What did Spain's AI sandbox reveal about high-risk compliance?
Eleven of 12 selected systems completed a five-phase regulatory pilot. The official report identifies documentation and governance as major implementation challenges, but the purposive sample cannot estimate readiness across Europe.
The short version
- What happened
- Spain selected 12 high-risk AI systems and 11 completed the pilot, which ran through five structured phases over more than a year.
- Who it affects
- Applicants, patients and claimants need contestable decisions and effective human oversight, not documentation alone.
- What remains unproven
- The purposively selected group of 12 systems is too small and non-random for an EU-wide compliance estimate.
Keep in mind
The purposively selected group of 12 systems is too small and non-random for an EU-wide compliance estimate.
The ministry led and evaluated the pilot, so the report is not an independent assessment.
Aggregate progress during a supported pilot does not establish sustained real-world compliance after exit.
Source analysis · 2 linked sources. Primary analysed source date: 28 September 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
The pilot and its denominator
Spain's digital-transformation ministry has published the final results of its first artificial-intelligence regulatory sandbox. The 111-page report gives September 2026 as its publication month, while the ministry's results and download pages were updated on 28 September. The English edition describes a government-led pilot designed to test practical preparation for the EU AI Act's requirements for high-risk systems. It is an official evaluation, not a peer-reviewed research paper or an independent audit of the ministry.
Twelve systems were selected through a public process and 11 completed the pilot. They covered medical devices, biometrics, employment, access to services, machinery and critical infrastructure. The process began in April 2025 and used five phases: training and interpretation of technical guides, self-diagnosis and an adaptation plan, implementation, self-assessment, and post-market monitoring. A 40-person multidisciplinary advisory group and market-surveillance authorities contributed technical, legal and supervisory perspectives.[1][2]
What the report says was difficult
The ministry's central conclusion is that participants generally found it harder to translate obligations into internal procedures, documentary evidence and working governance than to understand the rules in principle. The report says effort was associated more with system complexity and organisational maturity than with company size alone. It also highlights post-market monitoring: designing a monitoring concept was not enough unless it could be incorporated into routine operational and organisational processes.
Those observations are useful because compliance work often disappears behind a final certificate or policy statement. A provider may understand that human oversight is required yet still need to decide who receives an alert, what information that person sees, when they can stop the system and how the intervention is recorded. Similar operational questions apply to data governance, incident reporting and technical documentation. The report's value lies in exposing that translation work, not in declaring the participating systems safe.[1]
What organisations can use now
A developer or public buyer can adapt the pilot's sequence without assuming it guarantees legal conformity. Start with one defined high-risk use, identify the provider and deployer responsibilities, and map each requirement to a named process, record and accountable owner. Test whether the evidence can actually be produced: a policy that says performance will be monitored is weaker than a dated log showing the metric, threshold, review and response. Any unresolved question should remain visible rather than being converted into a completed checkbox.
For people affected by a high-risk system, the operational details determine whether rights work in practice. A job applicant, benefit claimant or patient needs a route to understand and contest a consequential decision. Staff need enough time, authority and information to provide meaningful oversight. A sandbox can help organisations discover where these arrangements fail before wider deployment, but affected people were not a statistical sample in this pilot and their outcomes should be evaluated directly in later work.
Limits and what would change our assessment
The 12 systems were purposively selected for maturity, innovation and sector diversity. Although the report describes the group as representative of high-risk providers in the Spanish ecosystem, it is not a probability sample from which a compliance rate can be estimated. One system did not complete, and the published aggregate cannot tell readers how every organisation would perform without intensive government, expert and supervisory support. Participants also knew they were in a structured pilot, which may change effort and documentation.
The next evidence should follow systems after the sandbox. Our assessment would strengthen if independent audits found that identified controls remained in use, serious incidents were detected and handled, affected people could exercise rights, and smaller organisations could meet requirements at a sustainable cost. It would weaken if documentation improved while operational behaviour did not, or if support costs made the approach inaccessible outside a small selected group. Other EU countries can learn from Spain's methods, but must test them against their own authorities, sectors and administrative capacity.