How the AEPD uses generative AI: eleven guidelines a firm can apply

The Spanish Data Protection Agency (AEPD) does not only supervise how others process personal data with artificial intelligence: it has also explained how it uses it itself. Between November 2025 and January 2026 it published three documents on the internal use of generative AI and, in February 2026, guidance on agentic AI.

These documents refer to the Agency's own administrative processes. They are not a rule for firms. But they come from the supervisory authority, and their guidelines can easily be carried over to any organisation that handles confidential information, such as a law firm.

The documents

According to the November press release, the policy is presented as the first of its kind in the public sector. The Agency itself clarifies that it does not amount to a verification, assessment or certification under the Artificial Intelligence Regulation.

The eleven guidelines of the January 2026 summary

The summary is addressed to decision-making bodies and teams implementing generative AI use cases. It warns that its list is not exhaustive. These are its guidelines, grouped:

On what is entered into the tool

1. Prevent users from entering personal, sensitive or confidential information into systems not authorised to process it.
2. Inform users that they must not enter private data, non-public information or personal data into those systems.
3. Train users in techniques for phrasing queries without details that compromise privacy or confidentiality.

On configuration

1. Configure systems with minimisation and limited data retention, including metadata and system memory.
2. Take use cases involving personal or confidential information preferably to internal or specific systems that guarantee the organisation's control.
3. Update data protection impact assessments when systems change significantly.

On outputs

1. Manually review and validate outputs before using or publishing them.
2. Keep effective human control mechanisms when the decision may affect rights or freedoms.
3. Do not include AI-generated content in official documents without human review and validation.
4. Make automated decisions understandable and provide notices of mandatory human review.

On suppliers

1. Assess before contracting the supplier's use of metadata, logs and training data, and include clauses on non-reuse of data and on compliance with the General Data Protection Regulation, the AI Regulation and the National Security Scheme.

How to apply them at a firm

For a firm, the translation is direct:

• Professional secrecy: the first guideline matches the duty of confidentiality. No client data should enter a tool not authorised by the firm.
• Filings: the seventh and ninth guidelines amount to a basic rule: no filing leaves the firm without a lawyer having reviewed what the tool generated. Courts have already sanctioned non-existent citations, as we reported in the review of the sanctions for fake case law.
• Supplier contracts: the eleventh guideline offers a list of points to negotiate with the tool's provider.
• Internal policy: the set of guidelines is a good skeleton for the AI policy we argued for in the editorial on AI policy at the law firm.

For a lawyer who practises alone, the same ideas fit with their general data protection obligations, which we review in the guide to data protection for the self-employed lawyer.

Agentic AI

The February 2026 guidance refers to systems that do not just respond but act autonomously to achieve their objectives. According to the Agency's note, they pose new data protection challenges, and the document proposes measures for controllers and processors. Anyone considering tools of this type at the firm should read the full document before rolling them out.

More analysis in Legal AI.