The Three Questions
- Can it harm my credibility with my colleagues, my manager, or worse, with clients and partners?
- Can it harm my company's credibility? The two are often linked, but not always.
- Is it irreversible? If the AI acts, can it permanently modify my files, my code, my data?
The Decision Rule
A single yes signals a high or irreversible risk. In this case, you write it yourself. AI can then act as a proofreader, coach, or devil's advocate, but never as the delivery mechanism. No “yes” answers: the risk is low, and you can let the AI perform the task with more autonomy.
“If the answer is yes to any of these three questions, you cannot let the AI do it for you. You write it yourself, then you use AI as a proofreader, a coach, or a devil's advocate, but not as a delivery mechanism. It's not complicated, and it works.”
Making Room for Effort
The methodology comes with a three-step workflow: first, write or think for yourself, then delegate, and finally, review with a specific objective. This order is not arbitrary. In the MIT Media Lab study published in 2025, participants who first wrote on their own before using AI showed better recall and stronger brain engagement than those who started with AI. See the studies in detail.
Who Signs Off?
The question of sign-off makes the method concrete. Can an employee who did not participate in the process be held responsible for an outcome? According to Mustapha Benkalfate, no. Thus, in trying to save time, one might have to redo everything: “instead of a productivity gain, we see losses.”
Why Doubt is a Career Skill
Generative AI is probabilistic: a request can work ten times and fail on the eleventh. It is also designed to please, with a confident tone even when it is wrong. The study by Microsoft Research and Carnegie Mellon University presented in 2025 to 319 knowledge workers points in the same direction: the more you trust AI for a task, the less you report exercising critical thinking; the more self-confident you are, the more you do it.
Mustapha Benkalfate has made this his criterion for recognizing truly competent people: “Incompetent people say that AI can do it. Competent people reply that they need to check what AI can do for them here, and not in general.”
Three Examples
- A business proposal for a strategic client. Yes to questions 1 and 2: you write it, then run the text through SCORE©.
- An agent authorized to modify shared files. Yes to question 3: no autonomy without minimal rights and a defined escalation path.
- Summarizing an article for your own monitoring. No “yes” answers: delegation is possible, provided you open the cited sources.