Machine learning teams in Germany, Austria and Switzerland often want external engineering capacity for their ML platform but hesitate because the training data contains personal data. The underlying question is simple: can a team in India build and operate pipelines on data that must stay in the EU? The answer is yes, but only if you treat remote access as what it legally is, a transfer to a third country, and design the architecture so that the transfer is minimal, documented and justified. This guide covers the legal baseline and four architecture patterns, ordered from least to most data exposure.
Data protection as the practical constraint
The constraint is real and widely felt. In a Bitkom survey of 605 German companies published in September 2026, 50 percent said data protection hinders their use of AI, and 42 percent said they had avoided products or solutions from countries outside the EU because of data protection concerns. AI use keeps spreading regardless: Bitkom's September 2026 AI study found that 57 percent of German companies with at least 20 employees now use AI, up from 36 percent in 2025. The platform work behind that growth needs engineers the domestic market cannot fully supply, with Bitkom counting 79,000 unfilled IT positions in Germany in 2026. India's technology workforce, estimated by NASSCOM at 5.8 million people in fiscal year 2025, is an obvious source of capacity. That is why the transfer question has to be answered in the architecture rather than avoided.
The legal baseline in four points
- Remote access counts as a transfer. The European Data Protection Board's Guidelines 05/2021 on the interplay between Article 3 and Chapter V GDPR treat making personal data accessible to a recipient in a third country, including by remote access, as a transfer. Storing the data in Frankfurt does not avoid Chapter V if engineers in Bangalore can read it.
- India has no adequacy decision. Transfers therefore need appropriate safeguards under Article 46 GDPR. In practice this means the Standard Contractual Clauses adopted by Commission Implementing Decision (EU) 2021/914, typically Module 2 (controller to processor) or Module 3 (processor to processor).
- Clauses alone are not enough. Since the Schrems II judgment (C-311/18), the exporter must assess the law and practice of the destination country (transfer impact assessment) and add supplementary measures where needed.
- Processing terms are separate. If the external team processes personal data on your behalf, you also need a data processing agreement under Article 28 GDPR. The SCCs and the Article 28 agreement can be combined, but both functions must be covered.
Pattern 1: No personal data leaves the EU environment
The external team builds and maintains code, pipelines, infrastructure definitions and evaluation harnesses, but never sees production training data. They work against synthetic data or schema-identical fixtures generated inside the EU. Training jobs run in EU compute, triggered by CI, and only aggregated metrics and model artefacts flow back.
This pattern removes the transfer question for the data itself. It works well for feature pipelines, orchestration, model registry, serving infrastructure and monitoring. It works poorly for tasks that require looking at real samples, such as error analysis, labelling quality review or debugging a data drift incident.
Pattern 2: Anonymised or aggregated outputs only
The team receives outputs that are anonymous in the sense of Recital 26 GDPR, for example aggregated statistics, confusion matrices or embeddings that have been tested for re-identification risk. Anonymous data is outside the scope of the GDPR. The difficulty is proving anonymity: many ML artefacts, including some embeddings and memorising models, can leak personal data. Document the test you used and repeat it when the pipeline changes.
Pattern 3: Pseudonymised extracts with keys in the EU
Direct identifiers are replaced by tokens, and the mapping table stays with you in the EU. Pseudonymised data remains personal data for the controller (Article 4(5) GDPR). The Court of Justice indicated in 2025 (C-413/23 P, EDPS v SRB) that whether pseudonymised data is personal data for a recipient can depend on whether that recipient can realistically re-identify individuals. That nuance helps, but it does not remove your obligations as controller, and it should not be the only basis of your transfer assessment. Treat pattern 3 as a transfer covered by SCCs, with pseudonymisation as a supplementary measure.
Pattern 4: Controlled remote access to EU systems
Engineers work inside an EU-hosted environment through a virtual desktop or bastion host, without the ability to download data. Copy and paste, file transfer and local storage are disabled, sessions are logged, and access is time-bound. This is still a transfer, because data is displayed in India. It is, however, the most auditable form of transfer and often the right choice for incident response and error analysis.
Decision table
| Task | Recommended pattern | Notes |
|---|---|---|
| Pipeline code, orchestration, CI/CD | 1 | Synthetic fixtures must match schema and edge cases |
| Model registry, serving, autoscaling | 1 | No data access needed |
| Training runs and hyperparameter search | 1 | Jobs run in EU compute, metrics flow out |
| Evaluation and fairness reporting | 2 | Aggregates only, check small cell sizes |
| Error analysis on misclassified samples | 4 | Time-boxed, logged sessions |
| Labelling quality review | 3 or 4 | Prefer 4 for free text and images of people |
| Production incident on data drift | 4 | Break-glass procedure with approval |
Special categories and the AI Act
If training data includes special categories under Article 9 GDPR, such as health data, the bar rises sharply, and pattern 1 should be the default. For high-risk AI systems, Article 10 of the AI Act adds data governance requirements, and Article 10(5) allows processing of special categories for bias detection only under strict conditions, including that the data is not transmitted or otherwise accessed by other parties. That condition is difficult to reconcile with remote access from a third country, so bias analysis on special category data should run inside your environment, with only results shared.
Implementation checklist
- Classify every dataset: personal, pseudonymised, anonymous, special category.
- Decide the pattern per task using the table above and record the reasoning.
- Conclude SCCs (Decision 2021/914) and an Article 28 agreement before any access is granted.
- Complete a transfer impact assessment and list the supplementary measures you rely on.
- Generate synthetic fixtures inside the EU and version them alongside the code.
- Configure the remote access environment: no download, no clipboard, session logging, named accounts, expiry dates.
- Define a break-glass procedure for incidents, with approver and maximum duration.
- Review access lists monthly and remove dormant accounts.
- Update your record of processing activities (Article 30 GDPR).
What this means for staffing an ML platform
Most MLOps work is infrastructure and code, and most of it fits pattern 1. A remote team can therefore carry a large share of the platform without routine access to personal data, as long as synthetic fixtures and EU-side execution are set up first. That set-up is a sensible first sprint for any external team. A typical deployment sequence is described in our 72-hour protocol, and the access model is documented on our security page.
This article is general information and not legal advice. Have your data protection officer or a qualified lawyer review the transfer assessment for your specific datasets.
Data residency is an architecture decision first and a staffing decision second. niogod™ pods treat the EU-side execution environment and synthetic fixtures as the first sprint deliverable under a Werkvertrag, so that most of the platform work can proceed without routine access to personal data. If an ML platform build is planned for next year, it is a conversation worth having before the first dataset is shared.
Talk to niogod™
Planning engineering capacity in DACH? Book a 30-minute consultation with the niogod™ team.
Book consultation