Privacy and artificial intelligence: where your data goes

I start there because that's what I see in almost every mandate.

Management has not approved any AI tools. Officially, the company does not make any. In fact, several employees use consumer services on personal accounts, with company documents inside. A quote to be reformulated. A contract to summarize. A customer complaint to be answered politely.

Nobody's talking bad faith. They are effective people who try to do their job well. But each of these actions sends information from your company to a service you have not evaluated, under conditions that no one has read.

This is the real issue of confidentiality. Not the project you haven't started yet. The use which already exists and which no one regulates.

To judge a risk, you need to know what is really happening. Here is the path of a document you submit to a cloud AI system.

• The departure. The text leaves your station or server, encrypted in transit.
• The treatment. It arrives at the provider, who cuts it and gives it to the model. The model produces a response.
• Conservation. According to the contract, the content can be kept for a few days for reasons of security and abuse, or not at all. On business offers, this duration is usually short and documented.
• Training. The critical point. Does your content serve to improve the model? On serious business offers, no, and it is written in the contract. On free personal accounts, the answer is often yes, or varies depending on the settings.
• The return. The answer comes back to you.

At no time does the template "hold" your document as a database would. But the information has indeed travelled, and it is this journey that you must frame.

Your level of control depends on the structure chosen. There are three, at least the most controlled.

General public service. An employee uses a public tool with an individual account. Minimum control, no visibility for the company, generic conditions. Acceptable for non-sensitive content. Unacceptable for everything else, yet it is the most common architecture at the moment, by default.

Company offer with contractual commitments. The company contracts directly, with non-training, limited retention and sometimes localization clauses. Reasonable control, moderate implementation effort. This is the appropriate solution for the majority of Quebec SMEs, including for sensitive business data.

Controlled infrastructure or hosted model. The processing takes place in an infrastructure that you control, possibly in Canada, with an open model. Maximum control. Real cost and complexity. This is justified for health information, employee files, information covered by a professional order, or industrial secrets.

Many companies end up with a combination of the latter two, which avoids paying the price of the most restrictive case everywhere.

What this implies technically, in plain language, with the reminder that the legal interpretation belongs to your advisor.

When an AI system processes personal information, the same principles apply as for any other processing. In concrete terms, it raises technical questions that you need to be able to document.

• What personal information actually enters the system. Often more than expected, because an "anonymous" document contains a name in a signature or address in a footer.
• Where they are treated and preserved. Localization question, to be documented.
• Who has access to it. At your place, and at the provider's.
• Communication outside Quebec. If treatment is done elsewhere, it triggers evaluation obligations. Your advisor will tell you which ones apply to your case.
• The right of access and withdrawal. If an individual requests the deletion of their information, you can do so in your AI system, including search indexes.
• Minimization. Please send more information than necessary to the model. Often, yes, for ease of design.

These are technical questions to which GXN answers through implementation: filtering before sending, logging, access control, flow documentation. The legal qualification of your obligations is up to your advisor.

Honest answer: sometimes a lot, often not.

It really counts in four situations. When you process health information or data covered by a professional order. When your business customers or calls for tenders contractually require it. When you are a public body or work for one. When your sector has strong expectations, like financial services.

It means less to marketing content, public documentation or internal processes without personal information.

What matters in all cases is being able to answer the question. The day a business customer asks you where their data goes when you use AI, “I don’t know” is a bad answer. Documenting the journey, even when it passes elsewhere, is better than knowing nothing.

Let's go back to the real problem.

You can choose the most rigorous provider on the market, sign the best clauses, host in Canada. If three employees continue to contract in a public tool on their personal account, all this work is bypassed.

What works, in order.

• Look before you ban. Ask, without threat, who uses what and why. You'll get a free map of your internal frictions.
• Provide an approved option. Prohibition without solution produces clandestinity. A sanctioned tool, even imperfect, brings the use back to light.
• Write a short policy. Three pages that people read beat twenty pages that sleep. What is allowed, what is forbidden, what to do in case of doubt.
• Train on concrete cases. No theory. Examples from their actual work, with what is acceptable and what is not.
• Review periodically. The tools change, as do the uses.