What data to share with AI, and how to protect it

The question is no longer whether to use AI: your teams are probably already doing it. The real question is which data can be sent to an AI model, through which channel, and with what guarantees. Here is the grid I use to sort data, the safeguards I build into an integration, and what the French authorities (the CNIL for data protection, ANSSI for cybersecurity) recommend. This page is not a substitute for advice from a lawyer or your DPO.

Sort your data by sensitivity

Not all data is equal. A public product sheet and an HR file do not call for the same precautions. I sort data into four levels before choosing a tool.

This table is a decision aid, not a legal rule. The right level also depends on your contract with the provider and on how the answers are used.

What the CNIL and ANSSI recommend

On July 18, 2024, the CNIL published a Q&A on using generative AI systems (in French, checked October 27, 2026). The points that matter for a company:

• only submit information the user is authorized to share, and keep personal data, confidential data and information covered by professional secrecy out of consumer services;
• use business accounts and turn off data reuse by the provider;
• for personal or sensitive data, including a document base the AI searches, prefer a deployment you control;
• sign a data processing agreement with the host and, where relevant, the AI provider, and put a legal basis in place for any transfer outside the EU;
• adopt an internal usage policy and train users on the system’s limits.

On the security side, ANSSI published security recommendations for a generative AI system on April 29, 2024. They focus on architecture: isolation, control over what data the AI system can reach, logging, and the full model lifecycle.

The core legal text remains the GDPR (Regulation (EU) 2016/679): data minimization, security of processing (Article 32), processor contracts (Article 28), and impact assessments when the risk is high (Article 35).

The technical safeguards I put in place

In a custom integration, you control what goes to the model. That is the main advantage over letting people use consumer tools freely.

1. Minimization. Only the fields needed for the task are sent. To classify a customer request, the request text is often enough; the name, address and contract number stay in your database.
2. Pseudonymization before sending. Names, emails and reference numbers are replaced with tokens before the call, then put back into the answer on the server.
3. Keys and calls on the server. The API key never sits in the browser. Calls go through your server, which applies the checks.
4. Inherited access rights. When the AI searches your documents, it does so with the logged-in user’s permissions. A sales rep cannot pull an HR document out with a clever question.
5. Region and contract. We pick a provider and a processing region that meet your requirements, with training on your data turned off. Choosing the model itself is covered in choosing an AI model for professional use.
6. Logging. We keep a record of who asked what and what the AI answered, with a defined retention period.
7. Human review. For any decision that commits the company or affects a person, the AI suggests and a human approves.

The contractual and regulatory framework

Three documents to check before connecting AI to personal data:

• the data processing agreement (DPA) of the AI provider and the host: purposes, sub-processors, location, retention period;
• the record of processing activities: the new use has to be listed;
• the impact assessment, if the processing involves sensitive data, evaluates people or poses a high risk.

The EU AI Act (Regulation (EU) 2024/1689) adds its own rules. Since August 2, 2026, an AI system that interacts with people has to let them know they are dealing with AI (Article 50). Companies that use AI must also take measures to support their staff’s AI literacy (Article 4, as amended by Regulation (EU) 2026/1744, published on July 24, 2026). The GDPR obligations of a business tool are covered in detail in GDPR and business apps.

When it is better not to send data to AI

I will be straightforward: some data has no business going into an AI model, even a well-governed one.

• When a simple rule does the job. Routing a request based on a keyword or a form field does not need AI.
• When errors are unacceptable. Payroll, invoice amounts, medical decisions: AI can help prepare the work, not make the call.
• When the volume is low. For ten sensitive documents a month, a trained person is faster than an integration project.
• When your contracts forbid it. Some customers or principals rule out any processing by a third party. Check your commitments before starting.

Where to start

1. Take stock of current usage: which AI tools your teams already use, with which data.
2. Classify your data with the grid above and write a short usage policy.
3. Pick a first use case with high value and well-controlled data.
4. Scope the integration: provider, region, contract, access rights, logging.

I handle this kind of project as part of AI integration, starting at $2,300 (€2,000)excl. VAT for a first use case. For AI built into a broader business tool, the estimator gives you a range in one minute. If what you need is an assistant that answers from your own documents, the page on an internal AI assistant connected to company documents explains the architecture.