# AI Agents and GDPR, compliance Guide for European Companies | Alher Tech

> European companies can run AI agents in full GDPR compliance, but not by default. The 7-point checklist, legal bases, DPAs, data minimization, retention on memory and traces, erasure that works, human oversight, plus the EU AI Act duties and compliant architecture patterns.

- Canonical page: https://alhertech.com/en/ai-agents-guides/ai-agents-gdpr-compliance/
- Site: Alher Tech (custom software, AI agents and SEO engineering, https://alhertech.com/)
- Contact: https://alhertech.com/en/contact/

---

Yes, European companies can deploy AI agents in full GDPR compliance, but not by default, an agent that reads customer emails, queries your CRM and calls a US-hosted model touches personal data at every step, and each step needs a legal answer. The good news is that the requirements are now well understood and every major model provider offers the contractual and technical tools to meet them. This guide walks through the roles, the legal bases, the seven-point compliance checklist, the interaction with the EU AI Act, and the architecture patterns that make compliance an engineering property instead of a legal afterthought.

## Why Agents Raise GDPR Questions That Chatbots Did Not

A classic chatbot processes the text a user types. An agent processes far more, it pulls customer records from your CRM, reads order histories, opens attached documents, keeps memory across sessions and writes detailed traces of everything it did. Each of those flows can carry personal data, and several of them send it to a third party (your model provider) and store it in a new place (your logs and vector database).

Concretely, an agent deployment creates four data flows that need review, personal data sent to the model provider in prompts, personal data retrieved from your systems as context, personal data persisted in agent memory and vector indexes, and personal data captured in observability traces. Most compliance gaps we find in audits live in the last two, because teams remember the model provider and forget their own logs.

## Roles, controller, Processor and Joint Controllers

Under GDPR, your company is the data controller. You decide why and how personal data is processed by the agent. Your model provider (Anthropic, OpenAI, Google or a European host) acts as your processor when serving API traffic, which is why a Data Processing Agreement (DPA) with them is mandatory, not optional. The same applies to your hosting, vector database and observability vendors.

If two entities jointly decide the purposes and means of processing, they are joint controllers under Article 26 and must define their responsibilities in a transparent arrangement. This is not exotic. Alher Tech itself operates under a joint controllership arrangement between its two principals, Francisco Herran Cerezo and Alejandro Alonso García, so we design agent architectures with the controller's obligations in mind because we carry them ourselves.

The practical takeaway, before any technical work, list every vendor that will see personal data through the agent and confirm you can sign a DPA with each. A vendor that cannot offer one is disqualified, whatever its benchmarks say.

## The 7-Point Compliance Checklist

These are the seven requirements we implement on every European agent deployment. Treat them as acceptance criteria for the project, not as a post-launch cleanup:

- **1. Legal basis per purpose**: Identify the lawful basis for each processing purpose, usually contract performance for support agents, legitimate interest (documented with a balancing test) for internal automation, consent where nothing else fits. Write it into your records of processing activities (Article 30).
- **2. DPAs with every subprocessor**: Model provider, cloud host, vector database, observability platform. Verify international transfer mechanisms (EU SCCs or the EU-US Data Privacy Framework) for any non-EU vendor, or choose EU-hosted endpoints.
- **3. Data minimization at the prompt boundary**: The agent should receive the minimum personal data needed for the case. Redact or pseudonymize identifiers before they reach the model where the task allows. Most classification and drafting works equally well on masked data.
- **4. Zero-retention and training opt-outs**: Configure API-level zero data retention or the shortest available retention with your model provider, and confirm contractually that your data is not used for model training. All major providers offer both in 2026; you must actually enable them.
- **5. Retention limits on memory, indexes and traces**: Agent memory, vector indexes and observability traces are personal data stores. Give each a retention schedule and an owner. Traces with full prompts are invaluable for debugging and radioactive for compliance, cap their lifetime and restrict access.
- **6. Data subject rights that actually execute**: Access, rectification and erasure requests must reach every store the agent created, cRM, memory, vector index, logs. Build the deletion routine during development; retrofitting erasure across embeddings and traces is painful.
- **7. Human oversight for consequential decisions**: Article 22 restricts decisions with legal or similarly significant effects made solely by automated means. Credit denials, contract terminations, claim rejections, keep a human decision point, and make the agent's recommendation and reasoning visible to that human.

## GDPR Meets the EU AI Act

The EU AI Act applies alongside GDPR, not instead of it, and 2026 is the year its main obligations bite for deployers. For most business agents the classification is straightforward, a support or back-office agent is limited-risk, which triggers transparency duties (users must know they are talking to AI) and the general-purpose AI provisions your model provider carries. Agents used in employment decisions, credit scoring, insurance pricing or essential services can be high-risk, adding requirements like risk management, logging, human oversight and conformity assessment.

The overlap is workable. The GDPR checklist above covers a large share of AI Act deployer duties. The two additions that matter for most companies are explicit AI disclosure in the user interface and documented human oversight procedures. Budget both from the start; they are cheap to build in and expensive to bolt on.

If your agent interacts with consumers, disclose that it is an AI system at the start of the interaction and offer a path to a human. This is both an AI Act transparency duty and, in our experience, good for customer trust and NPS.

## Architecture Patterns for Compliant Agents

Compliance is far cheaper when it is an architecture decision. The four patterns we use, roughly in order of adoption:

- **EU-region inference**: All major providers offer EU-hosted endpoints in 2026. Latency is equivalent, price is usually identical, and it removes the transfer analysis for the inference flow entirely. This should be your default.
- **PII redaction gateway**: A filtering layer between your systems and the model that masks names, emails, phone numbers and IDs, and restores them in the response. Adds a few milliseconds and removes a large share of exposure for tasks that do not need real identities.
- **Scoped retrieval with access control**: The agent's retrieval layer enforces the same permissions as the human it assists. It can only fetch records the requesting user or team could see. Prevents the classic failure of an agent leaking data across departments.
- **Self-hosted models for sensitive domains**: For health data, legal files or workforce analytics, open-weight models running on your own EU infrastructure remove the external processor entirely. You trade some capability and take on MLOps burden; for special-category data it is often the right trade.

## Questions to Ask Any Agent Vendor

Whether you buy a SaaS agent or hire an agency, these questions separate compliant vendors from hopeful ones:

- Where does inference run, and can you guarantee EU-only processing in the contract?
- Which subprocessors see personal data, and can we get the full list with DPAs?
- Is our data used for training any model, and where is that excluded in writing?
- How do you execute an erasure request across memory, embeddings and traces?
- What is logged, for how long, and who on your side can read it?
- How is human oversight implemented for consequential decisions?

## Compliance Is a Design Input, Not a Blocker

None of the requirements in this guide prevents you from deploying agents. European companies run compliant agents in production today at every scale, from ten-person firms to enterprises. What the requirements do is punish improvisation. The teams that struggle are the ones that built first and asked legal later.

Map the data flows, sign the DPAs, minimize what reaches the model, control retention, keep humans on consequential decisions and write it all down. Do this in the design phase and compliance costs you 10-20% of the project. Do it after an incident and it costs the project.

## Frequently asked questions

### Can we send customer data to US model providers at all?

Yes, through valid transfer mechanisms. The EU-US Data Privacy Framework for certified providers or Standard Contractual Clauses plus a transfer impact assessment. Using the provider's EU-region endpoint is simpler and removes most of the analysis, so prefer it when available.

### Do we need a DPIA for an AI agent?

Often, yes. A DPIA is required when processing is likely to create high risk, and agents that profile customers, process data at large scale or handle vulnerable users' data usually qualify. Even when not strictly required, a lightweight DPIA is the cheapest way to find design problems early.

### Does GDPR forbid the agent from making decisions automatically?

No. It restricts solely automated decisions with legal or similarly significant effects (Article 22). An agent can resolve tickets, draft answers and execute routine actions autonomously. For consequential decisions, keep meaningful human review, which is good engineering practice anyway.

### What about the data the agent remembers between sessions?

Agent memory is a personal data store like any other, it needs a purpose, a retention limit, access control and inclusion in erasure routines. The common mistake is letting memory grow indefinitely because it improves the agent; define a retention schedule from day one.

### Are fines actually a risk for a small company using agents?

The bigger short-term risks are a customer data leak through a careless integration and a lost enterprise deal because your compliance answers were weak. Fines scale with severity and negligence; a documented, honest compliance effort dramatically reduces both the probability and the size of sanctions.

## Related guides

- [What is an AI agent? Complete business guide](https://alhertech.com/en/ai-agents-guides/what-is-an-ai-agent/)
- [How much does an AI agent cost in 2026?](https://alhertech.com/en/ai-agents-guides/ai-agent-cost/)
- [How to build an AI customer support agent](https://alhertech.com/en/ai-agents-guides/build-ai-customer-support-agent/)
- [Our AI agent development services](https://alhertech.com/en/services/ai-agents/)