Running AI agents in the EU: data residency, GDPR, and the EU AI Act

EU hosting alone does not make AI agents meet the GDPR or the EU AI Act: every hop outside the EU needs a legal basis, and the AI Act asks what the agent does.

Updated GATE

This guide walks through both laws for a business that puts AI agents to work, with the EU AI Act’s dates as they stand after the changes that entered into force on 27 July 2026. It is the practical map GATE uses when customers ask, not legal advice.

Does EU hosting, Ireland included, make an AI agent GDPR compliant?

Storing data in Ireland, Germany or any other EU or EEA country keeps that copy inside the area where personal data moves freely, so the transfer rules in Chapter V of the GDPR do not apply to it. That matters, and it is also where many reviews stop too early.

An AI agent has more hops than a normal app. Every time it reasons, it sends the task’s context to a model, and that context often holds personal data: the email it is answering, a support ticket, a customer record it pulled in. Around the model sit more services: the email that sends sign-in links, the edge network in front of the servers, the integration that holds your mailbox login, the people who support the system. Each hop that reaches outside the EEA is a transfer of its own.

So the question to ask a vendor is not “is the database in the EU” but “which hops leave the EU, and under what mechanism”. A vendor that names every hop and its legal basis is easy to assess. One that answers “EU-hosted” and stops there is not.

What does the GDPR need for a model call outside the EU?

Chapter V allows a transfer outside the EEA on a few bases. The two that matter for AI vendors:

  • An adequacy decision. The European Commission has found that some countries protect personal data essentially as well as the EU does. For the United States this covers only organisations certified under the EU-US Data Privacy Framework; a US vendor that is not certified gets no benefit from it.
  • Standard Contractual Clauses. The Commission’s pre-approved contract terms for transfers without adequacy. Since the Court of Justice’s Schrems II judgment, the exporter is expected to assess the destination’s laws and add safeguards where they fall short.

A transfer is doable, but it brings paperwork and assessments that need keeping up. A model served inside the EEA takes that hop out of Chapter V and leaves the ordinary duties: a lawful basis, data minimisation, a processor contract under Article 28 and the ability to honour access and deletion requests.

What does the EU AI Act ask of a business that deploys agents?

The EU AI Act regulates AI systems by risk. It does not ask where they are hosted. What it asks depends on your role: a provider develops an AI system or places it on the market under its own name, and a deployer uses one under its own authority. A business that puts agents to work in its own operations is usually a deployer, and its duties come in layers:

  • AI literacy, since February 2025. Article 4 asks providers and deployers to take measures to support the AI literacy of the staff who use AI on their behalf. Since the July 2026 amendments it no longer requires them to guarantee any specific level of literacy.
  • Transparency, from August 2026. Article 50 makes providers design systems that talk to people so the people know they are dealing with AI. Deployers must disclose deepfakes, and AI-written text published to inform the public, unless a person reviewed that text and holds editorial responsibility for it.
  • High-risk duties, from December 2027. Article 26 adds the heavy deployer duties for high-risk systems only: use the system as instructed, give competent people the job of overseeing it, keep the logs it generates for at least six months, and tell workers before it is used on them.

Most ordinary business automation is not high-risk. The list in Annex III is specific: biometrics, critical infrastructure, education, employment and worker management, access to essential services such as credit, law enforcement, migration and justice. An agent that answers support mail, drafts posts or keeps the books sits outside it; an agent that screens job applicants does not. The two laws also stack: the AI Act governs the system, and the GDPR still governs the personal data inside it.

When do the AI Act’s rules apply after the July 2026 changes?

The AI Act entered into force on 1 August 2024 and applies in stages. The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the high-risk dates. As the European Commission’s AI Act page lists them:

  • 2 February 2025: the bans on prohibited practices and the AI literacy duty.
  • 2 August 2025: the rules for general-purpose AI models.
  • August 2026: the transparency rules and most of the rest of the Act.
  • December 2026: a new ban added by the Omnibus, on AI that generates non-consensual intimate images or child sexual abuse material.
  • 2 December 2027: the high-risk rules for systems listed in Annex III.
  • 2 August 2028: the high-risk rules for AI built into products covered by EU product safety law.

How long can an agent keep personal data?

The GDPR’s storage limitation principle in Article 5 applies to everything an agent keeps: the logs of what it did, the memory it builds, the files it stores and the backups behind them. Each needs a retention period you can state, and a deletion that reaches the backups too. Sector rules can set minimums on top, such as bookkeeping law for financial records, so write down which rule sets which period.

Retention is also a question for the vendor: how long a deletion takes, and how long a copy survives in backups.

What should an EU-hosted agent setup give you?

EU hosting does not make you compliant on its own, but a well-built setup makes the rest manageable. Look for:

  • A named list of hops. Where the agent runs, where each model is served, where memory, files and logs sit, and which services sit around them, each with its transfer mechanism.
  • Data minimisation. The model receives the context of the task at hand, not your whole mailbox or customer base.
  • A way to see and stop the agent. What it did and in which order, and a person who can step in, which is what Article 26’s human oversight asks for once a system is high-risk.
  • Contracts in writing. A Data Processing Agreement under Article 28 of the GDPR and a published list of subprocessors.
  • Retention you can state. Deletion times, backups included.

Where GATE stands

GATE’s production servers, where agents run and memory, files and logs are stored, are in Nuremberg, Germany. Encrypted backups are kept in Falkenstein, Germany, and a few support servers run in Helsinki, Finland. The embeddings behind memory search are computed on GATE’s own server. A model receives only the context of the current task. By default, model calls and a few services around them, such as edge traffic, sign-in email and the logins for connected tools, run in the United States under the Standard Contractual Clauses, plus the Data Privacy Framework where the provider is certified.

GATE does not use customer content to train models. Two switches stay off unless an organization turns them on: providers in China, and providers that may train on your data. On Edge, GATE’s default release track, named GATE Intelligence staff can open chats, files and memory to operate, support and improve GATE; on Stable and on a dedicated server there is no standing staff access. An EU-only setup, with the models inside the EU too, is scoped on request, and the default setup applies until it is agreed and in place.

Every hop and its mechanism is listed on the subprocessors page and in Annex 2 of GATE’s Data Processing Agreement, and the controls are mapped in the Trust Center. Where GATE keeps your data shows the same picture on a map.

The takeaway

Treat EU hosting as the start of the answer, not the end. List every hop your agent’s data takes, give each one outside the EEA its legal basis, keep only what the purpose needs, and know which AI Act layer you are in: AI literacy now, transparency from August 2026, and the high-risk duties from December 2027 only if your use is on the list. If you want your own agent for the boring half of running your company, with that map already written down, GATE is invite only while we onboard a small number of businesses.

Common questions

Does hosting AI agents in Ireland or elsewhere in the EU meet the GDPR and the EU AI Act?

Not on its own. Hosting in Ireland or any other EU country keeps stored data inside the EEA, so the GDPR's Chapter V transfer rules do not apply to that copy. Every other hop still counts: a model called outside the EEA, an email or integration service abroad, support staff abroad. The EU AI Act does not require EU hosting at all; it asks what the AI system does and who uses it. This is not legal advice.

Does an AI agent's model call count as a transfer of personal data outside the EU?

It can. Each time an AI agent reasons, it sends the task's context to a model, and that context often holds personal data: the email being answered, a customer record, a support ticket. If the model runs outside the EEA, that call is a transfer under Chapter V of the GDPR and needs an adequacy decision or safeguards such as the European Commission's Standard Contractual Clauses. A model served inside the EEA keeps that hop at home.

What does the EU AI Act require from a business that only uses AI agents?

A business that uses an AI agent under its own authority is a deployer under the EU AI Act. Since February 2025 deployers must take measures to support the AI literacy of the staff who use AI. From August 2026 the transparency rules apply, such as disclosing deepfakes. The heavier deployer duties in Article 26, human oversight, keeping logs for at least six months and informing workers, apply only to high-risk systems.

When do the EU AI Act's high-risk rules apply after the July 2026 changes?

The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the EU AI Act's high-risk dates. Systems listed in Annex III, such as AI used in hiring, education or credit decisions, must comply from 2 December 2027. AI built into products covered by EU product safety law follows on 2 August 2028. The bans and the AI literacy duty have applied since February 2025.

Is a small business's use of AI agents high-risk under the EU AI Act?

Usually not. The EU AI Act's high-risk list in Annex III is specific: biometrics, critical infrastructure, education, employment and worker management, access to essential services such as credit, law enforcement, migration and justice. An AI agent that answers support mail, drafts posts or keeps the books falls outside it. An agent that screens job applicants or scores credit falls inside it, so check the use, not the tool. This is not legal advice.

What are the data residency requirements for AI agents?

No EU law is called a data residency requirement; for AI agents the duty assembles from the GDPR, contracts and sector rules. Check four hops: where the agent runs, where the model it calls is served, where its memory and files are stored and where its logs are kept. Then check the services around them, such as email, integrations and support access. Any hop outside the EEA needs its own transfer mechanism.

How long can an AI agent keep personal data under the GDPR?

Only as long as the purpose needs it. The GDPR's storage limitation principle applies to an AI agent's logs, memory, files and backups like to any other copy, and sector rules can set minimums on top. Write down a retention period for each. GATE's draft Data Processing Agreement commits to completing a deletion within 30 days of the request, with copies in backups rolling off within 90 days after that.

Does GATE keep AI agents and their data in the EU?

Partly, and GATE names which part. GATE's production servers, where agents run and memory, files and logs are stored, are in Nuremberg, Germany, with encrypted backups in Falkenstein, Germany, and support servers in Helsinki, Finland. Model calls and a few services around them, such as edge traffic, sign-in email and logins for connected tools, run in the United States under Standard Contractual Clauses. An EU-only setup is scoped on request.

Does GATE train AI models on customer data?

No. GATE does not use customer content to train machine learning models. A provider that GATE classifies as training on the data it receives is used only after 'Allow providers that may train on your data' is turned on for the organization, and that switch is off by default. GATE's Data Processing Agreement names each such provider, and the two whose terms allow training by default where the switch does not yet apply.

  • EU AI
  • GDPR
  • EU AI Act
  • data residency
  • compliance
← All articles

Want your own AI agent for your business?

GATE does the boring half of running a company: the books, contracts, social posts and your support inbox. Invite only while we onboard a small number of businesses.

Invite only. We'll meet you at the door. Prefer to talk? Talk to the founders.