Trust

Pseudonymisation as a supplementary transfer measure

This document describes how filterit's pseudonymisation corresponds to the supplementary measure described in Recommendations 01/2020 of the European Data Protection Board (EDPB) on transfers to third countries, so that an organisation can use it in its own transfer impact assessment (TIA) when it sends text to an artificial intelligence provider outside the European Economic Area. It is written from what the code and the deployment do, and it says explicitly what remains with the organisation.

Related: the technical and organisational measures (Article 32), the control mapping for ISO/IEC 27701 and 27001, the data processing agreement (section 7 on transfers), the sub-processor list, the AI transparency notice, the accuracy page and its method.

1. What this is and what it is not

EDPB Recommendations 01/2020 (adopted 10 November 2020, final version 18 June 2021) say that, where the law of the third country does not ensure equivalent protection, the exporter may add supplementary measures to the standard contractual clauses. In their annex, use case 2 is the transfer of pseudonymised data: the data leaves in such a way that it can no longer be attributed to a specific person without additional information, the additional information stays exclusively with the exporter inside the EEA, and it is protected by technical and organisational measures.

filterit implements exactly that scheme for one flow: the text that staff send to an artificial intelligence tool. Before the text leaves, the personal data that is detected is replaced with placeholders of the form [NAME_1], [AFM_2]. The mapping of placeholder to original value, which is the "additional information" of the Recommendations, is kept encrypted on infrastructure inside the EU or on the user's device, and is not part of what reaches the provider. The model's reply comes back with the placeholders and the values are restored at the organisation.

What it is not:

• It is not anonymisation. The mapping exists, so the data remains personal for the organisation. For the provider, whether the pseudonymised text is personal data is judged from the provider's own position and the means it may reasonably use (CJEU, C-413/23 P, 4 September 2025). This document supports no claim of anonymity towards the provider.
• It is not complete coverage. Detection is probabilistic. The published accuracy per category and detection level is on the accuracy page, with its method. Free text may hold indirect identifiers (a position, an employer, a rare event, a small community) that identify without a name.
• It is not the organisation's transfer impact assessment and it is not legal advice. It is the material for one section of it, the one on the supplementary measure.

2. What corresponds to which condition

Paragraph 85 of the Recommendations sets four conditions for use case 2. For each one, what filterit does and where the evidence is.

Contractual basis next to the measure: the model provider receives the pseudonymised text under standard contractual clauses contained in its data processing agreement, with terms that exclude use for training, as the sub-processor list states. Pseudonymised text that still holds identifiers that were not detected is thereby covered contractually, not technically.

What is not cited as evidence, because it ships switched off: row-level restriction in the database, the policy engine, session-scoped restore and the second version of the encryption envelope. An organisation that turns them on in its own deployment adds them to its own assessment.

3. What remains with the organisation

• The assessment of the third country's law for the specific provider and the specific use (steps 3 and 4 of the Recommendations). For providers in the United States, the organisation notes whether it relies on the EU-US Data Privacy Framework or on the provider's standard clauses, and follows the pending Latombe appeal (C-703/25 P) before the Court of Justice.
• The choice of provider and its terms: no training on the content, zero or minimal retention, known sub-processors. An organisation that uses the desktop agent with its own account at the provider has its own contract with that provider.
• Condition (d): for which categories of data and which uses the residual that may escape detection is acceptable, and for which it is not (health data, criminal convictions, cases in small communities). The decision is written into the assessment and into the organisation's acceptable use policy.
• Informing data subjects about the use of AI tools, the pseudonymisation and the recipients (Articles 13 and 14). Under C-413/23 P, the duty is judged from the controller's position at collection, regardless of what the recipient sees.
• Review, whenever one of the items below changes.

Open items as of the date of this document (2026-10-14), which the organisation follows:

• EDPB Guidelines 01/2025 on pseudonymisation were in public consultation; we have not confirmed a final version. The draft Guidelines 02/2026 on anonymisation are in consultation until 30 October 2026.
• Judgment C-413/23 P (EDPS v SRB) held that whether pseudonymised data is personal data for the recipient is judged from the recipient's position. The consequences for AI providers have not yet been reflected in guidance from an authority.
• The Commission's "Digital Omnibus" proposal on the definition of personal data has not been adopted.
• The Latombe appeal (C-703/25 P) against the adequacy decision for the United States is pending.

4. Review

This document is reviewed at least once a year, after any material change to the architecture, and when one of the open items in section 3 changes. Its version and date are at the top; the change history is in the repository. The document has not been reviewed by a lawyer; when it is, the table will say so.