GDPR and LLMs and AI flows
Regulation · GDPR and LLMs
EU GDPR and LLMs: review the complete mission processing chain with the client
A mission may read a file, prepare a reply and pass its result to another tool. Personal data can appear in sources, prompts, outputs, logs and material retained for recovery. Reviewing only the model provider leaves part of the processing outside the assessment.
This guidance concerns the European Union GDPR and LLM use, with French CNIL guidance. It helps integrators prepare verifiable questions with the client and its DPO. British English does not make this a UK GDPR assessment, and the article is not a compliance assessment of DUBSAR.
In brief
The question and the useful result.
- Question
- Which data may the mission process, under whose responsibility and with which safeguards to establish?
- Useful result
- A processing and copy-review record for work with the client’s DPO and responsible teams.
GDPR and LLMs and AI flows
Separate legal obligations from work approvals
The EU GDPR requires, among other things, a defined purpose, an appropriate lawful basis, minimisation, limited retention and processing security (Articles 5, 6 and 32). Special-category data requires examination of Article 9 conditions.
Roles depend on actual decisions about purposes and means, not just contractual labels (Article 4). Where a provider acts as processor, examine Article 28 conditions. Information to individuals, exercise of rights, protection by design, a DPIA where required and transfers outside the European Economic Area are further matters to address as applicable.
A manager’s approval to send a message does not itself establish a lawful basis for processing. Nor does it prove the data subject’s consent. The CNIL usage FAQ helps examine which data may be shared and the deployment approach.
GDPR and LLMs and AI flows
Describe paths and copies
| Processing component | Matters to review |
|---|---|
| Purpose and roles | Intended result, controller, providers and decisions actually taken by each. |
| Source data | Necessary information, exclusions, consultation rights and special conditions. |
| Recipients | Model, checking services, business tools, processors and support access. |
| Copies | Prompts, outputs, temporary files, indexes, logs and evidence actually retained. |
| Retention and rights | Access, retention, deletion or other required handling; ways to locate relevant copies. |
| Transfers and safeguards | Processing and access locations; Chapter V assessment where a transfer is involved. |
Do not fill unknowns with assumed provider guarantees. A European region does not describe the entire chain. Replacing a name with an identifier does not establish anonymisation either: pseudonymised information can remain personal data.
GDPR and LLMs and AI flows
Scenario: a file is too broad for the intended reply
In this illustrative example, an agent must prepare a progress update for a client. The selected file also contains health information with no established need for that reply. The team identifies the mismatch before authorising that processing scope rather than assuming a model instruction will resolve it.
Appropriate responsible people establish the necessary data, relevant basis and conditions, then the permitted sources and recipients. If data has already been transmitted, removing a local document does not prove deletion of other copies. Establish what left and assign follow-up assessment.
The incoming colleague needs the scope version, decisions and open questions without receiving the entire sensitive file by default. This scenario announces no automatic DUBSAR detection, anonymisation or deletion.
GDPR and LLMs and AI flows
Retain proportionate evidence
Define what is needed to understand a decision or result: mission reference, reviewed version, authorisation, attempt and the response’s meaning. Evidence can itself contain personal data; its access and retention need consideration.
For a DPIA, Article 35 depends on processing risks and characteristics: using an LLM does not alone settle the requirement either way. Article 22 concerns certain solely automated decisions with legal or similarly significant effects; a formal approval step does not establish effective human involvement.
DUBSAR aims to connect mission decisions and evidence durably. That role remains to be qualified. It determines neither the lawful basis nor the absence of personal data or compliance of the entire chain.
GDPR and LLMs and AI flows
Frequently asked questions
Is human approval sufficient for GDPR compliance?
No. It can contribute to control of a specific action, but replaces neither the lawful basis, other processing duties nor checks that the relevant measures work in practice.
Must every prompt be retained as decision evidence?
Not as a general rule. Define the information needed, its access and retention. Copying all content can increase the data processed without improving the required evidence.
Does DUBSAR anonymise files or guarantee GDPR compliance?
This article establishes neither automatic anonymisation nor a compliance guarantee. Transformations, flows, providers and measures require assessment and verification within the client’s actual scope.
GDPR and LLMs and AI flows
Where DUBSAR fits
DUBSAR is developing a governance foundation for integrators and teams deploying agents for clients: durable mission state, explicit authorisations, human decisions, evidence and continuation of work. It aims to connect what was requested, what was permitted and what is established about the outcome.
Selected components and workflows have been tested; full B2B integration and validation are still underway. Pilots are not currently open.
GDPR and LLMs and AI flows
Sources and scope of review
Official sources reviewed on 20 October 2026. Scenarios and records are preparation methods that can be used without DUBSAR. They establish neither an available integration nor compliance.
- EUR-Lex — Regulation (EU) 2016/679 (EU GDPR) — Articles 4–6, 9, 13–22, 25, 28, 32, 35 and Chapter V.
- CNIL — Utilisation d’un système d’IA générative — Usage guidance, data boundaries and deployment choices; FAQ dated 18 July 2024, reviewed in 2026.
GDPR and LLMs and AI flows