Skip to content
Flopia
Data and GDPR

Your website's AI assistant: where your customers' data goes, and how to choose

Local model or a major provider's API: where your customers' messages go, what the GDPR says, model training, and how to choose a provider.

Illustration: a server connected to a chat bubble, with a shield and a padlock.

When a visitor writes to the assistant on your website, they sometimes share much more than a question: their name, address, phone number, sometimes a personal detail. These messages are personal data, and you are responsible for them. So “where are they processed?” is not just a technical detail.

Where do an AI assistant's messages go?

Most AI assistants do not run the model themselves: they send each message to a large model provider, often based in the United States, then display its answer. In practice, a single conversation can pass through:

  1. the chatbot tool and its hosting provider;
  2. the AI model provider, which receives the text needed for the answer;
  3. their own subcontractors (storage, logs, monitoring).

This is not illegal, but each step must be covered by proper safeguards, and you must be able to explain it to your customers.

Local model or major provider: two approaches

A “local” model is an open AI model that the tool installs on its own servers. Messages do not leave that machine, which makes the question of transfers simpler. On the other hand, the models that fit on an affordable server are smaller: their answers are often less reliable, slower, and they follow precise instructions such as “say you do not know rather than make something up” less well.

A major provider's API gives access to much more capable models, which are fast and better at following your rules. On the other hand, the messages needed for the answer are sent to that provider, often outside the European Union. You then need to check its terms: use of the data, retention period, transfer safeguards. A US company is also subject to US law, for example the CLOUD Act.

Neither approach is perfect. What matters is that your provider tells you clearly which one it uses, and what happens to the messages.

What the GDPR says

The General Data Protection Regulation applies as soon as you process the data of people in the European Union, including, in many cases, when your business is based outside the EU but serves customers there. For a chatbot, three points matter most:

  • The data processing agreement (Article 28): your chatbot provider processes data on your behalf; a contract must state what it is allowed to do with it, and which subcontractors it uses itself.
  • Transfers outside the European Union: they are possible, but only with safeguards (an adequacy decision, standard contractual clauses). The more intermediaries there are, the harder it is to check.
  • Informing people: your privacy policy must say who receives the messages, why, and in which country.

The CNIL, the French data protection authority, publishes guidance on AI and personal data: it is a good basis for checking any tool.

The question of model training

Another point is often overlooked: some services may use conversations to improve their models, depending on the plan and the settings chosen. Your customers' messages could then help train another company's AI. Always read the terms: the wording should be clear and, ideally, this use should be excluded by default. This is often the case for business API plans, unlike some consumer apps.

What Flopia does

We first tried a local model on our servers. For the quality and reliability of the answers, Flopia now relies on Claude, Anthropic's AI model, known for its reliability. In practice:

  • the website, the database, accounts and your settings are hosted in France, at OVHcloud;
  • for each answer, only the visitor's message, the latest exchanges and your relevant information are sent to Anthropic (United States), without the visitor's IP address or email;
  • under Anthropic's commercial terms, data sent through its API is not used by default to train its models;
  • this transfer outside the European Union is framed by the safeguards provided by the GDPR (standard contractual clauses in Anthropic's data processing terms);
  • nothing is sold, and nothing changes on your website without your confirmation.

Details are in our privacy policy.

Questions to ask any provider

Before installing a chatbot, whichever one it is, ask:

  1. Which AI model do you use, and who runs it?
  2. In which country are the messages processed, and with what safeguards?
  3. Who are your subcontractors, and where are they based?
  4. Are conversations used to train a model?
  5. How long are they kept, and can I delete them?
  6. Do you offer a data processing agreement that complies with Article 28 of the GDPR?

A serious provider answers these six questions without hesitation. If the answers are vague, that is already an answer.

Sources: text of the GDPR on EUR-Lex; CNIL resources on AI; Anthropic's commercial terms; Anthropic's privacy policy. Article updated on 14 October 2026. This article is for information only and does not constitute legal advice.

Keep reading

AI assistant

AI chatbot: how to stop it making up answers

Wrong opening hours, invented prices, impossible promises: why a chatbot sometimes makes things up, and 6 settings to keep it reliable.