In many companies, the question “should we get into AI?” has already been answered, without management knowing. A salesperson rewrites proposals with a chat assistant. An accountant has supplier contracts summarized. A developer pastes code snippets into an online tool to find a bug. Nobody asked, nobody approved, and everyone saves time.
This is what we call shadow AI: the use of AI tools outside any framework defined by the company. The principle is not new, we already talked about shadow IT for cloud tools adopted without IT approval. But generative AI made it widespread within months, because these tools are free, available from any browser, and immediately useful.
Why shadow AI exists
The first reflex is often to see shadow AI as a discipline problem. That is the wrong diagnosis. If teams use these tools without waiting, it is because they meet a real need: writing faster, summarizing long documents, translating, unblocking a technical task. Shadow AI is first and foremost a sign that the internal offering is not keeping up with demand.
A flat ban solves nothing. Usage moves to personal phones and home computers, becomes invisible, and the company loses any ability to govern it. The risk does not disappear: it just moves.
Shadow AI is not a discipline problem. It is a real need that has not yet found a framework.
What shadow AI really exposes
- Data confidentiality: a customer contract, a salary grid or source code pasted into a consumer tool may be retained and, depending on settings, used to improve the model. The company has no contract with the provider, so no guarantee.
- Compliance: sending personal data to an ungoverned service means processing with no documented basis, no processing agreement, sometimes a transfer outside the European Union. All points that are hard to justify during an inspection.
- Quality: a generated answer can contain plausible errors. Without a review rule, an approximation can end up in a sales proposal, a legal note or a customer email.
- Intellectual property: code, strategy, know-how, what makes the company valuable can leak without anyone noticing.
- Unknown tools: beyond the major assistants, there are browser extensions, meeting transcribers, “AI-powered” features built into other software. They are often the least visible and the least secure.
Start by looking, not by banning
Before writing any rule, you need to know what is really happening. A simple inventory is often enough: an anonymous survey of teams on the tools they use and what for, complemented if possible by an analysis of AI services accessed from the company network. Tone matters: the goal is to understand needs, not to find culprits. A team that fears sanctions will say nothing.
This inventory almost always reveals two things: usage is more widespread than expected, and most of it is legitimate. It also surfaces the few genuinely problematic cases, where sensitive data is flowing, on which to focus the effort.
The four pillars of governed use
- An official alternative: provide one or more tools under a business contract, with no reuse of data and a signed processing agreement. This is the most effective measure: when the approved tool is as convenient as the free one, unsanctioned use drops on its own.
- A short, clear usage policy: one page, not a thirty-page rulebook. Which tools are allowed, which data must never be entered (customer personal data, non-public financial information, technical secrets), and when human review is mandatory.
- Training: explain concretely why these rules exist, with examples from teams’ daily work. A rule that is understood is a rule that is followed.
- A fast approval process: when someone finds a useful new tool, they should be able to get it evaluated in days, not months. Otherwise, they will use it anyway.
A concrete example
In a services company of around a hundred people, the inventory reveals about ten different AI tools, used by roughly half of the teams. Most are used for writing and summarizing. Two uses are problematic: HR has named annual reviews summarized in a free tool, and the sales team uses an extension that reads the entire mailbox.
The answer is not a general ban. The company rolls out an assistant under a business contract, hosted in Europe and available to everyone, publishes a one-page policy, runs a one-hour training session per department, and handles the two sensitive cases specifically: the extension is removed, and HR gets an approved tool with pseudonymization. A few weeks later, most usage has moved to the official tool, and the company finally knows what is being done with its data.
Our recommendation
If you do not have an AI usage policy yet, assume your teams already use these tools. The question is therefore not whether to allow or ban, but to choose between invisible use and governed use. Start with a judgment-free inventory, offer an alternative as simple as the free tools, and set a few clear rules about data. It is a project of a few weeks, and often the best starting point for a serious AI initiative.