[ Compliance ]

GDPR & the EU AI Act:An AI-Infrastructure Checklist

What two overlapping regimes actually require of the infrastructure your AI runs on — and a checklist to pressure-test your stack.

April 2, 202611 min readUpdated June 18, 2026

Two regimes now govern AI deployments touching EU data subjects: the GDPR, which governs personal data, and the EU AI Act, which governs AI systems by risk tier. They overlap. This guide cuts through the legal text to what they actually require of your infrastructure — and gives you a checklist to pressure-test a deployment against both.

Not legal advice

This is an engineering-and-operations checklist, not legal advice. It is designed to support compliance work and surface the questions your DPO and counsel need to answer — it does not, by itself, make any organisation compliant.

OrientationTwo regimes, one stack

It helps to keep the two regimes' jobs distinct, then see where they intersect on your infrastructure.

GDPREU AI Act
GovernsProcessing of personal dataAI systems, by risk tier
Core questionIs this data lawful, minimised, secure?How risky is this AI system, and what controls apply?
Infra implicationResidency, access control, encryption, deletionTransparency, logging, human oversight, robustness
OverlapBoth demand audit logging, data governance and security-by-design

GDPRWhat GDPR asks of your infrastructure

GDPR is principles-based, but several principles translate directly into infrastructure requirements:

  • Lawfulness & purpose limitation — you can name the legal basis and purpose for every category of data the AI processes.
  • Data minimisation — the system retrieves and exposes to the model only the data necessary for the task; PII is masked where it isn't needed.
  • Security of processing (Art. 32) — encryption in transit and at rest, access control, tenant isolation, and the ability to demonstrate them.
  • Storage limitation & erasure — data can be deleted on request and on retention expiry, including from vector stores and logs.
  • Records & accountability — you can produce a record of processing activities and demonstrate the controls, not just assert them.
  • Transfers — where data leaves the EU (e.g. a closed API), an appropriate transfer mechanism (such as SCCs) and a transfer risk assessment are in place.
Residency is configurable, isolation is not

GDPR rewards keeping data in a known jurisdiction with a clear legal basis. An architecture that pins inference to a chosen region — EU, or on-premise — and enforces tenant isolation by construction makes the GDPR story far easier to tell.

GDPRWhen you need a DPIA

A Data Protection Impact Assessment is required where processing is likely to result in a high risk to individuals — which large-scale AI processing of personal data frequently is. A DPIA documents the processing, assesses necessity and proportionality, evaluates the risks, and records the mitigations. Treat it as the artefact that ties your infrastructure controls back to the legal requirement.

EU AI ActThe risk tiers, and what they demand

The EU AI Act classifies systems by risk and attaches obligations accordingly. The tier your use case falls into determines how heavy the infrastructure burden is:

Risk tierExamplesHeadline obligations
UnacceptableSocial scoring, manipulative systemsProhibited
High-riskSystems affecting access to employment, credit, essential services, justiceRisk management, data governance, logging, human oversight, accuracy & robustness, technical documentation
Limited / transparencyChatbots, generative contentDisclosure that users are interacting with AI / that content is AI-generated
MinimalMost other applicationsVoluntary codes of conduct

Most general-productivity AI sits in the limited or minimal tier, where the headline duty is transparency. But the moment a system influences a decision about a person — hiring, lending, eligibility — high-risk obligations attach, and your infrastructure must produce logging, traceability and human-oversight hooks.

ChecklistThe AI-infrastructure checklist

Data & residency

  • Inference and storage are pinned to a known, documented jurisdiction.
  • Personal data is minimised and PII is masked before it reaches a model where not needed.
  • Data can be deleted on request and on retention expiry — including from vector stores and logs.

Security & access

  • Encryption in transit (TLS 1.3) and at rest (AES-256), with documented key management.
  • Tenant and role isolation enforced in the retrieval layer, not just the UI.
  • Cross-border transfers covered by an appropriate mechanism and a transfer risk assessment.

Transparency, logging & oversight

  • Users are told when they are interacting with an AI system or AI-generated content.
  • Every AI interaction is logged immutably with enough detail to reconstruct what happened.
  • For high-risk uses, a human-oversight path and technical documentation exist.

Governance & paperwork

  • A DPIA exists where processing is high-risk.
  • A signed DPA is in place with every processor; a BAA where health data is involved.
  • A record of processing activities and a control matrix are available for review.
[ FAQ ]

Frequently asked questions

Does the EU AI Act replace GDPR?

No. They are complementary. GDPR governs the processing of personal data; the EU AI Act governs AI systems by risk tier. An AI deployment touching EU personal data must satisfy both — GDPR for the data, the AI Act for the system. They overlap on requirements like logging, data governance and security-by-design.

Do I need a DPIA for an AI system?

A Data Protection Impact Assessment is required under GDPR where processing is likely to result in a high risk to individuals, which large-scale AI processing of personal data often is. The DPIA documents the processing, assesses necessity and proportionality, evaluates risks and records mitigations.

What makes an AI system high-risk under the EU AI Act?

Systems that affect access to things like employment, credit, essential services or justice are classified as high-risk. They carry obligations including risk management, data governance, logging, human oversight, accuracy and robustness, and technical documentation. Most general-productivity AI instead falls into the limited (transparency) or minimal tiers.

Does keeping AI data in the EU make me compliant?

No single control makes you compliant. Pinning inference and storage to a known EU jurisdiction strongly supports the GDPR residency and transfer story and makes the compliance case easier to tell, but you still need the full set of controls — minimisation, security, logging, deletion, governance and the relevant AI Act obligations.

[ Talk to us ]

Bring private AI to your regulated workload

We'll walk your team through the architecture, the contracts and the controls — against your actual requirements, not a generic deck.