How we build GDPR-tight AI

Concrete tech architecture instead of marketing assurances: what we host where, how we encrypt, with which contracts — auditable in detail.

Data residency

All AI workloads run in EU regions. Standard stack:

• Model calls: Azure OpenAI in Germany West Central (Frankfurt) or Sweden Central — contract partner Microsoft Deutschland GmbH under the Online Services Terms.
• Application layer hosting: Hetzner Falkenstein, Upsun Frankfurt, or Azure Frankfurt — depending on requirements.
• Logs and telemetry: Azure Monitor in the same EU region as the model call. Never a US fallback.
• Backups: only in EU regions, with documented retention periods.

Encryption

Pseudonymization pattern

Every model call that might involve personal data runs through a local pseudonymization layer:

1. Detect: regex / NER / explicit field marking on the input form.
2. Replace: stable tokens (e.g. client name → Client_42) per request mapping.
3. Model call: the model sees only structured pseudo data.
4. Re-identification: after the model response, locally in your domain.
5. Audit log: mapping kept with access protection so each call stays reproducible.

DPA chain

• You as data controller (Art. 4(7) GDPR).
• IQONEX as processor (Art. 28 GDPR), DPA signed with you.
• Microsoft Deutschland GmbH as subprocessor — Online Services Terms incl. documented subprocessor list.
• Optional further subprocessors (Hetzner for hosting, Sentry for error tracking) — all listed in the DPA with documented TOMs.

§203 StGB — professional secrecy

For professional-secrecy holders (lawyers, doctors, tax advisors) the architecture is tailored accordingly:

• Local pseudonymization before every model call — clear names never leave the firm/practice.
• DPA with Microsoft Deutschland (with §203-StGB-compatible confidentiality obligation).
• Profession-equivalent TOMs by the processors (documented).
• Audit log with user identity per call — the professional must be able to reconstruct any AI-supported decision in dispute.

Audit log

Every model call is logged: timestamp, user identity (via Entra ID), model + version used, pseudonym hash of input, pseudonym hash of response. Retention per agreed period (6–10 years in regulated industries, often 12–24 months in Mittelstand workflows). Logs are reproducible, tamper-evident (hash chain), EU-resident.

EU AI Act

We classify each use case under the AI Act risk classes:

If your use case lands in "high risk", we say so in the intro call — and decide together whether the conformity assessment fits the budget or whether to redesign the use case.

What we don't do

• No US models without an EU layer (Azure OpenAI EU or Bedrock EU minimum).
• No model training on customer data (RLHF / fine-tuning is a different conversation, with DPA extension).
• No persistent logging of content without explicit mandate — audit log stores hashes and metadata, not content.
• No real-time biometric identification, no social-scoring systems, no manipulative AI (AI Act prohibited).

Documents shipped on engagement

• DPA per Art. 28 GDPR with documented subprocessor lists.
• TOMs per Art. 32 GDPR, extended for AI specifics.
• DPIA per Art. 35, reviewed by your DPO.
• Architecture document with threat model and data-flow diagram.
• Runbooks (deployment, incident response, restore).