AI regulatory sandboxes have moved from policy theory to active practice faster than most observers expected. By 2026, more than 60 sandbox programs related to AI, data, and technology exist worldwide, spanning jurisdictions from Spain to Singapore to Texas. For businesses building or deploying AI systems, understanding how these controlled testing environments work, and where they operate, has become a practical compliance question rather than an abstract regulatory one.
The EU AI Act sits at the center of much of this activity. Its sandbox provisions set a formal framework that member states are now implementing at different speeds, while countries outside the EU have developed their own parallel models. This guide maps the current state of national AI regulatory sandboxes, explains what happens inside them, and outlines what participation means for businesses navigating AI regulation in the EU and beyond.
How AI regulatory sandboxes work in practice
An AI regulatory sandbox is a controlled framework set up by a competent authority that allows providers or prospective providers of AI systems to develop, train, validate, and test an innovative AI system under regulatory supervision for a limited time. The EU AI Act formalizes this definition, and most national programs follow a similar logic: a company applies to enter, agrees to a sandbox plan, and operates under close oversight from the relevant authority during the testing period.
The core exchange is straightforward. A participating organization receives a degree of regulatory flexibility, including protection from administrative fines for violations that occur during the sandbox period, provided it follows the guidance of the national authority in good faith. In return, the authority gains direct insight into how a real AI system behaves, which informs future regulation. Liability toward third parties remains in place throughout; sandbox participation does not shield a company from legal claims arising from harm caused during testing.
What the process looks like from entry to exit
Most sandbox programs run for six months to two years, depending on the complexity of the AI system and the regulatory body involved. Participants typically move through a structured sequence: an application and selection phase, an active testing and supervision phase, and a closure phase that produces an exit report. Under the EU AI Act, that exit report carries real compliance weight. Providers can use it as supporting documentation in conformity assessment procedures, and market surveillance authorities are required to take it positively into account.
The EU AI Act also permits the use of personal data within a sandbox under specific conditions. Data must serve a substantial public interest, be necessary for the AI system’s development, be kept in a separate and protected environment, and be deleted once participation ends. A summary of the project must be published on the competent authority’s website, unless the work involves sensitive law enforcement data. These conditions make the sandbox a genuinely regulated space, not a loophole.
Countries that have launched national AI sandboxes
Spain was the first EU member state to operationalize a national AI regulatory sandbox, established under Royal Decree 817/2023 and administered by the national AI supervisory authority AESIA. By 2026, the program has processed more than 20 AI systems across multiple cohorts, covering high-risk applications in healthcare, biometrics, employment, and critical infrastructure. Spain’s early start has made it a reference point for other EU members still building their own programs.
Across the EU, progress has been uneven. As of mid-2025, only Spain had a fully operational sandbox, with a small number of other member states actively implementing their programs and the majority yet to communicate concrete plans. The EU Digital Omnibus agreement, finalized in mid-2026, extended the deadline for all member states to establish at least one national AI regulatory sandbox from August 2026 to August 2027, acknowledging the implementation gap while keeping the obligation in place.
Active programs outside the EU
Singapore launched its Global AI Assurance Sandbox in July 2025, overseen by the Infocomm Media Development Authority and the AI Verify Foundation. The program covers emerging risk categories including agentic AI, data leakage, and vulnerability to prompt injection. In a separate initiative, Singapore and Google ran a joint AI Agents Sandbox starting in August 2025, testing computer-use agents across quality assurance, social assistance, and AI safety use cases over approximately four months.
Norway has operated its Datatilsynet AI sandbox since 2020, one of the earliest in the world, and by early 2024 had entered its fifth round with a focus on generative AI. Brazil’s National Data Protection Authority runs a sandbox focused on algorithmic transparency and compliance with the country’s General Data Protection Law, with three AI companies selected in October 2025. In the United States, Utah and Texas have both enacted state-level sandbox legislation, while the federal SANDBOX Act introduced in September 2025 proposed a cross-agency program allowing companies to apply for regulatory waivers of up to two years.
Industries and use cases approved inside sandboxes
Healthcare, financial services, and public administration consistently appear as priority sectors across national sandbox programs. These are domains where AI systems carry significant risk, where regulatory requirements are dense, and where the potential gains from supervised testing are highest. Spain’s first cohort covered six sectors: essential services, biometrics, employment, critical infrastructure, machinery, and healthcare products.
The UK’s MHRA AI Airlock pilot, which ran through the 2024 to 2025 financial year, focused specifically on standalone AI medical devices. One notable outcome from the UK’s broader healthcare sandbox activity was the approval of an AI-powered virtual physiotherapy clinic for musculoskeletal conditions, where digital physiotherapists conduct video consultations and create personalized exercise plans. The program received roughly 40 applications, with around 44% coming from micro-SMEs with fewer than 10 employees.
Emerging use cases in 2025 and 2026
Singapore’s AI Agents Sandbox produced findings that extended the conversation beyond traditional regulated industries. Testing computer-use agents in real-world settings revealed strong potential for automation in citizen services while surfacing concrete risks in oversight, cybersecurity, and privacy governance. These results are shaping how regulators think about agentic AI systems, which behave more autonomously than earlier generations of AI tools.
France’s data protection authority CNIL worked with France Travail in 2024 to assess how a generative AI tool functioned within its employment services platform. Brazil’s sandbox focuses on algorithmic transparency in data-driven decision systems. Delaware’s state sandbox, signed into law in 2025, targets corporate governance, biotechnology, healthcare, chemicals, and finance. The pattern across all these programs is consistent: sandboxes concentrate on areas where the stakes of a failed deployment fall on individuals or critical systems.
What businesses gain from sandbox participation
Sandbox participation offers businesses four concrete advantages: regulatory certainty before a product launches, reduced time to market, direct access to regulators, and documentation that supports compliance processes. For companies building AI systems in regulated industries, the ability to test under supervision and receive written confirmation of that testing from a national authority is a meaningful asset during conformity assessment.
Under the EU AI Act, access to national AI regulatory sandboxes is free of charge for SMEs and startups, and member states are required to give these businesses priority access. This provision matters because smaller companies typically carry a disproportionate compliance burden relative to their resources. The exit report produced at the end of a sandbox program can accelerate conformity assessment procedures, reducing the time and cost of getting a product to market.
Competitive and investment signals
Sandbox participation also functions as a credibility signal. Evidence from the UK’s FCA fintech sandbox, which predates AI-specific programs, showed that companies completing the program received substantially more investment than peers who did not participate. While comparable figures for AI-specific sandboxes are not yet available, the dynamic is similar: demonstrating that a product has been tested under regulatory supervision signals to investors, customers, and partners that the organization takes compliance seriously.
Spain’s AESIA published more than 16 practical compliance guides in December 2025, developed directly from insights gathered during its sandbox pilot. These guides are now used as reference material by compliance professionals across the EU. For companies that participated in developing those insights, the reputational and knowledge advantages extend well beyond the sandbox period itself.
For businesses that want to stay visible as AI reshapes how customers discover products and services, producing well-structured, authoritative content on topics like AI regulation is a practical growth strategy. Scaling content output with the right hybrid approach helps companies build topical authority in areas that matter to their audiences.
Limitations and criticisms of the sandbox model
The empirical record on whether AI regulatory sandboxes achieve their stated goals remains thin. Critics, including analysts writing in The Regulatory Review, argue that the model can create an appearance of regulatory legitimacy without adequately addressing systemic risk. The concern is not that sandboxes are harmful, but that they may be less effective than their framing suggests.
Several structural tensions run through most sandbox programs. Participation does not suspend liability for harm caused to third parties. For high-risk applications in healthcare or employment, this raises genuine questions about the appropriate limits of supervised testing when the consequences of a failed experiment fall on individuals rather than on the company doing the testing.
Access and consistency challenges
Although EU sandboxes are free for SMEs, the compliance infrastructure required to apply and participate is not. Smaller firms may lack the legal and technical expertise to navigate the application process effectively, which risks concentrating sandbox access among companies with more resources. This would undermine the EU AI Act’s explicit goal of supporting SME participation.
Inconsistency across jurisdictions creates a separate problem. Differences in entry criteria, operational rules, and supervisory intensity between national programs create conditions for regulatory arbitrage, where providers select the sandbox that offers the most lenient requirements. Within the EU, fragmented enforcement capacity means some national authorities will have more resources and expertise than others, potentially producing uneven outcomes across the single market. The EU Digital Omnibus introduced an EU-level sandbox to complement national programs, which may help reduce this fragmentation over time.
How the global sandbox landscape is evolving
The global sandbox landscape is moving in two directions simultaneously: more programs are launching, and existing programs are becoming more interconnected. The Datasphere Initiative identified more than 60 AI, data, and technology sandboxes worldwide by early 2025, with 23 additional countries planning new programs specifically for AI. The OECD held a dedicated knowledge-sharing webinar in November 2025, bringing together government officials and regulators from seven countries, reflecting growing institutional investment in cross-border coordination.
The EU Digital Omnibus agreement, finalized in mid-2026, introduced an EU-level regulatory sandbox to sit alongside national programs and extended real-world testing rights outside sandboxes. The 2025 EU AI Continent Action Plan also committed over €220 million across a five-year period to EU-funded Testing and Experimentation Facilities, which will provide supervised testing infrastructure accessible through the European Digital Innovation Hub network.
Novel models emerging in 2025 and 2026
The UK’s AI Growth Lab, announced in October 2025, represents a distinct evolution in sandbox design. Rather than testing AI systems, it proposes testing legal interpretations of existing laws as applied to AI. The idea is that if a regulator’s interpretation of existing law proves too restrictive for AI deployment, that evidence can inform subsequent legislation. This approach treats the regulatory framework itself as the object of experimentation, not just the technology.
Africa had established 25 national sandboxes across 15 countries by late 2024, driven partly by the need to develop governance frameworks ahead of formal AI legislation. Latin America is following a similar trajectory. Japan, South Korea, and Singapore lean toward experimentation-driven models that emphasize real-world testing over prescriptive rules. The Future of Privacy Forum notes that these regional approaches reflect different institutional starting points rather than different end goals: all are trying to build trustworthy AI governance frameworks, but the paths vary considerably. For businesses operating across multiple jurisdictions, tracking these divergent models is becoming a core part of AI compliance strategy.
This content was generated with the help of AI and it may contain mistakes
