Handing documents to an AI without knowing where they go: the concern is legitimate. It is settled not by promises but by choices you can verify. This article sets out three options for processing confidential data, seven common habits, and what the law says.

The real question: where does your data go, and who can access it?

Where data is processed matters as much as where it is stored. First trap: the consumer tool, used outside any contract. ANSSI, the French cybersecurity agency, says sensitive data should never be entered into such tools (its guide on AI, in French).

Three options, strengths and limits

Each should be judged on five criteria: where processing happens, training, retention, applicable law and in-house effort. The offerings are described from their documentation as of September 2026.

1. A major US cloud, with a European region

Example: Google Cloud’s Vertex AI, renamed Gemini Enterprise Agent Platform in 2026. It is a US service that can process data in Europe.

• For: nothing to install, a wide choice of models. Google undertakes not to train its models on your data without your permission. According to Google, an endpoint located in the EU (the United Kingdom is not included), which must be called explicitly, keeps processing there, but only for the models and features it lists as covered.
• Watch out for: the second trap, the “global” endpoint (the address that receives requests), which may process a request anywhere in the world, with no data residency guarantee. Achieving “zero data retention” takes deliberate steps: requests flagged as suspicious may be logged for up to 90 days unless you obtain an exemption, and some features have retention periods of their own. And US law follows the provider: under US legal process, the CLOUD Act covers data it holds or controls, whether inside or outside the United States.

2. A European provider

• For: a provider subject to European law. According to Scaleway, its shared models are hosted in Paris and, by default, requests are not retained, except in cases of error or abuse. According to OVHcloud, data sent to AI Endpoints is never used to train models (page in French), and those models can be deployed on your own infrastructure: a way out.
• Watch out for: European does not mean nothing ever leaves Europe. Depending on the feature, Mistral AI may temporarily transfer data outside the EU to listed sub-processors, and conversations with its consumer assistant are used for training by default unless you opt out. The AI Endpoints page does not say where requests are processed: ask for it in writing.

3. AI on your own servers

So-called “open-weight” models can be installed on your premises. Read their licence: Meta’s Llama 4 acceptable use policy does not extend its multimodal models to companies whose principal place of business is in the EU.

• For: the data need not leave your premises, provided neither the server nor its tools call any outside service, and no provider can reuse it. For sensitive data, the French data protection authority, CNIL, considers on-premises installation generally more secure (in French), but costly.
• Watch out for: you need hardware, roughly 2 GB of GPU memory per billion parameters at 16-bit precision, more for long documents. And a local server is not secure by nature: ANSSI notably recommends isolation, safe model formats, controlled administrator access, and filtering of inputs and outputs. Add updates, monitoring and someone to run it.

Seven habits for confidential data, whichever option you choose

• Send only what is needed. Pseudonymisation is not anonymisation: the data remains personal data (in French).
• Get a written commitment that your data will not be used for training, then check the setting in each tool.
• Sign a data processing agreement: documented instructions, security, authorised sub-processors, deletion or return of the data at the end of the contract, audits (General Data Protection Regulation (GDPR), Article 28, in French). And plan your exit: getting your data back, switching models.
• Apply access rights to the AI itself. An assistant connected to your documents must only show each person what that person is allowed to open. And do not train a model on restricted documents: according to ANSSI, access to them can no longer be restricted afterwards. On our archiving platform, a read-only profile cannot trigger the AI’s suggestions.
• Keep a record of exchanges, and protect it. ANSSI recommends logging prompts and responses, which may contain sensitive data.
• Set retention periods according to purpose, records included. On one of our platforms, files from an abandoned application that was never submitted are deleted one week after the session expires.
• Keep a human in the loop, and a fallback that works without AI. On our platforms, an authorised person decides the outcome of a case or the shelf for an original (see our dedicated article); on one of them, a reversible setting switches AI off.

What European and Ivorian law say

These are pointers, not legal advice: check them with your data protection officer (DPO) or your lawyer.

European Union and Luxembourg

• The GDPR requires data minimisation, storage limitation and confidentiality (Article 5, in French).
• The AI Act has largely applied since 2 August 2026, and since 2 February 2025 for AI literacy among staff; the rules on high-risk AI have been postponed to 2 December 2027 or 2 August 2028, depending on the case. In Luxembourg, a draft law would give the CNPD, Luxembourg’s data protection authority, a central role.

Côte d’Ivoire

• Law No. 2013-450 (in French) sets out similar principles, requires processors to provide “sufficient guarantees” and treats any state outside ECOWAS (the Economic Community of West African States), Europe included, as a “third country”.
• A transfer to a third country requires an equivalent or higher level of protection and prior authorisation from ARTCI, the Ivorian authority responsible. For an organisation subject to this law, entrusting personal data to an AI hosted in Europe or the United States therefore requires this authorisation in principle, unlike hosting within ECOWAS. Check with your lawyer or with ARTCI.

Where to start

• List the AI tools your teams already use (consumer or business account, training setting), offer a managed alternative, set a usage rule and train people.
• Classify your data (public, internal, sensitive) and identify personal data: it triggers the data processing agreement, any impact assessment (GDPR, Article 35) and, in Côte d’Ivoire, transfer authorisation.
• Match an option to each level, demanding more verified safeguards the more sensitive the data.
• Test before adopting, away from the live service, on documents whose correct answers are known. Start with fictitious or public documents; use real ones only on-premises or once the contract and settings have been checked.

Planning an AI project involving sensitive data? Get in touch to talk it through.