As soon as a company seriously considers AI, the question of data comes up, often phrased too vaguely to be dealt with: “is it secure?” That question has no general answer. Three precise questions, however, do.

Question 1: what leaves your company?

Anything sent to an outside service leaves your perimeter. That isn't a problem in itself: a public business address or a piece of marketing copy is nothing sensitive. A complete customer file, health data or contractual documents are another matter.

The right reflex is to sort before you automate: what can leave freely, what must be anonymized, and what never leaves. This sorting takes half a day and prevents most accidents.

The risk doesn't come from AI. It comes from what gets sent to it without thinking it through.

Question 2: which account does the processing run on?

Running processing on your own accounts changes three things: you see exactly what is consumed, there is no hidden margin on usage, and you can cut everything off in a single step. It's a position of control, not a technical constraint.

Conversely, a provider that runs everything on its own accounts makes you dependent: you know neither what is consumed, nor what is kept, nor what happens to the setup if the relationship ends.

Question 3: who can read what?

An agent connected to your documents inherits the rights it is given. If it has access to everything, it can hand any user information that person should never have seen. The issue isn't confidentiality toward the outside world: it's internal compartmentalization.

  • Grant access to what the task strictly requires, and nothing more.
  • Keep a record of what was consulted and produced.
  • Plan explicitly for what happens when an employee leaves.

What the regulatory framework already requires

Existing obligations on personal data apply unchanged: informing individuals, a defined purpose, a retention period, regulated subcontracting. AI has created no exception. A company that is already rigorous on this point has no revolution to carry out, just one more line in its register.

The page to write before connecting anything

Before connecting anything, write down on a single page: the categories of data involved, those that must never leave, which accounts the processing runs on, and who can view what. This page is worth more than any verbal commitment, and it turns the subject into something you can discuss instead of something to worry about.

Control isn't promised. It's observed in the architecture.