Artificial intelligence does not need to be a fully general intelligence to wield significant power. For decades, the focus has been on how intelligent a machine could become. However, the emergence of AI agents has shifted the conversation to a more pressing issue: how prepared we are to allow them to act on our behalf. An AI does not need to be conscious or even general in its intelligence to have considerable influence. It only needs the right permissions. In 1950, Alan Turing suggested evaluating a machine's intelligence by observing its behavior rather than asking if it can think. A few years earlier, Isaac Asimov explored this from a different angle through his stories about robots obeying simple laws. His work highlighted how even simple rules can become complex when faced with the real world's ambiguities. One question focused on behavior; the other on limits. Seventy years later, a third question has emerged: what happens when we give a machine not just the ability to respond, but also the ability to act? This may become more critical for companies than the ongoing debate about the arrival of general artificial intelligence (AGI). For a long time, much of the AI discussion has revolved around AGI—the idea of a machine that can reason and learn across all domains like a human. Questions such as "When will it arrive?" or "Will it be more intelligent than us?" are fascinating. But they can distract from what is already happening: companies are being transformed by AI without needing a fully general intelligence. A company does not need a machine capable of reasoning like a human on all subjects. It needs a system that is sufficiently competent in specific tasks, connected to the right tools, and authorized to perform certain actions. This is a crucial distinction. Consider an AI that cannot cook or drive a car but can analyze advertising campaigns, modify a budget, generate new ads, and measure their performance. While not general, if entrusted with a 500,000 euro media budget, it unquestionably has economic power. An AI does not need to be AGI to have power. It needs capabilities, permissions, and autonomy—and it may be this combination that we should begin to measure. We first used generative AI as assistants: "Summarize this document," "Write me this program," "Analyze this table," "Propose a response to this customer." In these cases, the machine produces something, while the human remains the one who decides what to do with it. But gradually, some permissions are added. The AI no longer just proposes a response to a customer; it sends it. It no longer just recommends a change to a program; it opens the pull request, merges it, and triggers the deployment. The same model, the same company—but the nature of the system has completely changed. The difference is not mainly in its intelligence, but in the mandate we have given it. For instance, in a customer service department, an AI might start by analyzing a case and suggesting a response. Then it responds directly. Then it applies a commercial gesture under 20 euros. Then it chooses the compensation within a budget limit. Finally, it identifies the causes of complaints and modifies operational rules to reduce them. We have moved from a text generator to a system to which the company delegates a small part of its decision-making power. The evolution of AI agents is also evident in fields like marketing. A marketing manager might first ask an AI to "analyze the performance of my campaigns and make recommendations." Then, "Prepare the campaigns you recommend." Then, "Create them directly in the advertising platforms." Then, "Reallocate the budget daily to the most performing campaigns." Finally, "Here is my acquisition cost target and monthly budget. Optimize it." The last instruction contains only a few additional words, but it changes everything. The human no longer gives a task. He gives an objective and lets the machine choose the path to achieve it. AI agents are interesting because they no longer just execute simple instructions. They can break down objectives, use multiple tools, observe results, and decide on the next steps. Each new tool connected to them expands their action space—CRM, ERP, messaging, CMS, document base, advertising platforms, development tools, payment services, logistics systems. The question "What can the AI do?" becomes inseparable from another: "What does it have access to?" The "AI developer" who once only wrote code is now part of a broader evolution. A model can receive a functional request, explore a code base, identify the files concerned, write the fix, run tests, and modify its proposal based on the results. What matters is not just the quality of the code, but the length of the chain of actions we are willing to let the machine accomplish. The difference between "Write me this function" and "Fix this bug" is significant. In the second case, the system must begin to determine what it needs to inspect and modify. We could even go further: "Fix this bug and deploy the correction if all the tests pass." At this point, we introduce a delegation rule. This kind of evolution could have economic consequences that extend far beyond individual productivity. If creating internal applications, automating workflows, or developing features becomes much less costly, the traditional trade-off between buying software and developing it could shift. Some of the software that companies buy today because they lack the means to build it could be generated, adapted, and maintained more easily in the future. This does not mean the disappearance of software publishers, but it could change the boundary between standard software, specific development, and automation.