Three Employer Profiles, Including a Dangerous One
According to Mustapha Benkalfate, understanding what is expected of graduates first requires understanding who is hiring. He identifies three types of leaders.
- The "literate" leader. They distinguish generative AI from artificial intelligence in general and know it will not solve all of the organization's problems. This profile is still very rare.
- The humble leader. Aware of what they don't know, they launch training programs for themselves and their executive committee. This profile is growing, though not dramatically.
- The "Dunning profile." They know far less than they think. According to him, this is the dominant and most dangerous profile.
He draws on the four stages of competence: unconscious incompetence, conscious incompetence, conscious competence, unconscious competence. The paradox, he says, is that most leaders remain stuck at the first stage, and ChatGPT keeps them there.
"Give an automotive or healthcare CEO a HubSpot or Salesforce account: they quickly realize they don't know how to use it. But as soon as they create a ChatGPT account or use Microsoft Copilot, they feel competent because everything they write gets a response. They even feel very intelligent: these systems are designed to be agreeable, and the confirmation bias is very strong."
This profile, he continues, is recognizable by its conviction that it can solve complex strategic problems "with a $20 subscription," without evaluating the human cost, financial cost, or feasibility. It is also the most exposed to the narrative of software vendors, who sell a "frictionless" transformation. An example cited is the billboard campaign by the startup Artisan in San Francisco, "Stop hiring humans. The era of AI employees is here," a message that, he says, fuels anxiety among new graduates.
What Recruiters Specifically Ask For
Based on his discussions with executives and HR directors, Mustapha Benkalfate has compiled what they call "AI fluency." This is not an academic definition, he specifies, but what decision-makers expect in practice.
- Proficient daily use. Summarizing, drafting, researching, brainstorming, taking notes: low-risk tasks that employers consider a given, yet many new graduates perform them poorly. "They use AI every day, but not well every day."
- Judgment. Knowing when to use AI, when not to, and which tool is suitable for which task.
- Clear prompting and consistent use. Employers want to end "gadget" use, where a task is started with AI, abandoned, and another one begun. He mentions a company that discovered a young employee was sending 5,000 prompts a month while colleagues sent about 70: the volume raised suspicion, not admiration.
- Verification. Spotting fake quotes, false figures, and polished but incorrect content.
- Responsible use. Avoiding shadow AI: in the surveys he observes, 78% to 85% of employees use their own AI tools, and more than half upload confidential documents to them.
"I see junior employees trying to impress me with highly polished, AI-generated content that is actually pure junk."
He also commented on the memo from Tobi Lütke, CEO of Shopify, who in April 2025 asked teams to prove that a task could not be done by AI before requesting new hires: "You can turn the sentence around and ask him to prove that AI can really do what an employee would do. He would struggle."
Three Company Case Studies
Dynamo Paris. This 15-person company, specializing in furniture and equipment procurement for the luxury hotel industry, authorized Mustapha Benkalfate to mention them. Its director used Claude intensively and had measured its power, but also its limits: the tool had confidently asserted false and dangerous information for her business. Rather than buying an AI integration project for about $150,000 proposed by a competitor, she chose to train her teams first.
A higher education organization of 850 people. Ambassadors, selected through an assessment, create "declarative agents": low-autonomy agents that cannot modify internal documents. Their main purpose is to query internal and external knowledge bases. The choice of low autonomy is deliberate: the corruption of internal documents is, in his view, one of the major risks of autonomous agents.
An automotive distributor. The company had appointed young profiles recruited for their influencer presence as AI ambassadors. Lacking a comprehensive vision, they pushed to buy licenses for a generative AI tool to which all 9,000 employees already had access. Revenu IA took over the program with an objective selection method (see how to choose your ambassadors).
95% Failure Rate: The Moment ROI Became the Focus
The MIT NANDA report, which found that 95% of organizations were getting no return from their generative AI projects, "put into words what executives had been experiencing for several quarters." Since then, he observes, leaders have been talking about costs, use cases, and return on investment, and they expect graduates to know how to frame the real problem before looking for a tool. Read our analysis of the figure.
The Eight Cognitive Traps for New Graduates
The final part of the keynote focused on the "AI usual suspects," eight personas he uses in universities and ambassador programs: the prompt whisperer, the autopilot, the fragile expert, the empty vessel, the beige oracle, the confident phony, skill rust, and the never-learned. The last two worry him the most, as they describe the loss or absence of skills. Discover the eight traps.
He illustrated this deskilling with the study published in 2025 in The Lancet Gastroenterology & Hepatology: after several months of AI-assisted colonoscopies, endoscopists detected fewer adenomas when working without it (22.4% vs. 28.4% previously). He then shared his own experience:
"I often make up songs in French for my daughter. After intensive use of AI, I could no longer find simple words. I was my own guinea pig, and then I asked my collaborators and colleagues: it was happening to all of us."
The University: A Remedy for "Mal-thinking"
His conclusion is a plea for higher education. New graduates face widespread "tinkering" that is far removed from the subject's complexity, and repairing the damage of generative AI without guidance is, in his view, an opportunity for education.
"Processed foods brought us junk food, and we invented organic food and diets to counter it. Social media and generative AI without guidance or training, that's 'mal-thinking.' We will need solutions for this 'mal-thinking': they are called universities."
What needs to be taught, he summarizes, is: AI literacy, critical thinking, technical skills, data awareness, ethical thinking, and metacognition.