On 27 October 2026 the Spanish Data Protection Agency (AEPD) published Asistentes y agentes de IA. Que la IA actúe por ti, no sin ti ("AI assistants and agents: let AI act for you, not without you"), a decalogue for users, in Spanish. It starts from an idea its press release puts in two sentences: when the assistant only answers, a mistake is a wrong answer; when it acts, the mistake is an email sent, a file shared, a purchase, a contract or a consent given in your name.
The decalogue speaks to the person: which permissions to grant, what not to let it remember, when to demand confirmation. This article reads it from the other side — the organisation that person works for — and proposes which control matches each piece of advice.
Why advice to the person needs a decision by the organisation
The decalogue itself says so twice.
At work, according to the press release, the AEPD stresses the importance of using only the assistants authorised by each organisation and of not connecting work email or documents to a personal AI account. That advice can only be followed if someone has already decided which assistants are authorised. If nobody has, the person has nothing to go by.
On responsibility, the decalogue underlines that responsibility for AI assistants and agents cannot rest solely on the user. It addresses that to the provider of the service, but the reasoning holds just as well for whoever puts the tool in the hands of their staff. When someone uses an assistant with the data of customers, patients or students, the organisation answers for that data: that is the reading of agentic AI as a means of processing.
The ten points, as controls
The table sets each point of the decalogue next to a control from the Alethexis catalogue, by its code. The mapping is ours, not the AEPD's: the decalogue does not talk about controls, and the first column renders its headings in English as an index, not as a quotation.
| Point of the decalogue | Organisational control | What gets written down |
|---|---|---|
| 1 · Before delegating, check who you hand the keys to | M2-GOV-021 Master acceptable use policy for AI (in its first version, M1-VIS-033) | The list of authorised assistants, with the provider identified and what it does with the data |
| 2 · Grant access, not blank cheques | AGT-001 Scope boundary definition | Which tools and data each agent may use, and which it may not; reading is not the same as sending or deleting |
| 3 · Give it the minimum and protect what is sensitive | AGT-008 Protection of the model and AI assets | Where credentials are kept, and which data never goes into the assistant |
| 4 · Decide what it may remember about you | AGT-008 Protection of the model and AI assets | Which memory is switched on, what is deleted and how, using the service's own controls |
| 5 · Tell it what it must not do as well | AGT-001 Scope boundary definition and AGT-005 Time-boxed autonomy limit | The limits, set as permissions and not just as instructions; how long it may act without a person |
| 6 · Confirm what matters | AGT-002 Human-in-the-loop checkpoints and AGT-006 Pre-validation of irreversible actions | Which actions require confirmation before they run: sending, paying, contracting, deleting |
| 7 · Distrust what it reads too | AGT-007 Prompt injection defences | Which tasks combine outside content, private data and the ability to act, and what has been done so the three do not coincide |
| 8 · Protect other people's data too | M2-GOV-023 Policy for managing personal data in AI systems | Which sources holding third parties' data get connected, on what basis and with what limit |
| 9 · Protect your account and your device | AGT-008 Protection of the model and AI assets | Which add-ons and connectors are allowed, and from what source |
| 10 · Supervise it, learn to stop it and withdraw its access | AGT-003 Kill switch, AGT-004 Reasoning trace logs, AGT-009 Agentic change management and AGT-010 Anti-automation bias | How it is stopped, what record it leaves, what is reviewed when the agent is updated, and how to check that oversight is not a rubber stamp |
| If you use it at work | M1-VIS-039 Policy on personal accounts and corporate data in AI | That work email and documents are not connected to personal accounts, and what the corporate alternative is |
| What you can expect from the provider | The questions to the provider before authorising the assistant | Whether it offers what the AEPD expects of it (see below), in writing |
Three points in the table deserve one more line.
Points 5 and 7 say the same thing from two sides. The press release puts it this way: limits written into the instruction are guidance, but they do not replace permissions and confirmations. And on hidden instructions: since no measure removes this risk completely, the best protection is to limit what the agent can do if it is tricked. That is the idea behind the Rule of 2: untrusted content, sensitive data and the ability to act outwards should not meet in the same task. A limit written into a prompt guides; a permission never granted cannot be bypassed.
Point 6 has a premise you can copy as it stands. The AEPD urges clear limits on the premise that the agent prepares and the person approves, and advises against permanent authorisations. That is the design rule behind AGT-006: for anything irreversible, the agent leaves a draft and waits.
Point 10 opens with a short line from the decalogue: delegating is not washing your hands of it. Supervision means knowing how to stop the agent and withdraw its access before you need to, and reviewing its activity history. Circuit breakers are the technical form of the first; the log, of the second.
What the AEPD expects from the provider, as a list of questions
The decalogue closes with what can be expected of whoever offers the assistant. According to the press release, the provider should offer understandable permissions that separate consulting from acting, configurable confirmation for the highest-impact actions, a history of what the agent has done, an immediate way to stop it, management and deletion of its memory, information on who the data is sent to and whether it is used to train the AI, and a simple way to exercise data protection rights.
For an organisation, that is a list of questions to ask before authorising an assistant:
- Does it separate permissions to consult from permissions to act?
- Can confirmation be required before sending, paying or contracting?
- Does it keep a history of what the agent has done, and who can see it?
- How is it stopped, and how long does that take?
- How is memory managed and deleted?
- Who does it send the data to, and is the data used for training?
- Where are data protection rights exercised?
A "no" is not anyone's breach: it is information for deciding whether that assistant makes the authorised list, and with what limits.
What the decalogue is not
It is not an obligation. It is a set of recommendations for users. The controls in the table belong to the Alethexis catalogue and have their own basis: some rest on the GDPR, others are good practice, and others depend on whether the system is high-risk. None of them comes from the decalogue.
It is not the AEPD's guidance for controllers. That is a separate document: the guidance on agentic AI of February 2026, which we discuss in what the AEPD says about agents and data protection.
It does not cite the AI Act. It is a data protection document.
Where to start
With the point that makes the others possible: deciding which assistants are authorised and for what, and writing it into the AI use policy. There is an AI use policy template to start from. Where that step fits in the rest of the method is in the AI governance guide for SMEs.
The AEPD's reading is guidance, not law. The mapping of each point of the decalogue to a control is Alethexis's. Quotations from the AEPD, which publishes only in Spanish, are paraphrased in English.
This article is for information only and does not constitute legal advice.