# GDPR and AI Agents: Data Protection When Deploying AI at Work

> Once an AI agent reads or writes personal data, the GDPR applies in full. You need a legal basis, clear roles with your vendors, usually a data protection impact assessment, workable data subject rights across every store the agent uses, and a valid basis for any transfer outside the EEA.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-law-and-compliance/gdpr-ai-agents-business  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-10-27

An AI agent is not a chat window. It reads inboxes, queries the CRM, opens tickets and triggers the next step without being asked twice. That means it almost always touches personal data: customer contacts, staff records, names buried in documents. This guide sets out which GDPR obligations apply to agents, why agents raise the stakes compared with a simple chatbot, and the order in which to work through them. It is orientation, not legal advice. Protecting know-how and confidential business information, which matters even when no person is identifiable, is covered separately in [trade secrets and AI](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/trade-secrets-ai-protection); this guide is only about protecting people.

## What does the GDPR mean for AI agents?

The GDPR covers any processing of information about an identified or identifiable person. There is no special regime for agents: an agent is a means of processing, and the organisation that decides why and how it is used remains the controller, with every duty that implies - legal basis, transparency, security and data subject rights.

[Article 4(1) GDPR](https://eur-lex.europa.eu/legal-content/DE/TXT/HTML/?uri=CELEX:32016R0679) defines personal data broadly. A support ticket with a call-back number, a machine serial number linked to a delivery address, or a free-text note about a phone call all qualify, because they let someone work out who the person is.

Agents add three things that matter less for a plain chat assistant:

- **More sources per task.** A quotation agent reads the incoming email, searches past quotes, pulls contact details from the CRM and drafts a reply. Each source is a separate processing operation with its own purpose.
- **Hidden stores.** Agents produce logs, execution traces, vector indexes and sometimes a memory that persists across sessions. Personal data ends up there whether or not anyone planned it.
- **Actions with consequences.** An agent that sends emails or edits records acts directly on people. That brings in the rules on automated decision-making and on data accuracy.

Two regulatory documents are worth knowing. The European Data Protection Board (EDPB) adopted [Opinion 28/2024](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf) on AI models on 17 December 2024. In Germany, the joint body of federal and state data protection authorities (the DSK) published [guidance on AI and data protection](https://www.datenschutzkonferenz-online.de/media/oh/20240506_DSK_Orientierungshilfe_KI_und_Datenschutz.pdf) on 6 May 2024, written for organisations that select, roll out and use AI applications. It describes itself as a guide rather than a complete checklist, but it is one of the most practical regulator texts available on the subject.

## What are the legal building blocks?

Six building blocks make up the assessment: roles, legal basis, processing principles, risk assessment, data subject rights and international transfers. None of them is new or AI-specific. What changes with agents is the weighting, because data flows multiply and an external model provider is involved in most setups.

### Roles: controller or processor?

Your organisation is normally the controller because it sets the purpose and means. If a model provider processes prompts and outputs only on your instructions, that is typically a processor relationship under Article 28 GDPR, according to the DSK guidance. You then need a data processing agreement, and you may only use providers that give sufficient guarantees of secure processing. If the provider also uses the data for its own ends, such as training, the DSK points out that joint controllership under Article 26 may apply, and each party then needs its own legal basis.

### Legal basis

[Article 6(1) GDPR](https://eur-lex.europa.eu/legal-content/DE/TXT/HTML/?uri=CELEX:32016R0679) lists six legal bases. In day-to-day business, three matter most: performance of a contract (b), legal obligation (c) and legitimate interests (f). Consent (a) rarely fits agent workflows because it has to be freely given and can be withdrawn at any time. The EDPB stresses that there is no hierarchy between the legal bases. For legitimate interests it sets out a three-step test: identify the interest, check that the processing is necessary for it, and make sure it is not overridden by the interests, rights and freedoms of the people affected.

### Principles

Article 5(1) requires, among other things, purpose limitation, data minimisation, accuracy, storage limitation, and integrity and confidentiality. For an agent this translates into narrow access rights, fixed retention periods for logs and traces, and a review step before anything the model says about a person is reused.

### Risk assessment and DPIA

Every processing operation starts with an initial risk screening. If it shows a likely high risk, a data protection impact assessment (DPIA) under Article 35 is mandatory. The DSK expects this to be the case for many AI deployments. It also notes that you depend on the provider for information about how the system works, so ask for that before you sign.

### Data subject rights

The rights in Articles 15 to 22 - access, rectification, erasure, restriction, portability, objection and protection from solely automated decisions - apply unchanged. They have to work across the agent's own stores, not only in the system of record.

### International transfers

Under Articles 44 onwards, transfers outside the European Economic Area need either an adequacy decision (Article 45) or appropriate safeguards such as standard contractual clauses (Article 46). With US-based model APIs, this question almost always comes up.

## Which use cases carry more data protection risk?

Risk depends less on the technology than on whose data the agent handles and what it triggers. An agent that searches engineering manuals is a different case from one that sorts job applications. The patterns below cover the deployments most companies start with, ordered roughly from lower to higher risk.

**Knowledge and document agents.** These answer questions from internal material such as manuals, inspection reports or project files (see [AI agents for knowledge management](https://www.ruebstudio.de/en/ai-agents/knowledge-management)). Personal data appears as names in documents, author fields and the questions staff ask. The main control is permissions: the agent must never show someone content they could not open in the source system.

**Sales and quotation agents.** They read enquiries, pull contact details and draft offers (see [quotation agents](https://www.ruebstudio.de/en/ai-agents/quotation) and [sales agents](https://www.ruebstudio.de/en/ai-agents/sales)). The legal basis is usually pre-contractual steps or legitimate interests. Scoring prospects or contacting them automatically moves the case closer to Article 22.

**Service agents.** They classify tickets, suggest fixes and prepare field visits (see [technical service agents](https://www.ruebstudio.de/en/ai-agents/technical-service)). Ticket text often holds phone numbers, home addresses or, after an accident, health details. Filtering free text before it reaches a model is a sensible default.

**Agents handling employee data.** Once an agent evaluates people's performance, behaviour or applications, the risk rises sharply. The DSK gives a concrete example of an Article 22 breach: an application that screens all applications for a vacancy and sends interview invitations on its own. In countries with works councils or similar employee bodies, introducing such a tool may also require their involvement; the DSK notes this for Germany.

**Self-hosted agents.** If the model runs on your own hardware or with a provider inside the EEA, the transfer question disappears and you control logs and deletion yourself. Everything else still applies. The page on [sovereign AI](https://www.ruebstudio.de/en/sovereign-ai) explains how such setups are typically built.

## What needs to be in place before you start?

Three things come first: a written purpose for each agent, a map of every data flow including intermediate stores, and a named person responsible for data protection questions. Without them you cannot assess legal basis or risk with any credibility, and you have nothing to show a regulator.

**Data.** For each agent, list what it reads, what it writes and where intermediate results live. A typical inventory: mailbox (read), CRM (read and write), vector index over the document store, log database, the provider's model API. For each item, record location, retention period and who has access. This also feeds the record of processing activities required by Article 30.

**Systems and contracts.** Ask the model provider: Is there a data processing agreement? Are prompts and outputs used for training, and can that be switched off? Where is data processed, and which sub-processors are involved? The DSK prefers services that do not use inputs and outputs for training; where an opt-out is impossible, that training needs its own legal basis.

**Organisation.** The DSK recommends clear internal rules on which AI tools may be used and for what, ideally with examples of permitted and forbidden uses. Staff should work with company accounts and devices, not personal ones, so that no individual profiles build up on third-party services. Bring in your data protection officer early, and any employee representatives. How these rules fit into a lean overall framework is covered in [AI governance for mid-sized companies](https://www.ruebstudio.de/en/guides/ai-strategy/ai-governance-framework).

**Model provenance.** The [EDPB opinion](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf) says that a controller deploying someone else's model should carry out an appropriate assessment, as part of its accountability, that the model was not developed through unlawful processing. The depth of that check depends on the risk of your deployment. In practice: request the provider's documentation on training data and take into account any findings by regulators or courts against it.

## How do you work through it step by step?

Start with the purpose and work towards the technology, not the other way round. The eight steps below follow the order in which a regulator would ask the questions. Each step produces a document you can hand over on request, which is what accountability under the GDPR comes down to.

1. **Define the purpose.** Two or three sentences on what the agent does and does not do. For example: "The agent drafts quotations from incoming enquiries. It never sends anything itself."
2. **Map the data flows.** Every source, target and intermediate store, with location, retention and access.
3. **Settle the roles.** For each vendor, processor or independent controller? Sign Article 28 agreements and disable training on your data wherever possible.
4. **Document the legal basis.** Per data category and purpose. For legitimate interests, write down the three-step test.
5. **Assess the risk.** Run the screening and, where risk is high, a DPIA under Article 35. The examples in Article 35(3) - systematic evaluation of personal aspects, large-scale processing of special categories, systematic monitoring of publicly accessible areas - are a starting point, not a complete list.
6. **Build it in.** Data protection by design and by default under Article 25: least-privilege access for the agent, pre-filters for free text, log retention limits, training use and prompt history off by default. Security under Article 32, including defences against manipulated inputs that try to steer the agent into unintended actions (see [prompt injection](https://www.ruebstudio.de/en/glossary/prompt-injection)).
7. **Keep a human in charge.** Where outcomes significantly affect people, a person decides with genuine discretion. The agent supplies a draft and its reasoning, not the decision.
8. **Inform people and handle requests.** Update privacy notices under Articles 13 and 14. Define how access and erasure requests reach logs, vector indexes and agent memory.

Deeper guides on the legal basis for AI processing and on DPIAs for AI agents are planned; the [guides section](https://www.ruebstudio.de/en/guides) lists them once they are published.

## What are the limits and risks?

Most risk sits around the model rather than inside it: uncontrolled intermediate stores, broad access rights, human oversight that exists only on paper, and a transfer mechanism that can change. Some data subject rights are also hard to honour once data has shaped a trained model.

**Erasure in derived data.** Personal data that has been embedded in a vector index or written to logs does not vanish when the CRM record is deleted. The DSK expects rectification and erasure to be workable inside the AI application. Design the index so every entry points back to its source and is deleted with it.

**Accuracy.** Language models produce fluent statements that can be wrong (see [AI hallucination](https://www.ruebstudio.de/en/glossary/ai-hallucination)). When those statements concern people, the accuracy principle and the right to rectification are engaged. Check model output about individuals before it lands in a system of record.

**Oversight on paper only.** The DSK is explicit that a merely formal human role is not enough. Someone approving a hundred suggestions a day under time pressure is no longer really deciding. Following the [human-in-the-loop](https://www.ruebstudio.de/en/glossary/human-in-the-loop) principle, show reviewers the reasoning and make disagreeing with the agent a normal, expected outcome.

**Transfers can shift.** For the United States, the [EU-US Data Privacy Framework](https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en) adequacy decision has applied since 10 July 2023, covering only US organisations that take part in it. The General Court dismissed an action to annul it on 3 September 2025, as set out in [press release No 106/25](https://curia.europa.eu/jcms/upload/docs/application/pdf/2025-09/cp250106en.pdf). The two previous arrangements were struck down by the Court of Justice in Schrems I and Schrems II, so keep a fallback ready: standard contractual clauses or a model hosted in the EEA.

**Models with a past.** The EDPB does not treat every model trained on personal data as anonymous. To be anonymous, both direct extraction of training data and obtaining such data through queries must be insignificant in likelihood. If a model was developed unlawfully, that can affect your deployment, depending on the circumstances.

**When this approach does not fit.** If no sound legal basis exists for a purpose, no technical measure will fix that. The same goes for a provider that will not explain how it processes or trains on data: without that information you cannot complete a DPIA. In both cases, narrow the purpose or choose a different system.

**Fines.** Article 83 provides for fines of up to EUR 10 million or 2% of worldwide annual turnover, and up to EUR 20 million or 4% for infringements such as breaches of the basic principles, data subject rights or transfer rules - whichever amount is higher in each case.

## How do you get started?

Pick one agent with a narrow purpose and low-sensitivity data, such as a knowledge agent over engineering documentation. Use it to build your records, deletion paths and approval steps properly once, then reuse that groundwork for higher-risk use cases.

A practical three-week start could look like this:

- **Week 1:** Purpose statement and data flow sketch for the first agent; meeting with your data protection officer; request vendor documents (data processing agreement, sub-processors, training use, data centre locations).
- **Week 2:** Document the legal basis, run the risk screening, decide on a DPIA. Draft an internal policy on AI tool use.
- **Week 3:** Implement the technical measures (access rights, pre-filters, retention limits), update privacy notices, and test access and erasure with a real test record that passes through every store.

In our own operations, one habit has proved its worth: running the deletion path end to end before go-live. Create a test contact, let the agent process it, delete it, then check whether it can still be found in logs, the index or agent memory. That exercise surfaces gaps no design document mentions.

If you are planning several agents, it pays to give one person or role overall responsibility for AI, bringing together data protection, engineering and the business units. The page on a [fractional head of AI](https://www.ruebstudio.de/en/fractional-head-of-ai) describes how such a role can work. For an overview of agent types, see [AI agents](https://www.ruebstudio.de/en/ai-agents).

Status: October 2026. Case law and regulator guidance on AI are moving quickly; check the current position before you make decisions.

## Frequently asked questions

### Does every AI agent need a DPIA?

Not automatically. Article 35 GDPR requires one where processing is likely to result in a high risk, and German supervisory authorities expect that to be the case for many AI deployments. A documented initial risk screening is needed either way.

### Can an agent make decisions about people on its own?

Decisions with legal or similarly significant effects generally must not be based solely on automated processing under Article 22 GDPR. The human reviewer needs real discretion; a rubber stamp does not count.

### Is a large language model itself personal data?

It depends. The European Data Protection Board considers a model trained on personal data anonymous only if extracting that data, directly or through queries, is insignificant in likelihood. That has to be assessed case by case.

### Is this legal advice?

No. It is an orientation on the obligations and the questions to ask. For your specific project, involve your data protection officer or a law firm.

## Sources

- [Regulation (EU) 2016/679 (General Data Protection Regulation), EUR-Lex](https://eur-lex.europa.eu/legal-content/DE/TXT/HTML/?uri=CELEX:32016R0679)
- [EDPB: Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models (17 December 2024)](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf)
- [DSK (German Data Protection Conference): Guidance on Artificial Intelligence and Data Protection, version 1.0, 6 May 2024 (German)](https://www.datenschutzkonferenz-online.de/media/oh/20240506_DSK_Orientierungshilfe_KI_und_Datenschutz.pdf)
- [European Commission: EU-US data transfers (Data Privacy Framework)](https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en)
- [General Court of the EU: Press release No 106/25, judgment in Case T-553/23 Latombe v Commission (3 September 2025)](https://curia.europa.eu/jcms/upload/docs/application/pdf/2025-09/cp250106en.pdf)
