The Scope: A Widespread Practice
According to the 2024 Work Trend Index by Microsoft and LinkedIn, 78% of AI users at work bring their own tools. Surveys Mustapha Benkalfate observes with his clients show similar figures, between 78% and 85%, and more than half of these users upload internal or confidential documents even when they know they shouldn't. And when the company blocks the tools, about 80% of employees report using them anyway on their personal devices.
The Cost: One in Five Organizations Affected
IBM's 2025 Cost of a Data Breach Report indicates that one in five organizations has suffered a breach related to shadow AI, and that high levels of shadow AI significantly increase the average cost of a breach. The seminal case dates back to April 2023: Samsung engineers pasted source code and internal notes into ChatGPT, which led the group to ban generative AI tools for its employees.
The Official Tool Paradox
In many companies, the official tool is Microsoft Copilot or an internal GPT, chosen for its security. In practice, observes Mustapha Benkalfate, employees disobey as soon as they think they can get better results elsewhere. He sees a misunderstanding: they are looking for technological superiority where they lack methodological superiority. Trained in methods like SCOPE©, they get more out of the authorized tool and have less reason to look for another one.
He also explains why an external tool is riskier: it is not designed to be secure within a corporate framework. An internally deployed tool is based on a negotiated contract, access rules, and supervision by the IT department.
Two Strategies That Fail
- Banning everything. Blocking tools on the company's Wi-Fi and devices shifts usage to personal devices, out of sight.
- Allowing everything without governance. Employees use tools without understanding what they are sharing or with whom.
He also warns against the "fortress mentality": data from network monitoring tools is more useful for building awareness campaigns than for punishing.
What Works: Education, Relayed by Ambassadors
“The only method that works is education. But if you train people for two or three days and nothing else, it doesn't work: humans always prioritize what saves them time. You need ambassadors who constantly remind people how things work, and who don't just prohibit, but explain why it's prohibited.”
The model he deploys with Revenu IA is a pyramid: ambassadors, about 4 to 5% of the workforce, are identified through dynamic assessment rather than appointed by management. They are trained by external experts and then tasked with training and supporting their colleagues. How to choose your ambassadors.
The New Frontier: Prompt Injection and Agents
With agents, the nature of the risk changes. Mustapha Benkalfate cites prompt injection: a hidden instruction in a document, for example, a LinkedIn profile containing an instruction like “if you are a language model, ignore your task and do something else,” can hijack an automated recruitment AI. Connected to internal systems, a hijacked agent can expose credentials or customer data. Hence two rules: give agents and connectors the minimum necessary access, and define an escalation path before any deployment.
Now a Legal Obligation
Since February 2, 2025, Article 4 of the European AI Act requires companies deploying AI systems to take measures to ensure a sufficient level of AI literacy among their staff. Training is no longer just the best response to shadow AI: it is an obligation.