
An AI agent is given an objective, plans the steps and uses tools (email, ERP, search engines, APIs) to act on it; an assistant only answers. In an SME it makes sense for narrow tasks, with minimum permissions, reversible actions and human oversight. Regulation (EU) 2024/1689 applies to it with the timetable amended by Regulation (EU) 2026/1744.
Agent or assistant: what is the difference?
The Spanish Data Protection Agency (AEPD) published in February 2026 its guidance on agentic AI from a data protection perspective (version 1.2). It defines an AI agent as a system that uses language models to achieve an objective: it breaks complex tasks down into subtasks, carries them out in a planned way with various tools and perceives its environment by accessing internal and external services. Among the characteristics it lists are autonomy, action beyond generating text, proactivity and planning.
| Assistant | Rule-based automation | Agent | |
|---|---|---|---|
| Who decides the next step | The person, at each query | The rule written in advance | The model, within the limits set for it |
| What it does | Answers, summarises, drafts | Runs a fixed flow | Chains steps and uses tools with real effects |
| Typical error | Wrong answer | Badly written rule | A wrong action carried out without anyone seeing it |
Regulation (EU) 2024/1689 has no "agent" category. It defines an AI system as a system designed to operate with varying levels of autonomy (Article 3(1)), and an agent is an AI system. Annex XIV, added in 2026, mentions "agentic AI" only as a technology category for designating notified bodies. The obligations depend on what the system does, not on the name it is given.
Plausible use cases in an SME
They have in common a narrow objective, limited data and a person who reviews before anything has irreversible effects.
| Case | What the agent can do | What it should not do alone | Main risk |
|---|---|---|---|
| Customer service | Classify queries, look up the order, draft the reply and propose it; answer low-risk frequently asked questions on its own | Promise refunds or reimbursements, change customer data, respond to complaints | Invented answers; showing another customer’s data |
| Back office | Read invoices, match them to orders, flag differences, prepare the accounting entry and collection reminders | Pay, send communications, delete records | A reading error that spreads into the accounts |
| Document search | Answer with references to internal procedures, contracts or rules, according to each person’s permissions | Show documents to people who do not have access; answer without giving the source | Leaks through badly configured permissions |
| IT | Classify tickets, summarise alerts, propose a fix, run routine low-risk tasks | Change configurations in production, create users with privileges, delete data without approval | Excessive permissions; instructions hidden in a ticket or an email |
A gestoría (administrative and tax advisory firm) in Barcelona could start with the back office; a hotel in Andorra, with classifying booking enquiries and drafting replies that reception reviews. In customer service there is one specific obligation: Article 50(1) of the Regulation requires that people interacting with an AI system are told so, unless it is obvious from the context.
Risks to put on the table
Permissions. An agent acts with the credentials it is given. NIST (CAISI, request for information of 8 January 2026) describes these systems as capable of planning and acting autonomously on real systems, and points out that its research has shown risks of agent hijacking, when the agent processes data containing malicious instructions. The practices it cites are least privilege and a zero-trust architecture.
Prompt injection. The AEPD describes prompt injection, direct or indirect, as a threat that can make the agent ignore the organisation’s guidelines.
Irreversible actions. Permanently deleting data, sending communications and making payments are the AEPD’s examples of actions that require prior human approval.
Data. The AEPD speaks of "shadow leaks": the silent, gradual leaking of sensitive information through apparently innocent queries, and it asks that the agent’s memory be controlled.
Errors that accumulate. The AEPD explains that an agent’s accuracy drops as the task requires more steps, because an initial error (a badly made query, incomplete data) is processed as if it were correct. The same instruction can give different results, which makes verification and auditing harder.
The rule of two. The AEPD sets out a minimum threshold of safeguards: an agent should not combine all of the following three things at once. (1) It receives uncontrolled information. (2) It accesses sensitive information without restrictions. (3) It can execute automatic actions with real effects. Its example is an agent that answers email from any origin, with access to the whole system and no control over what it sends: it should not be allowed. With two of the three you can work under conditions: if there is uncontrolled information and sensitive data, no automatic action without human oversight; if there are sensitive data and actions, integrity safeguards; if there is uncontrolled information and actions, no access to personal data.
Human oversight: how to do it properly
The AEPD recommends defining, from the design stage, control points where human approval is required: high-impact actions, irreversible actions, atypical behaviour and limits set by the user, such as an amount above which a purchase must be approved. Whoever reviews must have competence, authority, training, independence and resources, and watch for automation bias. For high-impact processes it suggests the four-eyes principle: double verification by different people.
Our criterion for an SME is to work with three levels:
- The agent acts alone: low-impact, reversible actions, with a log.
- The agent proposes and a person approves: anything that leaves the company, moves money or changes customer data.
- The agent never does it: deleting, changing permissions or deciding about people.
There is also a legal basis. Article 22 of the GDPR and Article 25 of Andorra’s Llei 29/2021 on personal data protection recognise the right not to be subject to a decision based solely on automated processing with legal effects, with a right to human intervention where it is permitted. And the AI Act asks for effective human oversight in high-risk systems (Article 14), for those who carry it out to have competence, training and authority (Article 26(2)), and for the logs to be kept for at least six months (Article 26(6)).
The framework of Regulation (EU) 2024/1689 as of 25 October 2026
The Regulation distinguishes between whoever provides an AI system and whoever uses it under their own authority, the deployer (Article 3), which is the role of most companies. It also applies to providers and deployers in a third country when the output is used in the Union (Article 2(1)(c)). The timetable, as amended by Regulation (EU) 2026/1744 (OJ of 24 July 2026, in force on the 27th):
| Date | What applies |
|---|---|
| 1 August 2024 | Regulation enters into force |
| 2 February 2025 | Chapters I and II: general provisions, including AI literacy (Article 4), and prohibited practices |
| 2 August 2026 | General date of application, including the transparency obligations of Article 50 |
| 2 December 2026 | Two new prohibited practices on sexual material generated or manipulated with AI (Article 5(1), points (ba) and (bb)) and end of the deadline for marking synthetic content (Article 50(2)) for systems already on the market |
| 2 December 2027 | Obligations for the high-risk systems of Annex III |
| 2 August 2028 | Obligations for the high-risk systems of Annex I |
Article 4 has been reworded: it now asks providers and deployers to take measures to support the AI literacy of their staff, without requiring a specific level from any person. Annex III includes, for example, systems for recruiting or selecting staff (point 4(a)) and for assessing the creditworthiness of natural persons (point 5(b)). An agent that screens CVs falls into this group and would have obligations from 2 December 2027; an agent that answers questions about internal procedures, in principle, would not. This classification is a general reading; each case needs legal advice.
Where we have an interest, and where we don’t
To be clear: MIA is a BDR Group product, so we have an interest in this topic. According to the approved service page, MIA is a private corporate AI assistant that lets you search, summarise, compare and find answers about documents, procedures and internal data in natural language, with access by team, role or area, traceability of queries and deployment in a private cloud, a dedicated environment or on-premises. The page does not describe it as an agent that executes actions, and this article does not claim that it does.
What to do now
- List the AI tools already in use in the company and the actions they can take (read, write, send, pay).
- Choose a narrow, low-risk case and define in writing the objective, the data it sees and the actions it is allowed.
- Give it minimum permissions, with its own account, and check it against the rule of two.
- Decide which actions need approval and who approves them, and switch on logging of what it does.
- Tell the people who interact with it that AI is in use, and measure the results before extending it.
Frequently asked questions
Is an AI agent the same as a chatbot?
No. A chatbot converses and answers; an agent can also chain steps and use tools to do things in real systems. The boundary is gradual and the AEPD speaks of it as a scale of greys.
Is it legal for an agent to act without supervision?
It depends on what it does. No rule prohibits it in general, but if it takes decisions with legal effects on people, the GDPR and Llei 29/2021 recognise the right to human intervention, and if the system is high-risk, the AI Act requires effective human oversight.
Are agents banned or regulated by the AI Act?
They are neither banned nor covered by a rule of their own. They are AI systems and are subject to the Regulation according to what they do: transparency when they interact with people and high-risk obligations if they are used in the cases of Annex III.
Does an Andorran company have to comply with the Regulation?
If the system’s output is used in the Union, the Regulation also applies to providers and deployers in a third country (Article 2(1)(c)). For purely internal use in Andorra, the data framework is Llei 29/2021.
Sources
- Agencia Española de Protección de Datos. Inteligencia artificial agéntica desde la perspectiva de protección de datos, v1.2, February 2026 (in Spanish). Accessed: 25 October 2026. https://www.aepd.es/guias/orientaciones-ia-agentica.pdf
- NIST, Center for AI Standards and Innovation. Request for Information Regarding Security Considerations for Artificial Intelligence Agents. Federal Register, vol. 91, no. 5, 8 January 2026. Accessed: 25 October 2026. https://www.govinfo.gov/content/pkg/FR-2026-01-08/html/2026-00206.htm
- European Parliament and Council. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 2, 3, 4, 14, 26, 50 and 113 and Annex III. OJ L, 2024/1689, 12 July 2024. Accessed: 25 October 2026. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689
- European Parliament and Council. Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), Articles 1, 3 and 4 and Annex XIV. OJ L, 2026/1744, 24 July 2026. Accessed: 25 October 2026. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744
- European Parliament and Council. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22. OJ L 119, 4 May 2016. Accessed: 25 October 2026. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679
- General Council of Andorra. Llei 29/2021, del 28 d’octubre, qualificada de protecció de dades personals (Qualified Law 29/2021 of 28 October on personal data protection; in Catalan), Article 25. BOPA no. 119, 17 November 2021. Accessed: 25 October 2026. https://www.apda.ad/assets/pdf/CGL20211115_08_58_32.pdf
Want to know what AI can do with your documentation?
If you want to see how a company’s internal information can be queried with a private AI and access control, start with the product page: MIA, the private corporate AI assistant. There are more details at mia360.ad.



