The skills you need to use AI effectively

Generative AI has become part of everyday life in businesses, universities and the media. Yet, many people use these tools each day without really understanding how they work. This matters because it shapes the quality of the output, the way decisions are made and ultimately, the value AI can create.

The issue is not AI itself, but the way it is often used: quickly, intuitively and with little understanding of its underlying logic. Using a language model without grasping its basic principles is a bit like driving without full knowledge of the rules of the road. You may still get where you want to go, but with less efficiency and control, as well as a greater likelihood of wrong turns.

To use AI effectively, the following four skills are becoming essential.

• Understanding model architecture to choose the right tool for the task

Not all models are equal. And users can’t make informed choices if they are unaware of the differences between deep reasoning models, frontier models (such as the current Claude Opus, GPT-05, or Gemini Ultra) and lightweight models optimised for speed and cost. 

Frontier models use multi-step reasoning (chain-of-thought) capabilities, making them suitable for complex analytical tasks, such as strategic analysis, contract review and organisational diagnosis. Conversely, for a task involving rapid rephrasing or classification, a lighter model will be quicker, more energy-efficient and economically rational.

Understanding this granularity, knowing that every call to a large model consumes significant computational and energy resources, and that this consumption must be proportional to the value produced, is a sign of professional maturity in itself.

• Knowing how to orchestrate multiple AIs, rather than using just one

Value no longer lies in using a standalone tool, but in the ability to build intelligent workflows. According to McKinsey’s 2025 The state of AI survey, 23 per cent of organisations have already begun deploying agent-based systems at scale in at least one function, systems capable of learning, remembering and acting autonomously within defined parameters.

In practical terms, this means learning to connect a document analysis model, a synthesis agent and a fact-checking tool, via interoperability protocols such as the Model Context Protocol (MCP) or the Agent-to-Agent (A2A) standard to produce an analysis that none of the three could generate on its own. These protocols enable specialised agents to work in coordination rather than relying on monolithic systems. A user who knows how to design such workflows has a distinct advantage.

• Knowing how to configure AI to regain control

This is perhaps the least recognised skill, yet one of the most crucial. An unconfigured model tends to produce verbose responses, trigger unnecessary iterations and adopt a conversational mode which, while reassuring for novice users, is time-consuming and costly in terms of time, tokens and energy.

Knowing how to write a clear prompt, setting the right tone, defining the format and explaining exactly what is needed makes AI far more effective as a working tool. The same applies to tools such as Claude Code, which allow developers to interact directly with their coding environment to generate code or automate tasks more efficiently. The clearer the instructions, the fewer unnecessary exchanges, revisions or repeated requests for clarification, which also means lower energy consumption.

• Adopting a critical stance and agility in the face of technological disruptions

AI generates output fluidly and sometimes with a false sense of confidence. Hallucinations, confirmation bias and factual inaccuracies are structural flaws in current models. Nearly a third of respondents to the aforementioned McKinsey survey, conducted among 1,993 companies in 105 countries, reported having suffered negative consequences linked to the inaccuracy of AI within their organisations. Knowing how to evaluate a result, cross-check it against other sources and identify what the model failed to do involves exercising judgement that the machine cannot exercise on our behalf.

AI tools are evolving at an extraordinary pace and many of the platforms dominating today may not last more than 18 months. That is why students need more than technical familiarity; they need to understand the concepts behind these systems and how to question them critically. In France, 82 per cent of 5,069 student respondents to the 2025 Conference des Grandes Écoles (CGE) survey said they wanted more AI training from their institution.

An institutional responsibility

Bridging the gap between technical understanding and broader education is now an imperative. Training professionals who understand the nature of AI, its real capabilities, limitations, costs and internal logic, is to equip them with the means for true autonomy. It is also the best safeguard against passive or uncritical use. The future will not belong to those who use AI the most, but to those who truly understand how it works.