AI agents in a business: what the AI Act and the GDPR already say about software that acts

The AI Act does not define an AI agent. The European Commission's AI Act Service Desk says the term is used inconsistently, but that there is broad agreement an agent has the ability to receive and process input from its environment and to execute actions, based on that processing, that may interact with or affect its environment, such as issuing function calls. Its answer to how the Act treats agents is that they are not a separate category: the definitions of an AI system in Article 3(1) and of a general-purpose AI model in Article 3(63) are sufficient to cover them, so the rules for AI systems and models apply to agents as well.

The definition's own words cover software that acts. Article 3(1) describes an AI system as designed to operate with varying levels of autonomy, generating outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. The Commission's guidelines on that definition say a recommendation that is applied automatically becomes a decision, that a decision implies a fully automated process producing an outcome without any human intervention, and that virtual environments include digital spaces, data flows and software ecosystems (paragraphs 57, 58 and 60). On that reading, an agent that changes records in other software through function calls produces outputs of that kind.

The Act as amended in July 2026 mentions agentic AI once. Annex XIV, the list of codes that sets the scope of a notified body's designation, includes AI systems based on other emerging AI technologies, including agentic AI. It sets no rules of its own for agents. Whether the Act applies to your business at all is covered in the AI Act scope guide.

A business that uses an AI system under its own authority is its deployer (Article 3(4)). A business that buys an agent and runs it as supplied is the deployer, and whoever placed it on the market under their own name is the provider. A business that develops an AI system, or has one developed, and puts it into service under its own name is the provider (Article 3(3)), and putting into service includes supplying a system for its own use (Article 3(11)). The Act calls the provider of a system built on an AI model a downstream provider, whether the model is its own or another company's (Article 3(68)). A business that builds its own agent on another company's model can therefore be the provider of the agent, while the model's maker is the provider of the model.

Which one you are decides who owes several of the duties below. The provider or deployer guide works through the line, including how pointing a general-purpose system at a high-risk use makes the business that does it the provider.

AI literacy, Article 4, has applied to providers and deployers of any AI system, agents included, since 2 February 2025. Since 27 July 2026 it asks for measures to support the development of AI literacy, and it says it does not require anyone to guarantee a specific level of AI literacy for any individual. What counts as a measure is set out separately.

The prohibitions in Article 5 have applied since 2 February 2025. The Commission's Service Desk names the bans on harmful manipulation and on exploiting vulnerabilities, Article 5(1)(a) and (b), as particularly relevant to agents, and says that complying with them may require safeguards in the way agents are designed and developed.

Transparency, Article 50, has applied since 2 August 2026. An AI system intended to interact directly with people must be designed so that they are told they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person, and they must be told at the latest at the first interaction (Article 50(1) and (5)). That design duty is the provider's: the company that placed the agent on the market if you bought it, or yours if you built it and put it into service under your own name. The Article 50 guide sets out the disclosures that fall on the deployer, for deep fakes and for AI-generated text published on matters of public interest, and the wording that satisfies each.

The deployer's duties in Article 26 apply only where the agent is a high-risk AI system: from 2 December 2027 for the uses listed in Annex III, and from 2 August 2028 for AI in products covered by the laws listed in Annex I (Article 113, as amended). Annex III includes recruiting and selecting people, decisions on terms of work, promotion and termination, allocating tasks by individual behaviour or personal traits, monitoring and evaluating workers, credit scoring other than to detect financial fraud, and risk assessment and pricing in life and health insurance. An agent whose purpose is not listed in Annex III, and which is not part of a product covered by Annex I, is outside the high-risk rules.

Where Article 26 does apply, it includes human oversight by people with the necessary competence, training and authority, monitoring the system, and keeping the logs it generates (Article 26(2), (5) and (6)). What starts when has the full list and the calendar.

Article 22 of the GDPR, in force since 25 May 2018, gives a person the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects concerning them or similarly significantly affects them. Such a decision is allowed only where it is necessary for a contract with the person, authorised by law with suitable safeguards, or based on their explicit consent (Article 22(2)). In Ireland, section 57 of the Data Protection Act 2018 uses the second route only for decisions that an enactment authorises or requires. The Article 22 guide covers the exceptions, their safeguards and what counts as solely automated.

The Data Protection Commission's AI insights report, published on 25 September 2026, describes an AI system whose decisions could result in a user being banned from a service. The controller argued that Article 22 was met because another user triggered the system and there was an appeal with human review. The DPC considered that the decision fell within the prohibition, and its takeaway is that a human triggering an AI system, or an appeal with human intervention, does not necessarily ensure compliance (page 59). The controller changed the process so that a person was involved before any ban.

An agent that reads a mailbox, a customer record or a booking system is processing personal data, and the GDPR applies as it does to any other processing. The business that decides why and how the agent is used is the controller (Article 4(7)). A supplier that runs the agent or its model and processes the data on the business's behalf is a processor (Article 4(8)). It needs a contract with the terms in Article 28(3), including that it acts only on the business's documented instructions, and it may engage other processors only with the business's written authorisation (Article 28(2) and 28(3)(a)). Controller or processor sets out the test, and the processing agreement checklist goes through the terms.

The DPC's guidance on AI and large language models, of 18 July 2024, says an organisation using an AI product that relies on personal data could be a data controller, and that if so a formal risk assessment should be considered. Before using the system, it says, the organisation should understand what personal data it uses, where the data goes when a third party is involved, and whether the provider keeps or reuses it. The guidance asks whether the business could give a person access to their data held within the AI system, or delete it, and warns that relying on outputs without critical human analysis or intervention may introduce automated decision-making risks.

The AI insights report records two controllers that intended to launch AI agents working by capturing screenshots of a virtual browser. The DPC recommended more transparency, warnings tailored to the use of agents, and more clarity about the risks, including that personal data may be captured (page 42). It had engaged with five controllers on agentic AI launches in the EU, and says that, beyond a continuing lack of transparency and clarity on how the AI processes personal data, common themes were not yet clear (page 60).

Regulation (EU) 2026/1744, the Digital Omnibus on AI, is dated 8 July 2026, was published in the Official Journal on 24 July 2026 and came into force on 27 July 2026. It reworded Article 4, and moved the Annex III high-risk rules to 2 December 2027 and the Annex I rules to 2 August 2028. Article 50 still applies from 2 August 2026. The Omnibus also added two prohibitions to Article 5 and gave providers of generative systems already on the market more time to mark their output; the AI Act dates guide has when each starts. The Commission's AI Act page, last updated on 19 October 2026, gives the same dates.

The Digital Omnibus has a second part: a proposed regulation on data, cybersecurity and privacy rules which, among other laws, would amend the GDPR. The Commission's digital rulebook page, last updated on 27 July 2026, describes its changes as proposed revisions, and no act amending the GDPR has been published in the Official Journal. Until one is adopted and in force, the GDPR applies as described above.

The checks, in order: