Five layers of EU law reach an AI agent that acts. The EU AI Act regulates it by risk. GDPR governs the personal data it touches. DORA and NIS2 cover its operational resilience and security. Sector rules, such as PSD2 for payments, govern the actions it takes. Several supervisors enforce them; one set of controls can evidence most.
What does each law require, in plain words?
The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024, with obligations phasing in. It sorts AI by risk. Some practices are prohibited, from 2 February 2025. High-risk systems, such as credit scoring of individuals and pricing of life and health insurance listed in Annex III, carry duties on risk management (Article 9), data governance (Article 10), automatic logs (Article 12), information to deployers (Article 13), human oversight (Article 14) and accuracy, robustness and cybersecurity (Article 15). Deployers have their own duties under Article 26, and some must run a fundamental rights impact assessment under Article 27. Article 50 requires telling people when they interact with an AI system. Article 4 requires AI literacy among staff.
GDPR applies to every agent handling personal data: the Article 5 principles, a lawful basis under Article 6, limits on solely automated decisions under Article 22, records of processing under Article 30 and impact assessments under Article 35.
DORA (Regulation (EU) 2022/2554), applying from 17 January 2025, requires financial entities to manage ICT risk, report major ICT incidents, test resilience and manage ICT third-party providers.
NIS2 sets cybersecurity risk management and incident reporting duties for essential and important entities across critical sectors.
Sector and payment law. PSD2 requires strong customer authentication for payments; the Instant Payments Regulation requires payee name verification. Consumer credit, insurance distribution and investment rules govern what agents may tell and sell to customers.
Who supervises what?
| Area | Supervisor |
|---|---|
| AI Act compliance | National market surveillance authorities; for regulated financial firms, usually linked to the financial supervisor; the Commission for general-purpose AI models |
| Personal data | National data protection authorities, such as the CNIL, AEPD, Garante or the Data Protection Commission |
| Prudential and ICT risk | ECB banking supervision for significant banks; national authorities such as BaFin, ACPR, De Nederlandsche Bank, Banca d'Italia, Banco de España or the Central Bank of Ireland |
| Conduct and markets | National authorities such as the AFM, with EBA, ESMA and EIOPA setting common standards |
| Anti-money laundering | National authorities, with the EU AML Authority being established in Frankfurt |
The practical effect is that a single agent can be examined from three or four directions. Each examiner asks different questions about the same system.
What does this mean when the system is an agent that acts?
Agents raise one issue the laws treat separately: an agent both assesses and acts. The AI Act looks at the assessment, through its use case and risk class. GDPR looks at the data and any decision about a person. DORA looks at the ICT risk of the action and the platform. Payment law looks at the action itself. A credit agent that declines an application touches all four at once.
The second issue is classification by use, not design. Whether an agent is high-risk under the AI Act depends on what it is used for. A general support agent that one day starts recommending credit limits has changed its legal category without any code change.
How do you build one set of controls?
The efficient approach is to design controls around the agent and map each to every law it evidences. Our guide to designing an AI agent architecture for a regulated business covers the design.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| AI Act Art. 9; DORA ICT risk; NIS2 risk management | One risk assessment per agent capability, covering harm, security and resilience | Risk register by capability |
| AI Act Art. 10; GDPR Art. 5 | Sources approved per capability; data minimised | Source register |
| AI Act Art. 12; GDPR Art. 30; DORA | Append-only log of every action and decision | Log store with retention policy |
| AI Act Art. 14; GDPR Art. 22 | Named person can hold, override or stop each capability | Hold records; stop tests |
| AI Act Art. 15; DORA testing | Pre-release and periodic testing on real cases, including adversarial inputs | Test reports |
| AI Act Art. 50 | Disclosure that the customer is talking to an AI system | Disclosure text; logs |
| DORA third-party risk; GDPR processors | Platform contract, register entry and exit plan | Contract; register |
| PSD2 and payee verification | Payment actions use authenticated, verified instructions only | Authentication records |
| DORA and NIS2 incident reporting | Agent incidents classified against both regimes | Incident records |
What will EU supervisors ask?
- Which agents do you run, what does each do, and is any high-risk under the AI Act?
- Are you provider, deployer or both for each?
- Who can intervene in each agent, and how?
- Where are the logs, how long are they kept, and how do you reconcile that with GDPR?
- Is the agent platform in your DORA register of ICT third-party providers?
- Which incidents involving agents have you classified, and under which regime?
What do these laws not cover?
They do not license agents individually or require approval before deployment, outside certain high-risk conformity steps. They do not settle every member state detail; national law adds to GDPR and transposes NIS2. Application dates for some AI Act high-risk obligations have been subject to proposed changes, so check the current timetable. For the UK view, compare FCA expectations for AI agents in regulated firms; for a market without an AI law, see how AI agents are regulated in South Africa. For the full map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does the AI Act apply to a firm outside the EU?
It can. It applies to providers placing AI systems on the EU market and to deployers in the EU, and it can reach providers and deployers outside the EU where the system's output is used in the EU.
Is an agent built on a general-purpose model itself a general-purpose AI model?
Usually not. The agent is an AI system built on the model. The model provider carries the general-purpose model duties; the agent's provider and deployer carry system duties based on its use.
Do DORA and NIS2 both apply to a bank's agents?
DORA is the specific regime for financial entities and takes precedence over NIS2 on ICT risk and incident reporting for them. NIS2 matters more for non-financial essential entities and some service providers.