AI liability and operator duties in Spain: 2026 guide.

Spain built the agency first and the statute second, and the statute is not finished. AESIA was created in August 2023, before the EU AI Act itself was adopted, which made Spain the first Member State with a dedicated national AI body. What it did not do was give that body a domestic sanctioning statute, and Spanish administrative law requires one. The Proyecto de Ley Organica para el buen uso y la gobernanza de la inteligencia artificial was approved by the Council of Ministers on 26 May 2026, presented to the Congreso on 28 May, published in the Boletin Oficial de las Cortes Generales as A-97-1 on 12 June, and is in committee with the amendment deadline extended to 2 September 2026. This guide maps what actually binds an operator in Spain today, what AESIA can and cannot do, and what the liability picture looks like when an AI agent causes harm.

AESIA: the first dedicated AI supervisory agency in the EU

When the EU AI Act was published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024, most Member States were still designing their supervisory architecture. Article 70 of Regulation (EU) 2024/1689 required Member States to designate national competent authorities by 2 August 2025. Spain had already moved.

AESIA's statute was approved by Real Decreto 729/2023, de 22 de agosto, published in the BOE on 2 September 2023 under identifier BOE-A-2023-18911 and in force from 3 September 2023. An earlier version of this guide carried the identifier BOE-A-2023-18942, which resolves to an unrelated personnel appointment in a different ministry, and a sourcing note saying the decree could not be confirmed. Both have been corrected: the decree is real and was read at the BOE. Article 2 of the decree places AESIA's institutional seat in the city of A Coruna, in the La Terraza building ceded by the municipal administration. Its current Director is Alberto Gago Fernandez.

The decree provides that the agency effectively begins operating when its Consejo Rector is constituted, within a maximum of three months from entry into force. No June 2024 operational date appears at the BOE or at aesia.digital.gob.es, and the claim has been withdrawn rather than restated.

Here the earlier version of this guide overstated matters in a way that matters to an operator. Real Decreto 729/2023 predates Regulation (EU) 2024/1689 and does not designate AESIA under it. Article 10.1.k of the decree grants supervision and sanction functions only in accordance with what European rules stipulate. No BOE instrument designates Spain's Article 70 national competent authorities, and AESIA's own guides page still refers to potential national competent authorities. The two BOE instruments that actually give AESIA operating competences today are Orden TDF/774/2025 of 11 July 2025 delegating competences to the agency, BOE-A-2025-15133, and a Resolucion of 25 August 2025 on internal delegation, BOE-A-2025-18640.

The sanctioning point is sharper still. La Moncloa's own description of the bill states that under Ley 40/2015 an administrative sanctioning power requires a statute defining the infringements, the penalties and the procedure, and that this bill is what supplies it. Until it is enacted, there is no Spanish procedure by which an AI Act fine can be imposed. Article 99 of the Regulation sets the ceilings; Spain has not yet built the mechanism.

AESIA has published very little beyond the guides. Its Actualidad section carries a handful of undated items, all on Article 50 transparency and the code of practice on AI-generated content. Its Organos Ejecutivos, Normativa and Sandbox pages return an under-construction notice. There are no enforcement decisions and no register. An earlier version of this guide attributed a warnings-first enforcement posture to the Director General and said sanctioning powers had been in place since 2 August 2025; neither could be confirmed at source, and the second is contradicted by the Ley 40/2015 point above. Both have been removed.

None of this makes Article 5 aspirational. The prohibitions are directly applicable EU law and other routes to consequence exist, including the AEPD where personal data is involved and ordinary civil liability. What is absent is a Spanish AI fine.

The EU AI Act as it applies in Spain

Regulation (EU) 2024/1689 is directly applicable across all EU Member States without national transposition. It entered into force on 1 August 2024. Its obligations apply in a phased sequence. Understanding that sequence is the starting point for any operator deploying an AI system in Spain.

What is already in force

Article 5 prohibited practices have applied since 2 February 2025. For an operator in Spain, this means the following are illegal and enforceable by AESIA now. AI systems that use subliminal techniques operating below the threshold of a person's consciousness to materially distort their behaviour in a way that causes or is likely to cause harm. Systems that exploit specific vulnerabilities of a group of persons due to their age, disability, or economic situation to materially distort their behaviour. Systems used by public authorities for social scoring based on social behaviour or personal characteristics leading to detrimental treatment outside the context where data was generated. Real-time remote biometric identification systems in publicly accessible spaces by law enforcement (subject to specific, narrow exceptions requiring judicial authorisation). Emotion recognition systems in workplaces and educational institutions (with narrow exceptions for safety purposes and medical or research purposes). AI systems that create or expand facial recognition databases through untargeted scraping.

Article 4 AI literacy obligations have applied since 2 February 2025. Providers and deployers must ensure that staff who operate AI systems on their behalf have a sufficient level of AI literacy for their tasks. This means deployers are already obliged to have completed basic training or briefing of employees who use AI tools in their day-to-day work. AESIA's guidance on AI literacy sets out a competency framework that can serve as a practical benchmark.

GPAI model obligations under Chapter V have applied since 2 August 2025. If an operator in Spain deploys a system built on a general-purpose AI model (such as a large language model-based AI agent), the provider of that underlying model must have brought it into compliance with GPAI documentation, copyright transparency, and risk management obligations. Deployers do not directly bear GPAI provider obligations, but they rely on the GPAI provider's compliance to operate legally. If the GPAI provider is not compliant, the system the deployer has built on top of it may carry inherited risk.

What applies from 2 August 2026, and what moved

Article 50 transparency obligations applied from 2 August 2026 and were not deferred by the Digital Omnibus. They are in application now. From that date: any deployer operating a chatbot or conversational AI agent must disclose to users, in a clear and timely manner, that they are interacting with an AI system, unless this is obvious from context. AI-generated audio, video, image, or text content that could mislead people must be labelled as AI-generated. Emotion recognition and biometric categorisation systems must notify individuals subject to their processing. A narrow watermarking grace period to 2 December 2026 applies only for machine-readable marking of AI-generated content by systems already on the market before 2 August 2026.

High-risk deployer obligations under Article 26 no longer apply from 2 August 2026. Regulation (EU) 2026/1744 moved the Annex III stand-alone high-risk obligations to 2 December 2027 and the Annex I obligations for high-risk AI embedded in regulated products to 2 August 2028. The Annex III categories are unchanged: AI systems used in the management and operation of critical infrastructure such as electricity, water, transport and gas; in education, vocational training or student assessment; in employment, workers management and access to self-employment; in access to essential private and public services and benefits including credit scoring, insurance risk assessment and social benefit eligibility; in law enforcement and border control; in the administration of justice; and in democratic processes and electoral systems.

Two dates ahead of that. On 2 December 2026 the new Article 5 prohibitions added by the Omnibus take effect, covering AI systems that generate child sexual abuse material or that depict an identifiable person's intimate parts without consent, and the Article 50(2) machine-readable marking transitional period ends. Spain pressed for those prohibitions, with French support. By 2 August 2027 each Member State must operate at least one AI regulatory sandbox, which is a live question for Spain given that its own sandbox has closed.

The Digital Omnibus, as adopted

The Digital Omnibus on AI was proposed on 19 November 2025 as COM(2025) 836, reached political agreement on 7 May 2026, and entered into force on 27 July 2026 as Regulation (EU) 2026/1744, six days before the original high-risk deadline. An earlier version of this guide said the agreement had not been adopted or published and that the 2 August 2026 date therefore still bound, and told operators that relying on the deferral was a regulatory risk. That was wrong by the time it was read, and it is corrected here.

One practical note. AESIA's own guides page states that the sixteen guides will be updated once the digital Omnibus amending the AI Regulation is approved. Read them with that in mind: they were written against the pre-Omnibus timetable.

Spain's draft Organic Law on AI governance

The sequence is longer than the earlier version of this guide suggested. The anteproyecto was approved at first reading by the Council of Ministers on 11 March 2025, together with an agreement authorising urgent processing. The Council of Ministers approved the proyecto de ley organica for transmission to the Congreso on 26 May 2026. It was presented on 28 May, qualified on 8 June, and published in the Boletin Oficial de las Cortes Generales as A-97-1 on 12 June 2026. It sits in the Comision de Economia, Comercio y Transformacion Digital in the amendment phase, with the deadline extended to 2 September 2026. It is not law. An earlier version of this guide stated that the amending law entered into force on 27 July 2026, which conflated it with the EU Digital Omnibus, and told readers that Spain's organic law was in force. It is not.

Its purpose is not to supplement the substantive requirements of the EU AI Act, which is directly applicable, but to provide the national institutional and procedural architecture without which the Regulation cannot be enforced domestically.

The bill's key provisions include the following. Sectors already covered by product regulation keep their existing notifying and market surveillance authorities. For the rest, including employment, biometrics and education, supervision goes principally to AESIA, and also to the Agencia Espanola de Proteccion de Datos and the Consejo General del Poder Judicial, with AESIA as the single point of contact. The CGPJ's inclusion is worth noting because it is named in the official description and was missing from an earlier version of this guide.

The bill implements the penalties regime in Spanish administrative law. It does not simply adopt the Article 99 tiers. It classes infringements as muy graves, graves and leves, with sanctions reaching EUR 35 million or 7 per cent of turnover in the most serious cases and up to EUR 500,000 or 0.5 per cent of turnover in the least serious. That lower band has no equivalent in Article 99 and is a genuinely Spanish addition. The earlier anteproyecto proposed bands of EUR 7.5 million to 35 million or 2 to 7 per cent for prohibited practices, EUR 7.5 million to 15 million or 2 to 3 per cent for very serious high-risk breaches, and EUR 500,000 to 7.5 million or 1 to 2 per cent for serious breaches.

It creates an inventory of AI systems used by the public administration in administrative procedures, and not only for high-risk systems, together with a new statutory figure, the delegado de IA, both to be developed by Real Decreto. It provides legislative grounding for AESIA to operate the Article 57 sandbox, in the future tense. And it establishes formal coordination mechanisms between AESIA and sector authorities. An earlier version of this guide described that coordination as codifying an existing formal cooperation protocol between AESIA and the AEPD; no such protocol could be found at aesia.digital.gob.es, digital.gob.es or aepd.es, and the official description presents inter-authority coordination as something the bill will create.

Parliamentary processing takes several months in Spain, and the bill still has to clear the Congreso and the Senado. The practical implication for operators is precise: the EU AI Act's substantive obligations already apply by direct effect and do not wait for the organic law, but the Spanish machinery for enforcing them does. Until the law passes, Spain has designated authorities in a bill rather than in the BOE, and no domestic AI sanctioning procedure.

The sixteen compliance guides: what they are, and what they are not

Sixteen guides are published at aesia.digital.gob.es, in Spanish and in English, together with a checklists and examples archive. They are detailed and genuinely useful, and they are the most operationally specific national-level material available for the Regulation anywhere in the EU.

Three qualifications. AESIA's own page says they were developed within the Spanish AI regulatory sandbox pilot, and digital.gob.es describes them as technical guides prepared by the Secretaria de Estado de Digitalizacion e Inteligencia Artificial. They are sandbox outputs hosted by AESIA, not AESIA enforcement guidance. No publication date appears on the guides page or in AESIA's news section; an earlier version of this guide dated them to 16 December 2025, which could not be confirmed, and the currently served PDF of Guide 01 carries an internal creation date of 24 July 2026. And AESIA states that they will be updated once the digital Omnibus is approved, so they were written against a timetable that has since changed.

In order, the sixteen are: 01 introduction to the AI Regulation; 02 practical guide with examples; 03 conformity assessment; 04 quality management system; 05 risk management; 06 human oversight; 07 data and data governance; 08 transparency; 09 accuracy; 10 robustness; 11 cybersecurity; 12 records; 13 post-market monitoring; 14 incident management; 15 technical documentation; and 16 a checklist manual for the requirements guides. Note that human oversight is guide 06, not guide 10; an earlier version of this guide had them transposed in two places.

For a deployer in Spain operating an AI agent that falls within an Annex III category, the practical starting sequence is guide 05 on risk management, guide 07 on data governance, guide 08 on transparency, guide 06 on human oversight, and guide 14 on incident management. The checklist manual at guide 16 provides a documentation format built for these requirements.

The regulatory sandbox, which has closed

Spain established a controlled testing environment by Real Decreto 817/2023 of 8 November 2023, BOE-A-2023-22767, published on 9 November 2023. It was made to test compliance with the proposal for the Regulation, before the AI Act existed, and it was run by the Secretaria de Estado de Digitalizacion e Inteligencia Artificial rather than by AESIA. The bill now before Congress would provide that the Article 57 national sandbox will be operated by AESIA, in the future tense.

There was one call, opened by a SEDIA resolution of 20 December 2024, running from 23 December 2024 to 23 January 2025 and extended to 30 January 2025. It selected up to twelve high-risk AI systems. On 1 July 2026 digital.gob.es recorded that after several months of work the Spanish AI sandbox has finished, and published its conclusions on a results page.

This matters practically. An earlier version of this guide said the sandbox was in its third cohort, had processed more than twenty systems, and that operators uncertain about classification should apply to AESIA. There is no open sandbox to apply to. Article 57 requires each Member State to operate at least one AI regulatory sandbox, and the Omnibus moved that deadline to 2 August 2027; Spain's next sandbox is therefore a forthcoming obligation rather than a current facility. The guides and the published sandbox results are what remains usable from the pilot.

AESIA and AEPD: the dual-regime reality

Most AI systems deployed in Spain process personal data. This creates a dual-regime situation that operators must address structurally. AESIA supervises the AI Act obligations. AEPD supervises GDPR compliance and the national organic law on data protection (Ley Orgánica 3/2018, LOPDGDD). The two authorities operate under a formal cooperation protocol.

The practical implications for a deployer are the following. A Fundamental Rights Impact Assessment required by Article 27 of the AI Act for Annex III systems used in employment, credit, education, or essential services is not the same as a Data Protection Impact Assessment required by Article 35 GDPR, although they cover overlapping ground. AESIA expects the FRIA; AEPD expects the DPIA. An operator should conduct both and cross-reference them. AEPD has published updated recommendations specifically on AI-based voice transcription and automated decision-making under GDPR, which provide a practical framework for the GDPR side of a dual-regime assessment.

Where an AI system's outputs constitute a solely automated decision that produces legal or similarly significant effects on data subjects, Article 22 GDPR applies alongside Article 26 of the AI Act. AEPD enforces Article 22 rights (the right to human review, the right to an explanation, and the right to contest). AESIA enforces the AI Act's transparency and human oversight obligations. Both require documentation. The safest approach is a unified governance structure with a single record that addresses both regulators' expectations, rather than two siloed compliance programmes.

Civil liability when an AI agent causes harm in Spain

The EU AI Act establishes a public regulatory regime with administrative penalties. It does not itself create a private right of action for individuals harmed by AI systems. Civil liability for AI agent harm in Spain operates through three legal channels.

Spanish tort law under the Codigo Civil

Articles 1902 to 1910 of the Spanish Codigo Civil establish the general framework for extracontractual civil liability. The foundational rule under Article 1902 requires a person who causes damage to another through fault or negligence to repair that damage. Applied to AI agent harm, this means a claimant must establish: that the AI system produced an output or took an action that caused identifiable harm; that the operator of the system was at fault (failed to take reasonable precautions), or negligent in their deployment; and causation between the deployment and the harm.

The opacity of complex AI systems makes the causation and fault elements difficult to establish in practice. Spanish courts, consistent with other European jurisdictions, have not yet developed a substantial body of case law specifically addressing AI agent liability. The general principles of fault-based tort apply, but claimants face practical evidentiary challenges. The Mata v. Avianca case (Southern District of New York, 2023), while not a Spanish case, illustrates the legal consequences of AI-generated output causing harm in a professional context: the court imposed sanctions on the lawyer who filed the AI-generated brief containing fabricated citations without verification. Spanish courts would likely reach an analogous result under general professional liability principles. The Moffatt v. Air Canada case (British Columbia Civil Resolution Tribunal, 2024) illustrates the principle that an operator cannot disclaim liability for their AI agent's statements to customers: the deployer, not the AI system, bears the legal obligation. Spanish consumer law and general civil liability principles support the same conclusion.

Product liability under the LGDCU and the revised PLD

Spain's existing product liability framework sits in the Real Decreto Legislativo 1/2007 approving the consolidated text of the Ley General para la Defensa de los Consumidores y Usuarios, which carries the transposition of the original Product Liability Directive 85/374/EEC. An earlier version of this guide attributed this to Ley 3/2022; Ley 3/2022 of 24 February 2022 is the Ley de convivencia universitaria, a university conduct statute, and the citation was wrong. Under the existing framework, software and AI systems have generally not been treated as products for product liability purposes in Spain, leaving harm from AI outputs within the fault-based tort framework.

This will change on 9 December 2026, when Directive (EU) 2024/2853 (the revised Product Liability Directive) must be transposed into Spanish law. The revised PLD expressly includes software and AI systems within the definition of product. It applies to products placed on the market or put into service after 9 December 2026. Under the revised PLD, a Spanish claimant will benefit from rebuttable presumptions of defectiveness: where the defendant has not complied with EU-law safety obligations (which include the EU AI Act), or where the technical complexity of the system makes it excessively difficult for the claimant to prove the defect, the burden of proof shifts toward the defendant. The revised PLD also expands the categories of compensable damage to include data loss and medically recognised psychological harm, alongside existing categories of physical and property damage.

The practical consequence for operators placing AI agents on the Spanish market after 9 December 2026 is a materially more demanding civil liability landscape. Non-compliance with the AI Act not only risks administrative penalties from AESIA: it also creates a rebuttable presumption of product defect in civil litigation. The two regimes reinforce each other, and compliant documentation built for AESIA purposes (risk management under Article 9, technical documentation under Articles 16 and 17, incident records under Articles 12 and 73) is simultaneously the evidence base for defending product liability claims.

Contractual liability in B2B deployments

Many AI agents are deployed by a business operator as part of a service delivered to business clients. In these B2B contexts, the contractual framework between the operator and its client allocates liability risk as between the two parties. Spanish contract law under the Codigo Civil and the Codigo de Comercio governs these allocations. Standard AI vendor contracts typically attempt to limit the vendor's liability for AI outputs, exclude consequential loss, and require the deployer customer to use the system in accordance with instructions. These provisions are enforceable between commercial parties under Spanish law (unlike B2C contexts, where consumer protection rules limit exclusion clauses under the LGDCU). Operators on both sides of a B2B AI contract should review their liability allocation in light of the AI Act's deployer obligations: an operator who breaches Article 26 deployer duties cannot simply pass liability to the AI provider where the breach arises from the deployer's own conduct.

Insurance for AI operators in Spain

The commercial insurance market for AI agent liability in Europe is at an early and developing stage. Spanish operators face the same market structure as their counterparts in other EU Member States. No Spanish-specific AI liability product exists as a domestic policy class. Spanish-domiciled operators access the international AI insurance market through the same channels as other European operators.

The instruments available include Munich Re aiSure (performance-based cover, settling on measurable performance data rather than fault-based claims); Armilla, a Lloyd's of London coverholder offering AI liability and performance cover; HSB (Hartford Steam Boiler, Munich Re group), which has offered affirmative AI and algorithmic-risk cover to SMEs; and cover structured around the AIUC-1 standard, which sets out adversarial evaluation benchmarks across data and privacy, safety, security, reliability, accountability, and societal impact. The ElevenLabs policy announced on 12 February 2026 was the first coverage instance backed by the AIUC-1 standard for AI voice agents.

EIOPA's Opinion on Artificial Intelligence governance and risk management (EIOPA-BoS-25-360, published 6 August 2025) is addressed to national competent authorities in the insurance sector, including Spain's DGSFP (Dirección General de Seguros y Fondos de Pensiones). The Opinion does not create new insurance coverage requirements for policyholders but signals that European insurance supervisors expect insurers to address AI governance risk in their own underwriting frameworks. The practical implication is that insurers writing AI liability risks in Spain are likely to increase their documentation requirements for AI governance evidence as the market matures.

For operators preparing to approach the market for AI liability cover, the documentation built for AESIA compliance purposes is the core of the underwriting submission. The risk classification under Annex III, the Article 9 risk management record, the Article 14 human oversight mechanism, and the Article 73 incident reporting register are the primary evidence an underwriter will seek to assess the risk.

What operators deploying AI in Spain should do now

The regulatory environment in Spain is more defined than in most EU Member States because AESIA is operational and has published comprehensive guidance. The compliance pathway for operators is well-mapped.

The first step is classification. Every AI system the operator deploys should be assessed against Article 5, which prohibits outright, Annex III for high-risk with the full deployer obligations, Article 50 for transparency, and the GPAI provisions where the system is built on a general-purpose model. Guide 02, the practical guide with examples, is the most useful starting point.

The second step is to deal with what is enforceable now rather than what is deferred. Article 50 transparency obligations have applied since 2 August 2026: any chatbot or conversational agent must clearly and timely inform users that they are interacting with an AI system, and synthetic audio, image, video and certain published text must be marked. The Article 50(2) machine-readable marking transitional period for systems already on the market ends on 2 December 2026.

For Annex III systems, build the Article 26 posture against 2 December 2027 rather than treating the deferral as a reason to stop. The core elements are unchanged: a record of the provider's instructions for use and confirmation that the system is being used in accordance with them; a Fundamental Rights Impact Assessment under Article 27 where the use context requires one; a named human oversight arrangement under Article 14 with documented competence and intervention authority; an incident monitoring and reporting procedure under Article 73; and the Article 12 logging record where it is within the operator's technical control.

Do not plan on the sandbox. Spain's pilot has closed, and the Article 57 obligation to operate at least one now runs to 2 August 2027. For a genuinely borderline classification, the published sandbox results and the guides are what is available in the meantime.

For operators building commercial insurance into their risk management approach, the coverage gap between existing professional indemnity and cyber policies and AI-specific liability exposure should be mapped against the Annex III classification of the specific system. AI-specific endorsements and the emerging dedicated AI liability products from carriers including Munich Re, Armilla, and HSB provide the clearest route to affirmative coverage, with ISO/IEC 42001 certification increasingly used as an underwriting benchmark for governance maturity.

Frequently asked questions