Banco de España and AEPD expectations for AI agent oversight skills

Spain splits the question between two authorities. Banco de España expects banks to have sound governance, qualified staff and controls that work, and the AEPD expects personal data to be processed lawfully, with human involvement where automated decisions affect people. Staff overseeing AI agents in Spain need both skills: judging the agent's work and protecting the data inside it.

What does each authority expect?

Banco de España supervises credit institutions and other financial firms, with the largest banks under ECB banking supervision. Its expectations are the familiar prudential ones: a board responsible for strategy and risk, clear lines of responsibility, internal control functions that work, and people with the knowledge and experience their roles require. Technology risk sits inside that frame, now reinforced by DORA's requirements on ICT risk management, incident reporting, testing, third-party risk and staff training.

The AEPD is Spain's data protection authority. It enforces the GDPR and has given guidance on processing that embeds AI, with a consistent message: the controller must understand the processing, assess its risks, and be able to show that any human involvement is real. Under Article 22 of the GDPR, people have rights when a decision with legal or similarly significant effect is based solely on automated processing. Under Article 35, high-risk processing needs a data protection impact assessment.

The EU AI Act adds the cross-border layer. Article 4 asks providers and deployers to ensure sufficient AI literacy among staff. For high-risk uses, including credit scoring of individuals and life and health insurance pricing, Article 14 sets human oversight requirements and Article 26 tells deployers to assign oversight to people with competence, training and authority.

Who is in scope?

Any Spanish bank, payment institution, lender or insurer using agents, and any organisation, financial or not, that uses agents on personal data. The AEPD's reach is wider than the bank supervisor's: a utility or a retailer answering customers with an agent is under the AEPD but not Banco de España.

Within those organisations, the people who need oversight skills include reviewers, team leads, data protection officers, process owners and board members. The data protection officer has a particular role here. In Spain, as across the EU, they advise on DPIAs and monitor compliance, so they must understand how an agent reaches a decision well enough to judge whether the human review in the process is real. Our guide to AI literacy and certification for people who work with agents sets out the levels of understanding each role needs.

What does real human involvement look like with an agent?

This is where the two authorities meet. A reviewer who clicks approve on every held action is not human involvement in any meaningful sense, and both a bank supervisor and the AEPD will see through it. Real involvement has four marks:

1. The reviewer has the information. They see the facts the agent used, with provenance: sealed facts from the system of record, signed facts from a named party, declared facts from the customer, and external facts from outside. A decision built on external facts deserves a closer look.
2. The reviewer has the authority. They can decline, edit or escalate without needing anyone's permission, and the record shows they do so.
3. The reviewer has the skill. They know what the agent's depth allows, why an action was held, and that refused or withheld actions are not theirs to override.
4. The reviewer has the time. Volumes are set so review is possible. A queue of hundreds of held actions per person per hour is a rubber stamp by design.

For a team, a skills matrix for an AI agent team shows whether these marks hold across every shift, not just for the best reviewer.

How do you meet both sets of expectations?

Credentials should expire because the agents they relate to change. An agent's certificate lapses after 90 days and its depth can fall to 1 in a day; a staff credential that never lapses describes skills for a system that may no longer exist. The argument is made fully in why certificates for AI agent skills should expire.

What will a supervisor or the AEPD ask?

Banco de España inspectors will ask who governs the agent, who may approve its held actions, how competence is established and refreshed, and what the operation does when the agent drops to depth 1. The AEPD, typically after a complaint, will ask how the decision about the complainant was made, what the human reviewer saw, whether they could have decided otherwise, and whether the DPIA described this process accurately. Both will want the record: every verdict, filed and reconstructable.

Comparable expectations elsewhere in the EU are set out for De Nederlandsche Bank and AFM and for Banca d'Italia and IVASS.

What is not covered?

Neither authority prescribes a course, hours or an accredited provider. Neither decides what depth an agent may hold. The AEPD's remit is data protection, not prudential soundness, and Banco de España's is not data protection. Consumer conduct questions may involve other authorities. And training does not cure a badly designed process: if the agent's bounds are wrong, a well-trained reviewer is catching errors the design should have prevented.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

Where the Academy fits

Pryme Intelligence Academy's Operator track covers human review, reading the workspace and what to do when something goes wrong, then tests it in a two-week practicum of real review decisions with no missed red line allowed. That practicum is the kind of evidence the AEPD's question about real human involvement calls for: decisions made, checked and signed off by a Pryme assessor.