How to comply with the AI Act: obligations, costs and the political battle over regulating AI

What it is and who it applies to

The AI Act (Regulation (EU) 2024/1689) regulates artificial intelligence systems according to the risk they pose, not the technology they use. It does not only apply to European companies: it binds any provider that sells or puts into service an AI system in the EU, and also applies when the system's output is used in the EU, even if the company is based elsewhere.

The first thing a company needs to know is its role, because obligations change a great deal depending on it:

• Provider: whoever develops the system or model, or has it developed, and markets it under its own name. It bears most of the obligations.
• Deployer: the company or body that uses an AI system in its activity, for example a bank using a scoring model bought from someone else.
• Importer and distributor: those who bring it into the EU or resell it. They must check that the provider has complied.
• Watch out for a change of role: whoever puts their brand on a high-risk system, substantially modifies it or uses it for a different high-risk purpose becomes a provider (Article 25).

Four risk levels and general-purpose models

• Prohibited (Article 5). Manipulation that distorts behaviour, exploitation of vulnerabilities, social scoring, individual predictive policing based solely on profiling, facial recognition databases built by untargeted scraping, emotion recognition at work and in education, biometric categorisation by sensitive data and real-time remote biometric identification in public spaces by police, save for narrow exceptions. The 2026 reform adds systems that generate non-consensual intimate images or child sexual abuse material.
• High risk (Article 6). AI that is part of already regulated products (machinery, medical devices, toys...) and AI used in the Annex III areas: biometrics, critical infrastructure, education, employment and worker management, access to essential services such as credit or insurance, law enforcement, migration and borders, and justice and democratic processes.
• Transparency (Article 50). Chatbots must disclose that they are machines; AI-generated content must be marked in a machine-readable way, and deepfakes labelled.
• Minimal risk. Everything else, which is most uses (spam filters, recommenders, internal tools), has no specific obligations.

An Annex III system can fall outside high risk if it poses no significant risk: for example, if it only performs a narrow procedural task or prepares an assessment later made by a person. But the exception never applies if the system profiles people, and the provider must document why it applies it and register the system.

General-purpose models, such as large language models, have their own rules: technical documentation, information for those building on them, a copyright policy and a public summary of training data. If trained with more than 10²⁵ floating-point operations, they are presumed to pose systemic risk and must also assess and mitigate that risk, run adversarial testing, report serious incidents and protect their cybersecurity. A voluntary code of practice, published in July 2025, serves to demonstrate compliance; Meta announced at the time that it would not sign it.

The timeline, after the 2026 reform

1. 1 Aug 2024The AI Act enters into force.
2. 2 Feb 2025Prohibitions apply.
3. 2 Aug 2025Obligations for general-purpose models.
4. 27 Jul 2026The reform (Regulation (EU) 2026/1744) enters into force, postponing high-risk obligations.
5. 2 Aug 2026The Commission can demand information, evaluate models and fine model providers. Transparency obligations apply.
6. 2 Dec 2026New prohibitions (non-consensual intimate images and child abuse) and end of the grace period for marking content from systems already on the market.
7. 2 Dec 2027Annex III high-risk obligations (previously August 2026).
8. 2 Aug 2028High-risk obligations for AI built into regulated products (previously August 2027).

How to comply, step by step

For most companies, which use AI bought from others, the work is mostly inventory, classification and control. The heavy technical obligations fall on those who develop high-risk systems. This is a sensible order:

1. Inventory. List every AI system the organisation uses or sells, including those embedded in third-party software and those employees use on their own.
2. Classification. For each use, decide whether it is prohibited, high-risk, subject to transparency or minimal risk. Document the reasoning, especially if you rely on the Annex III exception.
3. Prohibited uses out. Remove any prohibited use: the bans have applied since February 2025 and carry the highest fines.
4. Role and contracts. Determine whether you are provider or deployer in each case, and require your suppliers to provide documentation, instructions for use and a contractual compliance commitment.
5. If you develop high-risk AI. You need a risk management system across the whole life cycle (Article 9), governance of training data including bias analysis (Article 10), technical documentation per Annex IV (Article 11), automatic event logging (Article 12), clear instructions (Article 13), design for human oversight (Article 14), accuracy, robustness and cybersecurity (Article 15) and a quality management system (Article 17).
6. Conformity assessment and registration. Most Annex III systems are self-assessed through internal control; biometric ones may need a notified body. Then an EU declaration of conformity, CE marking and registration in the EU database.
7. If you use high-risk AI. Use it according to the instructions, assign oversight to trained people, monitor input data, keep logs for at least six months, inform workers and their representatives before using it at work and tell affected people about its decisions. Public bodies and those assessing creditworthiness or pricing life and health insurance must also carry out a fundamental rights impact assessment (Article 27).
8. Transparency. Review chatbots, text, image or voice generators and any synthetic content: user notice, technical marking and labels on deepfakes.
9. Ongoing monitoring. High-risk providers must monitor the system after launch and report serious incidents to the authority no later than 15 days after becoming aware of them.

Fines are tiered by severity: up to 35 million euros or 7% of worldwide turnover for prohibited uses; up to 15 million or 3% for breaching other obligations; and up to 7.5 million or 1% for supplying incorrect information. The higher figure applies, except for SMEs, where the lower one does. Providers of general-purpose models face fines of up to 15 million or 3%.

Problems and advantages

The main problem is uncertainty. The harmonised technical standards, which allow a presumption that a system complies, were not finished in time, and several countries were slow to designate their supervisory authorities. Those two reasons explain the 2026 postponement. Companies notice it themselves: among those that considered using AI but did not, unclear legal consequences rose from 46% in 2023 to 54% in 2025 and are now the second barrier, behind only a lack of expertise.

• Problems: compliance costs for SMEs and start-ups, overlap with data protection rules, the Digital Services Act and sectoral rules, the risk that each country applies it differently, and a shifting timeline that penalises those who had already invested to meet 2026.
• Advantages: one rulebook for 27 markets instead of 27 different laws, rights protection already working for prohibited uses, more trust from customers and public bodies, and a standard that can serve as a reference for other jurisdictions and for large companies selling worldwide.

The rise in uncertainty is widespread: it affects 17 of the 18 countries with data for 2023 and 2025. The Eurostat survey does not distinguish between laws, so not all of it can be attributed to the AI Act.

The political battle: what was said at the UN

The UN General Assembly, whose general debate began on 22 September 2026, staged three ways of seeing AI regulation.

Secretary-General António Guterres devoted a central part of his last address to the Assembly to AI. He asked the countries leading the technology to cooperate on evaluating systems, share information on risks and agree common safeguards, drawing on the Independent International Scientific Panel on AI, which presented its first report in July, and the Global Dialogue on AI Governance, which met in Geneva that same month.

The United States went the other way. President Donald Trump told the Assembly that his country rejects any attempt to build a “globalist scheme” to control AI, and that it will keep leading what he called superintelligence. It is the same line his science adviser, Michael Kratsios, took at the Security Council a year earlier when rejecting “all efforts” by international bodies to govern AI.

The paradox came the next day, 23 September, at the Security Council: the chief executives of OpenAI, Sam Altman, and Anthropic, Dario Amodei, asked for international standards. Altman proposed common rules to measure capabilities, assess risks and preserve human oversight. Amodei suggested narrow global agreements, such as banning the use of AI to make biological weapons, systems to verify commitments between countries and a common incident notification mechanism.

The EU spoke on 24 September through European Council President António Costa, who called for AI governance that is “responsible, transparent and regulated” and a central role for the UN. China, for its part, has created its own forum: in July 2026 it launched the World Artificial Intelligence Cooperation Organization.

The outcome is awkward for Europe. It has the only general AI law in force and the UN looks in its direction, but the country home to most of the most advanced models rejects any international framework, and the EU itself has just postponed the most demanding part of its law. Europe has chosen to regulate; now it has to show it can do so without falling further behind.

Recent cases

• First requests for information. Since 2 August 2026 the Commission can act against model providers. At the end of August it sent its first requests for information to more than 30 providers, along two lines: the safety of the most advanced models, and copyright and transparency about training data. According to Executive Vice-President Henna Virkkunen, they targeted companies that had not taken part in informal dialogues with the AI Office.
• Grok and sexual deepfakes. In January 2026, the Commission opened proceedings against X over the Grok feature that allowed users to create fake sexual images of real people, including minors. It did so under the Digital Services Act, not the AI Act, and in February the Irish data protection authority opened its own investigation. The AI Act reform adopted months later expressly bans such systems from December 2026.
• The postponement itself. The 2026 reform is an enforcement case in its own right: the EU acknowledged that neither technical standards nor national authorities were ready and pushed back the key date for high-risk AI by 16 months.
• Beware of the “first fines”. Figures circulate online for AI Act fines allegedly imposed in August 2026 on recruitment, credit or retail companies. We have not found any confirmed by the Commission or a national authority, and some are impossible: the high-risk obligations they cite do not apply until December 2027.

For a Spanish company, the reference authority is the Spanish Agency for the Supervision of Artificial Intelligence (AESIA), based in A Coruña. The practical advice is not to wait for 2027: the inventory, the classification and the supplier clauses can be done today, and they are the hardest things to improvise.

Read the full theme: AI and digital