Challenge AI to make it a working tool — not the other way round

What is a language model? What is an agent — and why is it fundamentally different? Where classic automation stops and where AI begins. This module sets the common framework without which all the following ones stay blurry.

How language models really work, hallucinations, silent fabrication. Why blind trust is dangerous in a commercial context. GDPR and data sovereignty.

We prompt the way we speak: recent models no longer demand a formula. What really changes the result — the context provided, the examples given, and the ability to correct in two or three exchanges rather than hunting for the perfect sentence. What is folklore — magic formulas, imposed roles, incantations circulating online. And the honest question: in which professions does prompt precision remain decisive, and why yours is not one of them.

A prompt used once and never reused is lost work. How an organisation keeps its way of saying things — graphic charter, verbal charter, house phrasing — and passes it on to newcomers. Why a text file of prompts is useless: a prompt is judged by its result, and a document shows none. The anatomy of a reusable prompt: the fixed part, the fields that vary. Demonstration of a generator in production.

Customer data, supplier data, personal data: what can be sent to an AI model and what cannot. Digital sovereignty, storage outside the EU, liability in case of a leak. Sovereign tools (Albert, Mistral) and why they change the game for French companies.

Querying a sales file in plain language: what is rising, what is dropping, what sells together. Building the question before expecting the answer. And the decisive reflex — checking the figure the AI gives you, because a model that adds up wrongly never flags it. It is the textbook case of module 01, this time applied to your own data.

What AI genuinely speeds up: rough-hewing, structuring, formatting. What it fabricates as soon as it is left to its own devices: the plausible but false market size, the invented competitor, the statistic with no source, the quotation that does not exist. Getting a model to structure rather than assert. Cross-checking, sourcing, dating. What is public, what is paid, what is out of date.

Building a base of qualified prospects from public sources. Segmentation, scoring, personalisation at scale — and its limits.

Turning visit reports into usable knowledge. Why a CRM gets filled in badly — and how to fix that with AI. The tool is built on a computer and used on the phone, in the field.

The previous module says who to talk to; this one says what to tell them. Assisted writing, adapting the message to the segment, brand consistency from one channel to another. What shows immediately — and what undermines credibility.

AI from your phone, in real situations. Analysing a prospect before a meeting in a few minutes. Finding out about a customer on the fly at a trade fair. Generating a personalised sales pitch during the discovery phase.

Brochures, pitches, presentations: how AI produces generic output, and how to avoid it. Moving from a "made with AI" look that damages the brand to a distinctive, professional piece.

Can a text, image, code or logo generated by AI be protected? What French and European law says. How to protect a hybrid creation and document your creative contribution.

Reminders, quotes, deal follow-up. Where automation saves time, where it wastes it. Identifying the tasks never to hand over to a machine.

Counting the calls, pricing them, measuring the time really saved, and deciding. What a cost per run becomes when multiplied by daily use. The false gains: checking time, rework time, learning time. Some tools built during the course will not pass the test — and that is the most useful result of the session.

The project is carried out as personal work, outside class, on the chosen practical case — allow 12 to 15 hours to build the tool and prepare the presentation. The defence validates what has been learned: thirty minutes of presentation, twenty minutes of questions. The student shows the tool working, then explains what is behind it — structure, automations, API calls. Individual or collective, as you prefer; both formats can coexist within one cohort. The volume depends on the headcount: one defence per student, or per group of four to five in a large cohort. Deliverables are handed in according to a brief provided — application, JSON export, defined file formats — so that the criteria are identical for everyone.