Global Expansion EU

Cross-Border Data Transfer

China and the EU both regulate transfers, and the two regimes are not translations of each other. The work is finding where they diverge on the same data flow — and building one architecture that satisfies both.

Two regimes, one data flow, and no shared answer sheet

A company with EU customers and Chinese operations runs the same request through both systems. China’s Personal Information Protection Law and the GDPR govern that flow concurrently, but they were built on different logics: one starts from territorial and personal jurisdiction with a state-supervised export mechanism, the other from establishment and targeting with a transfer-instrument model. A document drafted for one does not satisfy the other, and translating it satisfies neither.

The practical consequence is that compliance has to be designed at the level of the data flow, not the document. Where the two regimes coincide, one control satisfies both. Where they diverge, the divergence has to be identified before the architecture is fixed — afterwards the cost is re-engineering.

What we do


Seven points where the regimes diverge

DimensionChina (PIPL)EU (GDPR)
Jurisdictional basisTerritorial and personal; extraterritorial where processing targets people in ChinaEstablishment, plus targeting of individuals in the EU
Export routeOne of three mechanisms: security assessment, certification, or standard contractAdequacy decision, appropriate safeguards, or derogation
ConsentSeparate consent required for export of personal informationTransfer is not grounded in consent as a general route; consent is a narrow derogation
LocalisationRequired for critical information infrastructure operators and over-threshold volumeNo general localisation mandate
DPONo mandatory DPO; a designated person required above volume thresholdsDPO mandatory in specified cases
Impact assessmentRequired in defined circumstances, with limited prescriptive guidanceDPIA mandatory for enumerated categories
Penalty ceilingUp to RMB 50 million or 5% of prior-year turnoverUp to EUR 20 million or 4% of global turnover

What is different for an AI company

Typical situations

Situation 1: An SCC was signed and treated as global compliance

The EU standard contractual clauses are executed, the transfer is documented on the EU side, and the China side is assumed to follow. It does not: the PIPL standard contract is a separate instrument with its own content and its own filing step, and an SCC satisfies neither requirement. Two filings, two documents, one data flow.

Our approach: treat the two as parallel obligations with a shared factual base. The data inventory, security measures and recipient description are written once and used in both filings, which shortens both and keeps them consistent.

Situation 2: Thresholds crossed without anyone noticing

Export assessment is not required for every transfer — it attaches above defined volume and sensitivity thresholds, or where the exporter is a critical information infrastructure operator. In an aggregated-API architecture, request volume accumulates across customers and crosses the threshold before anyone treats it as a compliance event.

Our approach: instrument the counter, not the intention. Establish a live measure of export volume by category, assign a trigger point below the legal threshold, and treat crossing it as a scheduled work item rather than a discovery.

Situation 3: Transfer impact assessment skipped because the SCC looked sufficient

Post-Schrems II, using SCCs requires an assessment of the legal environment in the destination country and, where necessary, supplementary measures. Whether China provides an adequate level of protection is the question most often avoided — and it cannot be avoided, because the assessment is what the transfer instrument depends on.

Our approach: run the assessment explicitly, document the conclusion either way, and where the conclusion is unfavourable, add the supplementary measures — encryption, pseudonymisation, split processing — that make the transfer defensible in practice rather than on paper.


Process and deliverables

StageWorkDeliverables
1Data flow mapping and categorisationTransfer map + legal basis per flow
2Dual-track assessment and gap analysisDivergence register + mechanism selection
3Documentation and filingsDPA set, standard contract or assessment pack, TIA
4Rights and governance buildRights response procedure + governance calendar
OngoingRe-assessment on change, annual reviewUpdated filings + change log

AI-specific deliverables

Frequently asked

Does a GDPR-compliant transfer architecture satisfy the PIPL? No. The mechanisms are institutionally different: one is a supervisory assessment or filing regime, the other a contractual or adequacy regime. They can run in parallel on one architecture, and that is usually the efficient answer, but neither substitutes for the other.

When is the security assessment required rather than the standard contract? Depends on the exporter’s status and the volume and sensitivity of what is transferred. The 2024 provisions raised thresholds and introduced exemptions, so the first step is measurement rather than assumption — the applicable mechanism follows from the numbers.

Is consent a workable route? On the China side it is a required element for export, not an alternative to the mechanism. On the EU side consent is a narrow derogation unsuitable for repetitive transfers. Relying on consent to solve the architecture problem generally fails on both sides at once.

How often should the assessment be refreshed? On any material change to the flow, the recipient, the data categories or the destination legal environment, and at minimum annually. Assessments that are never refreshed describe a system that no longer exists.


Does training a model abroad count as a transfer? Yes, if training data or user input leaves China. Corpora and input are assessed separately, and remote access by an overseas team is a transfer even without a file being exported.

Related Case Studies

All cases →

Each case states the obligation triggered, the judgment we made, and what was delivered.

Further Reading

Rule changes, enforcement cases and practical judgments around this service, listed in publication order.

Let's Talk About Your AI Company