Putting personal data into an artificial intelligence tool is not forbidden. What changes is that this use is a data processing operation like any other, and the General Data Protection Regulation (GDPR, or RGPD in Spanish) and the Spanish law that supplements it, the LOPDGDD, apply in full. The Spanish Data Protection Agency (AEPD) sums it up in its guide on AI: incorporating artificial intelligence removes no obligation, and also forces you to think about the risks specific to the technology.
This guide goes into practical detail: the contract with the provider, transfers to the United States, impact assessment, health data and what to do in the most common cases. If you are looking for the basics (legal basis, duty to inform and customer rights), they are explained in AI and GDPR with your customers' data.
What happens legally when you put personal data into an AI?
You start a new processing operation, or change one you already had. And that obliges you to answer five questions before switching anything on:
- On what legal basis? Performance of a contract, legal obligation, legitimate interest or consent. It has to fit the specific purpose.
- What exactly for? Data you collected to provide a service cannot, just like that, be used to train a model or for any other purpose.
- Who else touches it? Almost always, the AI provider.
- Where does it end up? If it leaves the European Economic Area, there are specific rules.
- What risk does it pose to people? Whether you need an impact assessment depends on that.
You are the controller: the company that decides what the data is used for and how. Responsibility does not pass to the provider, however much the tool is theirs.
What do you have to tell customers, patients and employees?
The GDPR obliges you to inform people about what you do with the data (Articles 13 and 14). With AI there are two additions that tend to be forgotten:
- The recipients. If the AI provider receives the data, your information has to mention that category of recipients and, if there are transfers outside the EU, say so.
- Automated decisions. If there are any, you have to inform people that they exist, give meaningful information about the logic involved and explain what consequences it has for the person.
The LOPDGDD allows this to be done in layers (Article 11): the basics at the moment of collecting the data and a link to the full information. On the web, that is usually your privacy policy, which has to describe what you do today, not what you did before using AI.
Why do you need a contract with the AI provider?
Because, in most cases, the provider is your processor: it processes data on behalf of your company. Article 28 of the GDPR says you may only work with processors that offer sufficient guarantees and requires a written contract. That contract must set out, at a minimum:
- That it processes the data only on your documented instructions.
- That its staff are bound by confidentiality.
- What security measures it applies.
- That it does not engage another processor (a sub-processor) without your authorisation.
- That it helps you respond to people's rights and meet your security and impact assessment obligations.
- That, at the end, it deletes or returns the data.
- That it gives you the information needed to show it complies and allows audits.
Large providers usually have this as a "data processing agreement" (DPA). Read it before accepting. Look at three things: whether it uses your data to train or improve its models, how long it keeps it and which sub-processors it has.
Another detail: the European Data Protection Board (EDPB) reminded in its Opinion 28/2024 that whoever deploys an AI model must make a reasonable assessment of whether that model was developed by processing data lawfully. It is not enough to trust the brochure.
Can you use an AI from a United States provider?
Yes, as long as the transfer is covered. Sending data outside the European Economic Area is an international transfer and the GDPR only allows it with safeguards. With US providers there are two usual routes:
- EU-US Data Privacy Framework. The Commission approved the adequacy decision on 10 July 2023. If the receiving company is certified under that framework, the data can travel without further transfer requirements. Check that your provider appears on the official list and that the certification covers the type of data you send.
- Standard contractual clauses. These are model contracts approved by the Commission. Many providers include them in their processing agreement.
Status as of 25 September 2026. The adequacy decision remains in force. The EU General Court dismissed an action for annulment against it on 3 September 2025, and that ruling has been appealed to the Court of Justice. On 29 June 2026 the US Supreme Court ruled in Trump v. Slaughter that the president can freely remove FTC commissioners, the authority that oversees compliance with the framework. On 31 July 2026 the EDPB wrote to the Commission asking it to assess whether this affects the functioning of the decision. Nobody has annulled it, but the ground is unstable. The practical recommendation: have your contract also include standard contractual clauses as a back-up, and review this point every few months.
When do you need an impact assessment (DPIA)?
The data protection impact assessment (EIPD in Spanish) is a prior analysis of the risks of a processing operation and of how you will reduce them. Article 35 of the GDPR requires it before starting when a high risk is likely, especially if new technologies are used.
The AEPD published a list to make this concrete. The rule: in most cases you must do one if the processing meets two or more of these criteria, among others:
| AEPD list criterion | Example with AI |
|---|---|
| Use of new technologies or innovative use of existing ones | Almost any project with generative AI |
| Profiling or assessing people | Scoring candidates or customers |
| Automated decisions, or ones to which it contributes heavily | Filtering applications with no human review |
| Special categories of data (health, among others) | Summarising medical records |
| Vulnerable people | Minors, people with disabilities, victims of gender-based violence |
| Combining databases for different purposes | Cross-checking the CRM with other sources to "enrich" records |
| Large-scale processing | Thousands of conversations or records |
Since AI already adds one, one more is enough for you to have to do it in most cases. The LOPDGDD (Article 28.2) adds its own list of higher-risk situations, which includes profiling, special categories, vulnerable groups and habitual transfers to countries without an adequate level of protection.
Can AI make decisions about people?
Article 22 of the GDPR gives anyone the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. Rejecting a candidate, denying a service or setting worse terms are good examples.
It is only allowed if it is necessary for a contract, if a law authorises it or if there is explicit consent. Even so, the person keeps the right to human intervention, to express their point of view and to contest the decision.
In practice: use AI to prepare, sort and summarise, and let a person decide. And make that review real, not an automatic click on what the machine proposes.
How do you reduce the data you give the AI?
The minimisation principle (Article 5 of the GDPR) says to process only data that is adequate, relevant and limited to what is necessary. With AI, this is what removes the most risk:
- Remove what is not needed. To classify an enquiry you do not need the ID number or the address.
- Pseudonymise. Replace names and numbers with codes and keep the lookup table separately. Pseudonymised data is still personal data, but the risk goes down. The EDPB has specific guidelines (01/2025) on how to do it properly.
- Anonymise when you can. Only truly anonymous data, from which nobody can re-identify the person by reasonable means, falls outside the GDPR. Deleting the name is almost never enough.
- Limit retention. Configure how long the tool keeps conversations and files.
If you want more detail on how to organise and clean your data before connecting it to an AI, see how to prepare your company's data for AI.
What changes if you process health data?
A lot. Health is a special category of data (Article 9 of the GDPR) and its processing is prohibited except in certain cases. A clinic usually relies on the one in Article 9(2)(h), healthcare, which requires the data to be processed by a professional subject to professional secrecy or under their responsibility.
That does not cover just any use. If you put medical records into an AI:
- The provider has to be a processor under an Article 28 contract and not use that data for anything else.
- You almost certainly need a DPIA: a special category plus new technology already add up to two criteria.
- Apply minimisation thoroughly: send only the part of the record the task needs.
- The clinical decision is taken by the professional, not the tool.
What must you never do?
- Paste customer, patient or employee data into an AI chat for personal use. Personal and free accounts do not usually come with a processor contract, and their terms may allow what you type to be used to improve the service. The AEPD says it clearly in its January 2026 ten-point guide: do not share personal data, sensitive information or customer and employee data.
- Let each employee use whatever tool they like. Define which ones are approved and for what. You will find more on this in the guide to AI security in the company.
- Reuse data for training without checking whether that new purpose has a legal basis.
- Accept the provider's terms without reading what it does with your data and where it stores it.
- Connect an AI agent to your systems with too many permissions. In February 2026 the AEPD published specific guidelines on agentic AI: the more the agent can do on its own, the more care about what data it reaches.
Situation by situation: what to do
| Situation | What to do |
|---|---|
| Customer service chatbot with customer data | Processor contract with the provider, transfers covered, a clear notice that the person is talking to an automated system and an up-to-date privacy policy. It should not ask for more data than necessary. See chatbots. |
| Transcribing calls with AI | Tell people at the start of the call about the recording and its purpose. Provider as processor, a defined deletion period and limited access to the transcripts. |
| AI to review CVs | Inform candidates, do not reject anyone automatically and do a DPIA (assessing people plus new technology). The AI Act treats these systems as high risk. |
| Analysing invoices with personal data | Minimise what is sent, provider as processor with no use for training and limited retention. If the invoices are from companies, still check the names of contact people. |
| Clinic using AI with medical records | Special category: DPIA, strict contract, minimum data sent, professional secrecy and the final decision by the healthcare professional. |
How does it fit with the EU AI Act?
They are two different rules that apply at the same time. The GDPR protects personal data. The AI Act regulates AI systems according to their level of risk, with obligations such as telling people they are talking to a machine or extra requirements for high-risk uses such as recruitment. Complying with one does not exempt you from the other. It is explained in the guide to the EU AI Act for SMEs.
A note on future changes: in November 2025 the European Commission proposed the Digital Omnibus, a package with two parts. The one on the AI Act is already approved and in force since 27 July 2026 (Regulation (EU) 2026/1744; we cover it in the guide to the Regulation). The one that amends the GDPR is another: as of 1 August 2026, according to the European Parliament, it was still under negotiation with no agreement. Until it is approved and published, the GDPR applies as it stands.
Checklist before putting personal data into an AI
- I know what I use the AI for and which legal basis covers that purpose.
- I have recorded the new processing in my record of processing activities.
- My privacy policy mentions the provider, the transfers and, if any, the automated decisions.
- I have signed the processor contract (Article 28) with the provider.
- I have checked whether the provider uses my data for training and have turned that off or opted out.
- I know where the data is stored and, if it leaves the EU, which safeguard covers the transfer.
- I have checked whether I meet two or more criteria of the AEPD list and, if so, I have done the DPIA.
- No important decision about a person is taken by the AI alone.
- I send the tool the minimum of data and pseudonymise when I can.
- I have set how long conversations and files are kept.
- My team knows which tools it can use and that no data is pasted into personal accounts.
- I have a procedure for handling rights (access, erasure, objection) that includes what is in the AI.
If you want to set up an AI agent that works with your company's data and do it right from the design stage, tell us your case.
This guide is for general information and is up to date as of 25 September 2026. It is not legal advice. For a specific processing operation, consult your data protection officer or a specialist professional.
Sources
Xora IntelligenceAI consultancy and custom software in Los Barrios (Cádiz). We write about what we build for businesses in the province. If you have a question about your case, write to us.




