AI agents and the processing of personal data: the keys to the new AEPD Guidelines

The Spanish Data Protection Agency (hereinafter, "AEPD") has recently published a document with guidelines to be considered by all data controllers and processors who use AI agents to carry out any kind of processing of personal data.

First of all, we must answer the following question: what is an AI agent? According to the AEPD Guide, it is an artificial intelligence system that uses language models ("LLMs") to fulfil an objective, i.e. these agents act appropriately according to their circumstances and objectives, are flexible in the face of changing environments and goals, and learn from experience and decision-making.

In other words, these agents are not just systems that only analyse information or generate responses, but can plan, execute actions, interact with multiple sources, store persistent memory and act with a significant degree of operational autonomy. In this context, the challenge is to realise the importance that their implementation requires conscious use and clear governance in processes and internal policies.

In view of the above, here are some relevant aspects to consider in case you use or plan to use AI agents in your data processing:

• Quality of the data used for training: First of all, the agent must be configured and fed with quality, coherent and updated contextual information, especially in its memory and context retrieval mechanisms, in order to achieve the most correct results possible. Otherwise, the agentic AI could start with small errors or hallucinations that little by little begin to divert the purpose of its use, thus causing damage to those responsible and in charge.
• Protection by design and by default: In consideration of the state of the art, its cost, nature, context, purposes and existing risks, the controller must implement appropriate technical and organisational measures both at the time of determining the means of processing and in the processing itself. Thus, the IA agent should be designed in such a way that it only uses the data that are strictly necessary, uses them exclusively for the purposes for which they were collected and protects this information throughout its life cycle. In addition, the level of autonomy of each agent in its dealings with the environment must be defined.
• Memory control and autonomy: The agent's memory allows it to accumulate information about the context, learn patterns and thus carry out autonomous work over time that will have results in accordance with the information collected. Thus, it is necessary to implement measures and policies to avoid excessive storage of unnecessary data.
• Review of external parties with which it is related: The use of the AI agent usually involves more external actors such as model providers, third-party APIs, cloud services, among others. For this reason, the controller must assess the functioning of the AI agent, the providers it uses and the technical and organisational measures determined in order to mitigate risks.
• Risk assessment and governance of the processing: The incorporation of AI agents may alter the context, purposes, means and risks of the processing, so the determination of responsibilities (controller/processor), the basis of legitimacy, the Register of Processing Activities (RAT) and, where appropriate, an Impact Assessment (EIPD) should be reviewed, especially if there are automated decisions in the terms of Article 22 of the GDPR. It should also be integrated within the internal governance and under the supervision of the DPO, avoiding uncontrolled deployments ("BYOAgentic").

For more information, we recommend you review the document here.