The Mass Replacement Narrative Doesn't Hold Up
Mustapha Benkalfate begins by dismantling the dominant narrative. He sees no evidence today of a mass replacement of employees by AI. Layoffs are happening, but their cause is primarily financial.
"Technology is used to justify layoffs and send a signal to the markets, and this predates AI. Some developers are replaced for certain tasks, but they become very useful for others, such as AI supervision."
He attributes this narrative to the funding needs of major AI players, who are asking investors for trillions of dollars. Having failed to convince them they will make money, they promise to save it. Hence his motto, which is also that of Revenu IA: "the value of AI lies in the humans who know how to use it."
Owning the Car Doesn't Mean Knowing How to Drive
Generative AI is now everywhere, even in WhatsApp, and its penetration in organizations is approaching 90%, he says. The question then becomes one of operation: knowing how to write a prompt, how to use one's brain while using AI, and how to use it safely. "If you buy a car and end up in the hospital after the first drive, was the investment worth it?"
He points out that generative AI is probabilistic: it can work ten times in a row and produce a serious error on the eleventh. It is also designed to please, with a confident tone even when it's wrong. From this, he derives a career rule: content produced with AI is only valuable if its author can defend it. He cites the case of an important report, perfectly structured and written, but entirely false; once forced to be accurate, it became correct but unusable, and its author could defend it in neither case. The simplest remedy: peer review.
Metacognition: Thinking About Your Own Thinking
Metacognition involves observing your own way of thinking while working with AI. Many training programs have focused on mastering the tools, neglecting how our brains change upon contact with them. He references the MIT Media Lab study published in 2025: participants who wrote with ChatGPT showed lower brain connectivity than those who wrote alone. But those who first wrote themselves, then used AI to improve their text, showed the opposite effect. Details of the studies are on our cognitive debt and AI page.
He also describes a symptom that anyone can verify: abandonment. You start a task with AI, you're not satisfied, you give up, you start over elsewhere, and you end up with dozens of open conversations on the same topic. "When we talk about productivity, that is unproductivity."
The usual advice, "keep a critical mind," is too vague to be effective, he argues. He prefers a reproducible routine: the méthodologie du doute©, three questions to ask before any task.
"My obsession is that you always remain the driver. I don't want humans to become dependent on AI. Some people are proud of a result produced with AI, but their contribution was so small that the AI should be proud, not the human."
Four Human Strengths That Are Gaining Value
- Judgment. Deciding what to delegate to AI, and recognizing that it's sometimes faster to do it yourself than to verify.
- Communication. Without the ability to explain and defend what you produce, there is no accountability.
- Ethical thinking. Every use case involves the organization. According to IBM's 2025 Cost of a Data Breach report, one in five organizations has experienced a breach related to shadow AI.
- Adaptability. The tools are constantly changing. "What survives is human intelligence built on solid foundations: method. Methods survive."
Why Maintain Friction
Focusing solely on the outcome is a modern mistake, he argues, because the process builds invisible skills: writing an important email yourself is also preparing for the subsequent meeting. The worst combination for heavy users is rising confidence while skills erode. He illustrates this with an analogy:
"Imagine I win the Kuala Lumpur professional tennis tournament by poisoning each of my opponents. No one would congratulate me. We congratulate the winner because we know everything it took to get there. It's the same thing with your brain."
Finally, he poses the question that increasingly worries organizations: who signs off? If you haven't participated in the process, can you take responsibility for the outcome?
What Works: Making Value Collective
If 95% of AI projects fail, it's because the value remains individual: everyone claims to save time, but the organization sees no benefit. Value becomes collective when prompting methods, assistants, agents, and honest feedback—including on what no longer works—are shared. He commends the host organization's culture of peer review, championed by its senior management, and concludes with his formula: "Responsible doesn't mean slow: it's strategic."
Questions from the Audience
Should we test every new tool? No, optimizing the ones you already have yields more. He cites Excel's =COPILOT() function, which allows you to interact with a sheet's data. "If a tool survives for three months, you'll hear about it." He also warns against influencers who praise amazing tools... for two-person companies.
Are MCP connectors secure? The risk always exists. It is reduced by giving minimal access to connectors and using a managed environment aligned with the IT department.