AI and privacy
What a European provider solves and what it does not: training by subscription, retention, data location, sub-processors and your own responsibilities.
No. An EU provider largely removes one problem: personal data transfers to a country outside the European Economic Area. The GDPR's other requirements apply in full. You still need a lawful basis, a data processing agreement, correct settings and a reason why this particular data needs to go to a chatbot.
“Where is it held?” is the question asked most often and easiest to answer. This makes it seem more important than it is. The GDPR contains 99 articles, and one chapter concerns transfers.
Anyone opening Mistral's terms today will barely find Le Chat. Mistral says in its help centre that Le Chat is now Mistral Vibe: one product for work, coding and everyday conversations. Accounts, subscriptions and conversations have moved with it, while the address and sign-in remain unchanged.
This article consistently uses Mistral Vibe because that is the name in Mistral's documents. Mistral's pricing page calls the subscriptions Free, Pro, Team and Enterprise, with a student offering alongside them.
The transfer question, and with it a substantial amount of paperwork.
Chapter V GDPR begins in Article 44 with the principle that personal data may go outside the EEA only if that chapter's conditions are met. In practice, this means a European Commission adequacy decision (Article 45) or appropriate safeguards such as standard contractual clauses (Article 46), alongside a risk assessment.
With a provider established in the EU and processing data there, this does not arise or arises much less. Mistral is a French company based in Paris. Your contracting party falls directly under the GDPR and European supervision.
Under Mistral's business terms, French legislation applies to customers outside the Americas and Asia-Pacific, and disputes go to courts in Paris (as of October 2026). For a Dutch organisation, that is easier to navigate than a contract governed by legislation outside the EU. It says nothing about the rest of the terms. Read those yourself.
Caution is appropriate here. Mistral says in its help centre that data is hosted in the EU by default. The same explanation also says the following, as of October 2026:
“European” is therefore the starting point, not a guarantee for every feature. We do not name the cloud services and other parties Mistral uses here. That changing list is in Mistral's Trust Center, referenced by the data processing agreement. Read it before recording anything about data location in your own record.
Everything unrelated to location. Five subjects return with every chatbot, wherever it comes from.
Every processing activity needs a lawful basis (Article 6), and special category data such as health is subject to a prohibition with exceptions (Article 9). “We wanted a summary” is a purpose, not a lawful basis. Which data types belong in a chatbot and which do not is in May you put client data into an AI chatbot?.
Article 28 requires an agreement with every party processing personal data on your behalf. At Mistral, the answer depends on the terms under which you use the product.
The EEA consumer terms apply to personal use. Mistral itself is controller there, and no data processing agreement exists. The business terms apply to an organisation and say the data processing agreement applies if Mistral processes personal data on the customer's behalf on its own infrastructure.
An employee's own Free or Pro account therefore falls under consumer terms even when used for work. For a paid organisational account, establish which terms you accepted. What else to check is in Data processing agreement with an AI provider.
This is where a European provider differs least from an American one. See the next section.
Article 5(1)(c) requires processing only what is necessary for the purpose. Uploading a sixty-page file to find one deadline sends fifty-nine pages too many. That is as true for a provider in Paris as anywhere else. Mistral also expressly treats uploaded documents as input, with everything that entails.
A lawyer, notary, accountant or doctor has a separate confidentiality duty alongside the GDPR. That asks who can learn the file's contents, not where the server stands. A provider is a third party, even when European. What this means for those professions is in Professional secrecy and AI.
| Question | What a European provider solves | What you still arrange |
|---|---|---|
| Does data leave the EU? | Largely: storage is in the EU by default | Check features involving temporary transfers and read the sub-processor list |
| Is there a lawful basis? | Nothing | The basis per purpose, and an Article 9 exception for special category data |
| Is there a data processing agreement? | Only under business terms | Take out a business account and read the agreement |
| Is input used for training? | Nothing: depends on subscription and settings | Disable the setting and agree nobody uses feedback buttons for confidential work |
| How long is it kept? | Nothing | A rule for deleting conversations and knowledge of what remains afterwards |
| Is too much sent? | Nothing | Share only the necessary passage and replace personal data beforehand |
| Does my confidentiality duty apply? | Nothing | Assess your professional rules |
| Is it documented? | Nothing | An entry in the record and a DPIA where required |
| Do employees know what is allowed? | Nothing | Policy, explanation and a business account instead of personal accounts |
Mistral's help centre says input and output are used to train models by default unless you disable it. The exception named is Enterprise: those customers are off by default. This reflects October 2026. Check before relying on it, because these pages change regularly.
| Subscription | Training on input | Disabling | Terms |
|---|---|---|---|
| Free | On by default | By the user in settings | Consumer terms for personal use |
| Pro and student offering | On by default | By the user in settings | Consumer terms for personal use |
| Team | Mistral does not name Team as off by default; the administrator can disable training organisation-wide | By the administrator in the admin panel | Business terms |
| Enterprise | Off by default; only the administrator can enable it | Unnecessary | Business terms with an individual contract |
As of October 2026, based on Mistral's help centre and pricing page. The Team row needs attention. The explanation of disabling training says only that Team administrators can disable training organisation-wide, while the pricing page lists the same choice for Team as Free and Pro. Do not assume a Team account automatically stays outside training. Open the admin panel and check the switch.
Until you delete them. Mistral's privacy policy gives no number of days for the chatbot: input and output remain until you delete the conversation or close the account (as of October 2026).
Mistral's help centre says it may retain data in underlying systems for a time after deletion, including to investigate abuse or meet legal requirements. Mistral says zero retention exists only for certain API calls, not the chatbot.
For an organisation, this means everything stays without your own rule. Record when conversations are deleted and by whom. With Team and Enterprise, the organisation may have its own settings for this.
The chatbot can connect to other services. Mistral distinguishes connections retrieving data only when requested from connections indexing files or folders for chatbot searches. With the latter, Mistral holds an index of your content, deleted when an administrator disconnects it. Connecting a shared drive is therefore a bigger decision than uploading one file. What connections generally share and do not share is in What is an MCP server?.
The same steps as for any chatbot. The provider's origin makes the list one item shorter.
The choice between providers' subscriptions is compared in Which subscription may you use for business?.
If things go wrong, the provider's origin matters little. The Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, says entering personal data into an AI chatbot against employer agreements is a data breach. That warning concerns AI chatbots generally.
PrivaSearch addresses what a European provider does not solve: how much goes along. Personal data is detected within PrivaSearch and replaced with consistent pseudonyms, and a colleague checks before a document becomes searchable. Search questions are answered with passages and source references, per project and only for authorised users. The steps are on How PrivaSearch works.
You can search entirely within PrivaSearch. Semantic search then uses a European AI provider and only pseudonymised, reviewed text. Processing and storage take place in data centres in the EU. Parties at each stage are in the sub-processor overview, and principles on Hosting in the EU.
An AI assistant can also search through a connection: it receives only released, pseudonymised passages, no original files. This works with Claude, ChatGPT Business and Mistral Vibe. What reaches the assistant's provider through the connection, and what you type yourself, is honestly described on ChatGPT, Claude, Mistral Vibe and the GDPR.
Pseudonymisation is not anonymisation. For your organisation, it remains personal data because you hold the key. The lawful basis, DPIA and assessment of your situation remain yours and your data protection officer's. The difference between the techniques is in Pseudonymisation or anonymisation.
Want to see a question about a file answered without sending the file along? Book a demo.
No product makes an organisation automatically meet the GDPR. A European provider largely removes the issue of transfers outside the EU. You arrange the rest: lawful basis, business account with a data processing agreement, training setting, conversation deletion rule and agreements about employee input. Assessing your situation remains yours and your data protection officer's responsibility.
Mistral's help centre says input and output train models by default unless you disable it. Enterprise customers are off by default. On Team, the administrator can disable training organisation-wide. Rating an answer with a feedback button allows the rating and associated conversation to be used on every subscription. This reflects October 2026; check your settings.
Mistral says data is hosted in the EU by default. Depending on the feature, it may temporarily go outside the EU; locations appear on the sub-processor list in Mistral's Trust Center. Enterprise organisations can disable some features. Read the list before recording data location in your record.
Yes. Mistral renamed Le Chat to Mistral Vibe in August 2026, a product for work, coding and everyday conversations. Your account, subscription and conversations moved with it, while the address and sign-in remained the same. Mistral's terms, privacy policy and help centre now use Mistral Vibe throughout. Search using that new name when looking it up.
That depends on its contents, your lawful basis and your account. Uploading sends the whole document to the provider, and Mistral counts uploaded documents as input. A personal account lacks a data processing agreement and has training enabled by default. That is unsuitable for a file containing personal data, even with a European provider. Ask a lawyer if in doubt.
On one point, yes: fewer arrangements are needed for transfers outside the EU. Origin makes no difference on the other points. Training, retention, the data processing agreement and how much you share depend on subscription, settings and your own agreements. Compare providers by subscription, not by country.
Copilot shows what employees can already open, including overshared files. What Microsoft documents, what the DPIAs say and what to arrange first.
For each Claude plan: what happens to your input, how long it is kept, whether an agreement is in place and where your data is stored. October 2026.
What you need to arrange for each type of data before using an AI chatbot: from names and email addresses to health, criminal records and BSNs.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →