GDPR and AI Agents: How to Stay Compliant When Automating

Nearly half of all German companies now use AI — according to Bitkom. But only one in three has thought about data protection beforehand. And as soon as you deploy AI agents that make autonomous decisions, the question gets really interesting.

Because an AI agent isn't a normal tool. An AI agent decides. It acts. It stores. And depending on its architecture, it has a memory. That brings legal questions many SMEs haven't considered.

Well-intentioned isn't well-compliant. And nobody wants a GDPR fine of up to €20 million or 4% of annual revenue.

So: what do you need to consider when deploying AI agents in your company? Without scaremongering, but with clear rules.

What Changes with AI Agents Compared to Normal Software

A CRM system stores data. An AI agent does too — but it processes it differently. It analyzes, evaluates, and makes decisions based on probabilities. And that's the crux.

Traditional software: Input goes in, output comes out. The path is deterministic, traceable, reviewable.

AI agent: You give a goal, the agent decides how to reach it. It chooses tools, prioritizes data, and sometimes — if it has memory — retains information from previous sessions.

From a GDPR perspective that means:

• Automated decision-making (Art. 22 GDPR) — does your AI agent make decisions with legal effect? That's regulated.
• Data processing on behalf (Art. 28 GDPR) — who hosts the agent? Who has access to data?
• Right to explanation (Art. 13-15 GDPR) — can your AI agent explain why it made a particular decision?
• Data minimization (Art. 5 GDPR) — an agent with memory collects lots of data over time. Is it allowed to?

Sounds like a lot of bureaucracy? It is. But it's doable. We've solved it for our clients several times.

Art. 22 GDPR: When Your AI Agent Becomes a Compliance Trap

Article 22 is the paragraph most people underestimate. It states: you may not make decisions that have legal effect or similarly significantly affect a person solely on automated processing — unless the person has consented or it's legally provided for.

Sounds abstract? Concretely that means:

• Contract decisions: Your AI agent rejects a loan application? → Art. 22.
• Performance evaluation: Your AI agent evaluates an employee's performance? → Art. 22.
• Customer classification: Your AI agent classifies customers as "good" and "bad"? → Very likely Art. 22.

What you need to do: If your AI agent makes such decisions, you need a legal basis. Consent is an option but hard to implement. The more practical approach: put a human in the loop. The agent suggests, the human decides. That way the automation isn't "sole" and Art. 22 doesn't apply.

Pro tip: Explicitly define in your system design which decisions the AI agent may make autonomously and which require a human in the loop. Document it. Not just in code — for the data protection officer too. Our experience: clear separation is what auditors want to see.

Data Processing: The Contract You Need

Every AI agent needs infrastructure. Hosted in the cloud or locally — someone operates the tech. And as soon as a service provider has access to personal data, you need a data processing agreement (DPA).

The hurdle with AI agents: Many AI providers don't have DPAs for operating autonomous agents. Their standard contracts cover normal API usage — but not what an agent does: reading data from your system, caching it, processing it, and putting it into new contexts.

Checklist for your DPA:

• Are my data used for model training? (If yes — no thanks)
• Where are the servers located? (Must be EU/GDPR-secure)
• What happens to data in the agent's memory? (Deletion periods!)
• Who has access to the data? (Support, admin?)
• Is there a data export function? (No vendor lock-in)

Our approach at raspb: We host AI agents either on customer servers or in German data centers. The customer signs a DPA with us — and we contractually bind our subcontractors. Fixed price, transparent data processing, no training with customer data. Sounds simple. It is — when you do it right from the start.

The AI Agent's Memory: Blessing and Curse

An AI agent with memory is powerful. It remembers past conversations, learns from mistakes, and gets better over time. But that's exactly the problem.

GDPR requirements:

• Storage limitation (Art. 5 para 1 lit. e): data may only be stored as long as needed. An agent memory that keeps everything forever is not permitted.
• Right to erasure (Art. 17): a customer says "delete my data". Can your AI agent do that? Is the information still somewhere in its internal state?
• Data minimization (Art. 5 para 1 lit. c): the agent should only store data it actually needs. Not everything it sees.

Practical solution: Don't build your agent's memory as a black box. Structure it. Session memory for the current process. Long-term memory only with explicit consent. And implement deletion mechanisms — at the push of a button.

Our WissensRetter does exactly that. It stores knowledge in a structured way, gives the customer full control over data, and can be reset at any time. No black box, no data waste. Fixed price, clear rules, and 333% ROI.

Right to Explanation: Your Agent Must Be Able to Give Answers

This is perhaps the most underestimated point. The GDPR gives data subjects the right to know how a decision was made. With traditional software it's simple: you show the database query or the business logic code.

With AI agents? Harder.

An LLM isn't if-then-else logic. It calculates probabilities over billions of parameters. "Why did you classify this customer as low-risk?" — The answer lies in mathematical space, not in readable rules.

What you can do:

1. Build in tracking: Every AI agent decision is logged. Which data were used? Which tool was called? Which model made the decision?
2. Cite sources: If the agent uses your knowledge base, it must be able to say which document the information came from.
3. Enable human review: The agent gives a recommendation with source citations. The human reviews and decides.

We build an audit log into our agents by default. Every action, every decision, every data access is documented. The customer can trace what their agent did at any time — and present it to the auditor. That doesn't just save you from fines. It builds trust.

Data Minimization: How Much Should Your Agent Know?

A common misconception: "My AI agent needs access to all data to work well." Not true.

An AI agent for invoice checking doesn't need personnel files. An AI agent for customer communication doesn't need production data. An AI agent for compliance checks doesn't need employees' private emails.

The principle of data minimization means: Give your agent exactly the data it needs for its task — nothing more. That's not just GDPR-compliant. It also makes the agent more efficient because it delivers more relevant results.

How to implement it:

• Role-based access: The agent only gets access to the data sources relevant to its task. Like an employee, the agent only has the permissions it needs.
• Time limitation: The agent only stores data for the duration of processing. After that it's deleted or anonymized.
• System separation: Separate data processing from data storage. An agent can make a decision without permanently storing raw data.

Another pro tip: Start with a small, clearly defined use case. An agent that only checks invoice data is easier to make compliant than an all-rounder that searches the entire company. Start small. Learn. Then scale.

The Roadmap: Making Your AI Agent GDPR-Ready

In summary, the plan in five steps:

1. Inventory: What data does your AI agent process? Personal data? Special categories (health, religion, etc.)? Sounds boring, but without inventory nothing works.
2. Update your processing activity register: Your register must include the AI agents. What they do, what data they process, on what legal basis.
3. Check legal basis: Is it contract fulfillment? Legitimate interest? Consent? AI agents often need a combination.
4. Sign a DPA: With your AI provider, your hosting provider, everyone who processes data on your behalf.
5. Conduct a DPIA: If your AI agent "is likely to result in a high risk" to data subjects' rights (which is often the case with autonomous decisions), you need a DPIA.

Sounds like work? It is. But it's one-time work. And it's many times cheaper than the fine waiting for you if you skip it.

Conclusion: Compliant Isn't a Contradiction to Efficient

I often hear: "GDPR slows us down in AI adoption." — My honest opinion: that's only true if you treat compliance as an afterthought.

If you think about GDPR from the start — when designing the agent, choosing infrastructure, defining its tasks — then it's not a roadblock. It's a quality mark. Your customers trust you more. Your employees work more safely. The auditor is satisfied.

At raspb we build AI agents that are GDPR-compliant from the ground up. Not because it's trendy. But because we're an SME ourselves and know that trust is the hardest capital.

Our promise: Fixed prices, no surprises, your code belongs to you. And your data stays where it belongs: with you.

Ready for your first GDPR-compliant AI agent? Book a free AI audit — 30 minutes, concrete recommendations, no hidden costs. We'll check where you stand and what your first step should be.