Menu
© 2026 World Compliance Association

03/03/2026

Agent-based AI and data protection: the AEPD shifts the debate from the model to the system

Agent-based AI and data protection: the AEPD shifts the debate from the model to the system

By Iván Martínez, CEO of Intedya

In February 2026, the Spanish Data Protection Agency published a document that marks a turning point in the way we must analyse artificial intelligence from a legal perspective: ‘Agentic Artificial Intelligence from a data protection perspective’

This is not a guide to generative AI in the broad sense. Nor is it a technical manual on architecture. It is something more significant: a structural warning. The risk no longer lies solely in the model, but in the architecture that orchestrates models, memory, tools, APIs and autonomy. The AEPD is shifting its focus: from the algorithm to the agentic system.

From the reactive LLM to the autonomous agent

In the first sections of the document (Chapters II and III), the Agency clearly defines what it means by an AI agent: a system that uses language models to achieve objectives through task decomposition, planning and autonomous execution. The key difference from a traditional LLM is threefold:

  1. Operational autonomy: the agent decides how to subdivide tasks and which tools to invoke.

  2. Persistent memory: it can store both short- and long-term context.
  3. Capacity for action: it does not merely generate text; it carries out actions on internal or external systems.

This last point is decisive. When a system can act on behalf of the user or the organisation, it ceases to be an assistant and becomes an automated operator with the capacity to have legal implications; the question is no longer simply whether it ‘generates content’, but whether it makes decisions, modifies data, executes contracts or alters operational workflows.

The real risk: architecture and attack surface

Chapter IV introduces one of the document’s most significant concepts: the expansion of the attack surface. Agent-based AI is not an isolated component; it integrates:

  • LLMs (in-house or external),

  • databases,

  • internal tools,

  • web services,

  • structured and unstructured memory,

  • multi-agent orchestration mechanisms.

The AEPD expressly warns against adopting a ‘merely additive’ view of vulnerabilities. The interaction between components generates multiplicative effects; in other words, the risk is not linear, but systemic.

Particularly critical is the concept of SPOC (single point of compromise): a breach of one component can compromise multiple data processing operations simultaneously. In environments where a single LLM-as-a-service processes prompts from various data processing operations, the system can become a node for the mass concentration of personal data.

Memory: the most sensitive frontier

The section on memory is probably the most strategic in the document. The Agency distinguishes between:

  • Working memory (short- and long-term),

  • Management memory (logs, traceability, activity records).

And it raises a critical issue: compartmentalisation. The document illustrates two architectural extremes:

  • A single repository where data from multiple processing operations converge.

  • Granular compartmentalisation by processing operation, case and user.

The difference is not technical; it is legal. A design lacking effective compartmentalisation may breach:

  • The principle of purpose limitation.

  • The principle of data minimisation.

  • The safeguards regarding the separation of processing operations.

Furthermore, persistent storage facilitates the creation of longitudinal profiles, even where the original processing did not provide for this. The AEPD conveys an implicit message here: storage is not neutral. It is a structural source of risk.

Autonomy and Article 22 of the GDPR

In its analysis of autonomy, the Agency sets out different levels of human-agent interaction:

  • The agent proposes.

  • The agent collaborates.

  • The agent operates with approval.

  • The agent acts under passive supervision.

The central question is whether certain actions constitute automated decisions with legal or substantially similar effects (Article 22 of the GDPR).

The AEPD does not provide definitive answers, but it does offer operational criteria:

  • Is the action reversible?

  • Is there effective human oversight?

  • Is there transparency and traceability?

  • Is the chain of reasoning reproducible?

The document also warns against automation bias – the human tendency to trust technological decisions due to their apparent consistency. Symbolic human oversight is not sufficient. It must be effective, informed and have a genuine capacity to intervene.

Non-reproducibility: a break with the classical control paradigm

One of the most sophisticated sections is that devoted to non-repeatable behaviour. In complex agent-based systems:

  • minor variations in the environment or latency can alter the chain of reasoning,

  • infinite loops may be generated,

  • compounding errors may occur,

  • inference may not be deterministic.

This undermines a classic premise of technological compliance: same input – same output. Without reproducibility, the following are weakened:

  • auditability,

  • traceability,

  • accountability,

  • ex ante validation.

From a legal perspective, this has far-reaching implications for liability and evidence management.

BYOAgentic: the new organisational risk

The Agency introduces an informal but highly illustrative concept: BYOAgentic. Just as the phenomenon was once BYOD (bring your own device), we now have BYOAI (bring your own AI). There is a risk that unqualified users may deploy agentic workflows outside the scope of corporate governance. No-code tools allow complex automations to be designed without the involvement of IT, security or the Data Protection Officer (DPO). The technical ease of use may lead to a dangerous trivialisation of risk.

A prudent, yet demanding, document

The text is not alarmist. Nor is it prohibitionist. The AEPD insists that agentic AI can be a tool for data protection by design. But it demands:

  • Formal governance.

  • Impact assessment where appropriate.

  • Strict memory management.

  • Granular control of privileges.

  • Effective human oversight.

  • Full traceability.

  • Identity and credential control.

  • Segmentation of processing operations.

It does not propose to hold back technology; it proposes to regulate it.

The document’s overarching message is clear: agentic AI is not an incremental evolution of generative AI; it is a paradigm shift in automation architecture. Data protection can no longer focus solely on:

  • the model,

  • training,

  • the dataset.

It must analyse:

  • orchestration,

  • memory,

  • service-to-service flows,

  • privileges,

  • autonomy,

  • reproducibility.

The AEPD has raised the bar in the debate. The question now is whether organisations are prepared to take on this level of technical and legal maturity. Because implementing an agent is straightforward, but governing it properly is another matter entirely.

reCAPTCHA protection