Guide
AI agents in your company under the GDPR and the EU AI Act
What the GDPR and the EU AI Act ask of a company that runs AI agents, what applies now and what comes in 2027, where your data goes, and how to design an agent that is easier to keep compliant.
Updated 2026-09-15
An AI agent reads your emails, your tickets and your customers' messages. It calls your systems and sometimes acts on its own. In the European Union, that brings two laws into play: the GDPR, which already applies in full, and the AI Act, which is arriving in stages.
This guide covers what each asks of a company that runs agents, and how to design an agent that is easier to keep on the right side of both.
The AI Act: what applies now
Checked in September 2026. This is a practical overview, not legal advice. For decisions that matter, talk to a lawyer or your data protection officer.
The AI Act entered into force in 2024 and applies in stages. In July 2026, the Digital Omnibus on AI (Regulation (EU) 2026/1744) moved the high-risk deadlines back. The timeline today:
| Since or from | What applies |
|---|---|
| February 2025 | Prohibited AI practices are banned |
| August 2025 | Obligations for providers of general-purpose AI models |
| 2 August 2026 | Transparency obligations (Article 50) |
| 2 December 2026 | Marking of AI-generated content, for systems already on the market before August 2026 |
| 2 December 2027 | Obligations for high-risk AI systems (Annex III) |
| 2 August 2028 | Obligations for AI inside products covered by EU product-safety law (Annex I) |
Transparency: this applies to you today
If people talk to your agent, they must know they are talking to AI, unless it is obvious. A support bot on your website, an assistant in WhatsApp, an agent that answers emails: each one should say what it is.
Content your agent generates, such as synthetic audio, images or video, and in some cases text published to inform the public, comes with labelling duties of its own.
Is your agent high-risk?
Most business agents are not. An agent that triages support tickets, runs your tests, drafts reports or answers questions about your product is outside the high-risk list.
It changes if the agent helps decide things that shape people's lives. The high-risk areas include:
- Employment: screening CVs, ranking candidates, evaluating performance, deciding promotions or dismissals
- Credit: assessing whether a person is creditworthy
- Insurance: pricing or risk assessment for life and health insurance
- Education: admissions, grading, monitoring students during tests
- Essential services: deciding access to public benefits and similar services
If your agent works in one of these areas, plan now for December 2027. Risk management, data governance, technical documentation, logging, human oversight and accuracy requirements all take time to build.
The GDPR: your agent processes personal data
Nothing about an agent is exempt from the GDPR. If it reads a name, an email address, a phone number or a customer record, it processes personal data, and the usual rules apply.
What that means for an agent
- A legal basis for every use. The reason you hold customer data does not automatically cover feeding it to an AI model for a new purpose.
- Only the data the task needs. An agent that answers invoice questions does not need the customer's full history in every prompt.
- A data processing agreement with every processor. Your AI model provider, your hosting provider and your messaging platform each handle personal data on your behalf.
- Transfers outside the EU. If a request reaches a model provider outside the EU, that is an international transfer and needs a valid basis.
- A DPIA when the risk is high. New technology used on a large scale, sensitive data, or systematic monitoring usually calls for a data protection impact assessment.
- Humans in decisions that matter. People have the right not to be subject to decisions based solely on automated processing that have legal or similarly significant effects on them.
- Retention rules for transcripts and memory. Agents keep conversation logs and memory files. They need a retention period like any other record.
- Data subject rights include the agent. When someone asks for access to or deletion of their data, the agent's logs and memory are part of the answer.
Follow the data
Before anything else, draw where a single request travels. For an agent that answers customers on WhatsApp, it might look like this:
One request, end to end
- The customer sends a messageAnother company
The messaging platform
- It reaches the agentYour hosting, EU
The machine the agent runs on
- A routine request is handled locallyNobody else
A local model on the same machine
- A hard request goes to a cloud modelMay leave the EU
The provider's servers
- The agent acts in your applicationYou
Your application and database
- The conversation is storedNeeds a retention period
Logs and agent memory
Every step with another company in it needs a processing agreement. Every step that can leave the EU needs a transfer basis.
Every row with another company in it needs a processing agreement. Every row outside the EU needs a transfer basis. Every row that stores data needs a retention period.
Design choices that make compliance easier
- Let a local model see the personal data first. A small model on your own machine can classify requests, extract fields and remove names or ID numbers before anything is sent to a cloud model.
- Host the agent in the EU. The machine that holds transcripts, memory and credentials stays under EU law.
- Choose API terms that fit business data. Check each provider's data retention and training terms for API use, and whether EU data processing is available.
- Require approval for actions with consequences. Sending an invoice, changing a record, replying to a complaint: let the agent prepare, and a person confirm.
- Keep an audit trail. For every action: which request triggered it, what the agent decided, which tool it called, what changed, and who approved it.
- Give the agent scoped credentials. Read-only where possible, limited to the systems the job needs, easy to revoke.
- Local AI models on a Mac mini: what runs on each memory size, for the first of these
Checklist
AI Act
- Users are told when they interact with the agent
- AI-generated content is labelled where required
- You have checked whether any use falls into a high-risk area
- If it does, you have a plan for December 2027
GDPR
- You know which personal data the agent reads, stores and sends
- Each use has a legal basis
- Processing agreements are signed with every processor
- Transfers outside the EU are covered
- A DPIA is done where the risk calls for one
- Decisions with significant effects have a human in the loop
- Transcripts and memory have a retention period
- Access and deletion requests cover the agent's data
Operations
- Every agent action is logged
- Consequential actions need approval
- Credentials are scoped and revocable
- Your privacy policy names the AI providers you use
Why MacDuty
- Hosted in the EU. Your Mac, its disks and every backup stay in the European Union, operated by an EU company.
- A processing agreement on request.
- One customer per Mac. Your agent's transcripts, memory and credentials sit on hardware nobody else uses.
- Local models on the same machine. Keep routine requests, and the personal data in them, off third-party AI services.
Sources
Next