Launching an artificial intelligence startup in Spain means complying, from day one, with part of the AI Act, the GDPR and Spanish intellectual property law. The prohibited practices have applied since 2 February 2025 and the Article 50 transparency obligations since 2 August 2026; the high-risk obligations arrive on 2 December 2027. In this guide we cover what role your startup plays under the AI Act, which obligations already apply to you, how to make sure the code belongs to the company, which contracts you need and what concessions the Regulation gives start-ups, with a practical example and a checklist.
If you need lawyers for your tech startup, request a free initial assessment.
The AI Act, Regulation (EU) 2024/1689, allocates obligations by role, not by size. Before looking at anything else, you need to know yours:
A startup that builds its product on the API of a large language model is usually both: deployer of the third-party model and provider of the system it sells to its customers. And Article 25 turns into a provider anyone who puts their name or trade mark on a high-risk system already on the market or substantially modifies it.
Two parts of the Regulation already apply to any startup, whatever the risk of its product.
Prohibited practices, since 2 February 2025. Article 5 prohibits, among others, subliminal or manipulative techniques that cause significant harm, exploiting vulnerabilities due to age or disability, social scoring, and emotion recognition in the workplace and in education, except for medical or safety reasons. This is the highest tier of fines in the Regulation. We explain it in our article on prohibited AI practices.
Transparency, since 2 August 2026. Article 50 requires the provider to design the system so that people know they are interacting with an AI, unless this is obvious, and to mark the audio, image, video or text it generates in a machine-readable format. The deployer must, among other things, inform people when it uses emotion recognition or biometric categorisation and disclose that content is a deep fake. There is some leeway: systems that generate synthetic content and were placed on the market before 2 August 2026 have until 2 December 2026 to comply with the marking in Article 50(2), under the Digital Omnibus on AI, Regulation (EU) 2026/1744. The detail is in our article on AI transparency under Article 50.
A system is high-risk if it is a safety component of a regulated product (Annex I) or if it is used in one of the areas listed in Annex III: employment and recruitment, access to education, creditworthiness assessment, essential services and biometrics, among others. Many HR, edtech and fintech startups fall here without realising it.
After the Omnibus, these obligations apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems. They include risk management, data governance, technical documentation, human oversight and a quality management system (Article 17). That system applies in proportion to the provider’s size, in particular for SMEs and start-ups, but without lowering the rigour the Regulation requires. Since 14 October 2026 there has also been a harmonised standard for building it, which we explain in our article on the EN 18286 AI Act harmonised standard.
The GDPR applies from your first user, and in an AI startup it comes up at three points:
How the two sets of rules fit together is explained in our article on the AI Act and the GDPR.
It is the first thing an investor looks at, and Spanish intellectual property law does not always settle it in the company’s favour:
On top of this come the open-source licences of the libraries the product uses, the terms of use of the third-party model, which often restrict commercial uses or sectors, and registering the trade mark before investing in it. How to put code ownership in order is explained in our article on software assignments from founders, freelancers and agencies.
An AI startup that sells software as a service needs, at a minimum, terms of service with a service level agreement (SLA), a data processing agreement for its customers and a policy on using customer data to improve the model. Since 12 September 2025, the EU Data Act has also required data processing services to set out switching between providers in writing. We explain it on our SaaS contract lawyer page.
On the supplier side, check who is liable if the third-party model fails or changes, what the supplier does with the data you send it and whether its contract lets you meet your own transparency obligations.
The Regulation requires each Member State to have at least one national AI regulatory sandbox operational from 2 August 2026 (Article 57), and to give priority access to SMEs and start-ups with a registered office or branch in the Union that meet the eligibility conditions (Article 62). For fines, SMEs and start-ups are subject to the lower of the fixed amount and the percentage of turnover (Article 99(6)).
This example is fictitious. Atenda is a Madrid startup with twelve employees that sells a customer service assistant built on the API of a large language model. Atenda is the deployer of that model and the provider of the assistant it markets under its own brand. Because the assistant talks to its customers’ clients, it must tell them they are dealing with an AI, and because it generates text, Atenda checks with its supplier how to comply with synthetic content marking before 2 December 2026. The assistant does not decide on credit or employment, so it is not high-risk. Atenda signs data processing agreements with each customer, checks that the model’s terms allow commercial use and discovers that part of the code was written by a freelancer without an assignment agreement: it puts this right before its pre-seed round.
The fines the Regulation provides for in each case are explained in our article on AI Act penalties.
Yes. Using a third-party model makes you at least a deployer, and if you market the system you build on it under your own name or trade mark, also the provider of that system. The obligations depend on that role and on the risk of the use, not on the size of the company.
The prohibited practices since 2 February 2025; the Article 50 transparency obligations since 2 August 2026, with until 2 December 2026 for synthetic content marking in systems already on the market; and the high-risk obligations from 2 December 2027 for Annex III and 2 August 2028 for Annex I.
Fines for prohibited practices can reach EUR 35 million or 7% of worldwide annual turnover for the previous financial year. For SMEs, including start-ups, each fine is set at whichever of the amount or the percentage is lower (Article 99(6)).
Only with a valid legal basis under the GDPR and for a purpose compatible with the one that justified collecting the data. You also need to inform users and, if the processing is high-risk, carry out an impact assessment. It is best decided before training, because undoing a training run is much harder than planning it.
With contracts. What salaried employees write in the course of their duties belongs to the company unless otherwise agreed (Article 97(4) of the Spanish Intellectual Property Act). What founders write before incorporating, or what freelancers or agencies write, needs a written assignment that specifies the rights, forms of exploitation, term and territory.
None of these obligations requires a large legal team, but all of them are easier to meet when designing the product than when correcting it later, with customers or investors waiting. At Innovatech we advise tech startups on the AI Act, data protection, contracts and the intellectual property of their product, as part of our tech startup legal services. Tell us about your case and we will give you a free initial assessment.
Marta Suárez-Mansilla is Managing Partner of Innovatech Legal and a Spanish lawyer (abogada), Madrid Bar (ICAM), working in technology law. She completed Harvard Law School's Copyright course and BerkeleyX's Blockchain programme, and has advised technology companies for more than eight years.
