What is shadow AI?
Shadow AI covers every generative AI tool used for work without the company’s approval: the free version of a chatbot, a transcription app on a personal phone, a browser extension that rewrites emails.
It inherits from shadow IT, the spreadsheet hosted who knows where or the messaging app installed without telling anyone. With one major difference: a shadow IT tool stores what you give it. An AI tool reads it, rewrites it and sends it back into a quote, a client email, a board note. The data goes out, and the answer comes back into your official documents.
What the figures say
On 14 September 2026, Cybermalveillance.gouv.fr published its recommendations for framing generative AI in business. The figures released alongside: 45% of employees use unapproved AI tools, and 36% of them paste confidential data into them.
Bpifrance Le Lab had set the scene in its 2024 study on French companies and AI: 54% of users worked with free, generic tools, and 73% of AI projects were carried by the owner alone.
Put together, these figures describe a company split in two. The owner tests alone, on their side. The teams do the same, on theirs. Nobody tells anybody.
What we see at our clients
The pattern comes back in almost every company we work with. The owner has a paid subscription to a consumer assistant, moves fast and is happy with it. The teams use free versions, tolerated as long as nobody asks.
The result: the owner does not know which documents left the company, nor which invented answers made it into a quote. Nor does he know which employees found uses that would save the whole department an hour a day.
One case taught us something. A client of ours, a small outfit on a monthly plan, left us for a Claude account. Then he bought licences for his colleagues. We do not count it as a loss: he did what most owners do not, give his team an official tool instead of letting everyone rig up their own.
Why bans fail
The reflex is to block chatbot sites on the company network. The next day, usage moves to the personal phone, on mobile data. You have lost the only thing you had: the ability to see.
We have seen this film before, with the CRM. A tool imposed from above, which does not serve the people who have to use it, produces a workaround: the parallel spreadsheet, the notebook, notes on a phone. We described it in Your CRM is useless. Shadow AI follows the same slope, faster.
An employee using AI on the quiet is not at fault. They found a way to work faster and nobody offered them better.
What shadow AI actually costs you
The data. A client name, an amount, a contract clause pasted into a free tool goes to a vendor with whom you signed no processing agreement. For personal data, Article 28 of the GDPR applies, and it is the company that is exposed. We set out the method in GDPR and AI assistants.
Quality. A generic chatbot knows neither your price list, nor your references, nor your contractual commitments. It fills the gap with something plausible. The plausible goes to the client under your logo.
Know-how. The employee who found the right way to write a site report in 5 minutes keeps the method on their phone. The day they leave, it leaves with them.
Taking back control in 4 weeks
Week 1, the inventory. Ask each team which AI tools they use, what for, with which data. Promise in writing that no declared use will be punished. Without that promise, you will only hear about the uses already approved.
Week 2, the official tool. Choose a paid tool, with a contract, known hosting and a written commitment that your data trains no model. Our questions to ask your AI vendor sort it out in one meeting.
Week 3, the one-page rule. What may go into the tool, what must never go in: client names, unpublished figures, signed contracts. One page, not a 20-page policy nobody will read.
Week 4, pooling. The time-saving uses reported in week 1 become team tools. That is the job of an assistant forged on your documents: one employee’s method becomes the department’s, with sources cited and a log you can read.
Three questions to ask this week
No audit, no monitoring software. Three questions, asked in a team meeting, with the promise that nobody will be punished.
The answer to the second tells you your level of risk. The answer to the third tells you which assistant to start with.
What is shadow AI?
The use, by employees, of generative AI tools the company neither chose nor framed: free chatbot versions, apps on personal phones, browser extensions. The main risk lies in the documents and data they paste into them.
Is shadow AI illegal?
Not in itself. The legal risk comes from the data: pasting clients’ personal data into a tool with no processing agreement under Article 28 of the GDPR exposes the company, not just the employee.
Should ChatGPT be banned at work?
Banning without an alternative moves usage onto personal phones. Provide a framed tool first, write a one-page rule, then ban what remains outside it.
How do I find out which AI tools my staff use?
By asking, with the commitment that no declared use will be punished. Network log analysis comes second: it misses personal-phone usage and breaks trust if it comes first.
Is an AI policy enough?
No. A policy with no tool provided stays a document. A one-page rule paired with a tool teams prefer to their free versions changes practice.
Sources
- Cybermalveillance.gouv.fr, recommendations for framing generative AI in business, 14 September 2026.
- Bpifrance Le Lab, « Les entreprises françaises et l’IA », 2024.
- Forgeron3 field observations with client companies, 2025 to 2026.
Bring the list of AI tools your teams say they use. We show you the assistant that takes those uses onto your own documents: what it reads, what it cites, and where it stops.
Request a demo→