The question arises at the first serious use, and many handle it by total avoidance or complete carelessness. Between the two there is a simple hygiene that lets you use these tools without exposing what matters.
Definition
Confidentiality with an AI model rests on three distinct things people tend to conflate: what you send, what the provider keeps, and what they do with it.
| Element | Who decides |
|---|---|
| What you send | You, entirely |
| Retention period | The provider's contract |
| Reuse for training | The contract, see Training data |
| Place of processing | The plan chosen |
The first line is the only one you fully control, and it is also the one with the most effect. The best contract in the world does not protect information you should not have sent.
The most useful habit is one question, asked before every send: would I put this content in an email to an outside supplier? If the answer is no, do not send it to a model either.
The four levels of sensitivity
No stakes. Public content, generic drafts, marketing copy. No particular precaution.
Internal. Procedures, notes, minutes. A business plan with a non-reuse commitment is enough.
Personal. Customer data, emails, named files. Processing agreement required, systematic minimisation, see GDPR and AI.
Critical. Health, data under professional secrecy, information under a confidentiality agreement. This is where Self-hosting with an Open-source model comes into its own.
The most frequent mistakes
The attachment sent without looking. A spreadsheet sent to analyse one column contains all the others, including those you had forgotten.
The screenshot. It shows far more than the subject: the rest of the screen, another tab, a notification.
The free tool chosen in a hurry. It is almost always the least protective offer that gets used for the most sensitive content, because the question did not arise at the time.
The colleague who did not know. Without a written, shared rule, everyone decides alone, and the least informed person decides like everyone else.
Check what your automation tool carries too. A Workflow sending a full file to a model to extract one line sends the whole file out, even if you only retrieve one value.
Frequently asked questions
Are business plans genuinely safer?
On contractual commitment, yes, and that is the main difference. They generally offer a non-reuse commitment, a short retention period and a processing agreement, which free offers rarely do.
What happens in the event of an incident at the provider?
Your data can be exposed as with any supplier. The reasoning is the same as for a hosting company or a management tool: it is not specific to AI.
Can you anonymise before sending?
Often, and it is very effective. Replacing names with identifiers rarely changes the quality of the analysis, and removes a large part of the problem.
How do you frame this in a small team?
With a one-page rule, written and known to everyone, listing what may leave and what may not. Our Claude Cowork course helps establish that rule before use spreads.