternative approaches (e.g., innovation hubs or forbearance) would be more effective. Sandbox designers and regulators must evaluate if adjustments are required, such as participant recruitment or resource allocation. Participants also need clarity on whether sandbox participation aligns with their goals. However, measuring success presents challenges, particularly in selecting appropriate metrics.

   Strategy. Success criteria should include impact on innovation, market competition, consumer protection and regulatory learning.

   The impact on innovation can be evaluated by examining whether the sandbox facilitated the development of new ideas into viable products or services. Relevant metrics include the number and diversity of innovations, the rate at which sandbox-tested solutions progressed to full market deployment,Footnote ⁴⁷ and the level of venture investment attracted.Footnote ⁴⁸ For example, research has found increased venture funding, enhanced patenting activity and improved survival rates for sandbox participating firmsFootnote ⁴⁹ in the case of the FCA Regulatory Sandbox.

   The sandbox’s ability to promote market entry and competition can be assessed through the speed and cost of obtaining licenses or regulatory authorisation, subsequent access to funding, the conversion rate (i.e., how many participants successfully transition from testing to full market operation) and the emergence of new business models, especially from startups or non-traditional actors. These new entrants often introduce more efficient or productive approaches, contributing to overall sector dynamism.

   Additional dimensions include consumer protection and risk mitigation, which can be assessed through metrics such as the number of safety incidents, consumer complaints or direct feedback on user satisfaction.

   Equally important is regulatory learning. Success in this area can be evaluated through post-sandbox assessments that explore whether regulatory staff gained insights from working with innovative firms (maybe through exit reports),Footnote ⁵⁰ whether the sandbox led to the identification and resolution of regulatory gaps,Footnote ^51, Footnote ⁵² and whether it fostered sustained engagement between regulators and innovators. The frequency and quality of feedback loops, such as workshops or scheduled check-ins, can serve as proxies for this interaction.

   Under the AIA, many of these metrics are explicitly or implicitly justified through its key sandbox objectives, such as improving legal certainty for innovators, facilitating knowledge-sharing across authorities and enabling evidence-based regulatory learning. Notably, Article 57(9)(e) emphasises the support of AI innovation – particularly from SMEs and startups – by encouraging safe experimentation and accelerating market entry.

   Thus, in the context of the AIA, a meaningful success evaluation framework would align with how well each of these objectives is met. When EU Member States implement these AI sandboxes, they should wonder: Did the sandbox lead to clearer regulatory guidance (regulatory learning)? Did it help an AI startup bring a product to market (innovation and market access)? And was this achieved without compromising users’ fundamental rights (consumer protection)? Such questions should also inspire their design.

3. c) Transition challenges. Entering a regulatory sandbox is not an end in itself; the ultimate goal for participants is a successful and sustainable market launch at scale. Consequently, some of the most significant supervisory and governance challenges arise from sandbox experimentation to full market deployment during the transition phase. To address these challenges, sandboxes, including those established under the AIA, typically require clear, structured exit plans. Nevertheless, certain transition challenges remain inevitable, and regulators should proactively guide and prepare participants for these realities.

   Strategy. Firstly, completing sandbox testing successfully does not guarantee immediate market approval. Companies often encounter a compliance “cliff edge” post-testing, meaning they must quickly establish comprehensive compliance measures or fulfill capital requirements previously relaxed within the sandbox. To mitigate this risk, firms should develop detailed plans for the post-sandbox phase well in advance, anticipating compliance demands and potential costs. Secondly, sandbox testing may reveal the need for legal or policy changes, aligning with regulatory learning objectives. However, if these legislative changes are delayed, companies exiting sandboxes might struggle to fully scale operations due to persistent regulatory constraints.Footnote ⁵³ We should not forget that delays in transition can result in losing first-mover advantages, potentially benefiting competitors who are not bound by sandbox limitations or larger firms capable of faster adaptation. Regulators must, therefore, remain agile, committing to prompt policy adaptations and, where necessary, granting temporary regulatory waivers to support innovators until permanent legislative adjustments are in place. Thirdly, companies leaving sandboxes frequently face resource gaps. Sandboxes usually provide structured support such as regulatory guidance, mentorship and industry connections. However, this support typically diminishes or ends after exit. Consequently, firms must independently secure funding and expertise to effectively scale their operations. Proactively addressing these resource needs – including fundraising, hiring compliance personnel and scaling technical systems – is essential for a smooth transition.

   Additionally, supervisory challenges previously highlighted, such as interagency coordination, substantially affect transition ease. Post-testing supervision and governance thus deserve equal attention alongside initial sandbox design.

   Under the AIA specifically, transition challenges may be particularly acute. While benefits for sandbox participants – especially SMEs and start-ups – are substantial during the entry phase (e.g., simplified eligibility criteria), fewer supports are available post-sandbox. Although Article 57(7) of the AIA mentions that conformity assessments might be accelerated based on sandbox exit reports, the vague formulation “accelerating […] to a reasonable extent” may invalidate the effort. Moreover, the AIA sandbox environment primarily evaluates regulatory compliance rather than operational scalability or large-scale performance. While the AIA allows sandbox testing under real-world conditions (Article 57(5)), inherent constraints limit full market replication (and scaling beyond EU borders introduces additional complexity).

   To strengthen the transition process under future implementations of the AIA, several solutions are recommended: (1) operationalise genuinely simplified and accelerated conformity assessments explicitly linked to sandbox exit documentation; (2) establish explicit regulatory feedback loops; (3) broaden conditions for real-world testing and encouraging its use – for example, by introducing provisions for “conditional market access” if real-world testing meets predefined benchmarks, similar to the supportive role played by sandbox exit reports; (4) encourage proactive consumer trust-building initiatives within the sandbox plans and (5) develop funding pathways, including a European AI Innovation Fund, for successful sandbox alumni.