AI at work, without blindly believing in it

This book starts from the actual mechanics to derive a working method. It sells neither enthusiasm nor distrust: it provides the criteria to decide, task by task, if the tool has a place.

The six parts

I. The Mechanics — What a language model really is, in one page. Tokens and context window, and why this explains billing, calculation errors, and the forgetting of instructions in long conversations. The complete typology of inventions — reference, precision, completion, sycophancy — and classified risk areas.

II. The Judgment — The verification rule, which sorts all uses into one sentence. A green / orange / red grid of twenty-five uses. The three questions to ask before automating, including the one no one asks: will the error be isolated or systematic?

III. The Practice — The four factors that genuinely change the quality of a response, in order of impact. Six reproducible instruction patterns, including the source constraint that eliminates most inventions. Folk techniques that are useless. RAG explained from beginning to end, with the five points that fail in practice. A decision tree between instruction, RAG, and fine-tuning.

IV. Industrialize — Build an evaluation set and measure instead of admiring. Design the three levels of human supervision and avoid automatic validation. The monthly cost formula and the five levers to reduce it. Choose a model based on seven criteria, and guard against dependency.

V. The Risks — What you send and where it goes. GDPR and the European AI regulation: what concretely applies to a small structure. Instruction injection explained with an attack example and six protections. Intellectual property and editorial responsibility.

VI. The Usage — Ten sustainable uses, detailed. Eight that cost more than they bring in. How to recognize an unedited generated text, in eight signs. What will remain when the hype dies down.

What you take away

• A library of ready-to-use instructions: document analysis, rewriting, critique, structured extraction, subject mapping
• A 10-point checklist before using a model for a task
• A 12-point checklist before putting a system into production
• The three-color usage grid
• The instruction / RAG / fine-tuning decision tree
• A glossary of 12 terms

For whom

Freelancers, small business leaders, and professionals who already use these tools daily and want to know where the safe zone ends. No technical skills are required for parts I, II, V, and VI; parts III and IV assume you are building something.

For whom it is not

Those looking for a list of "magic prompts": the book explains precisely why such lists do not work. Nor machine learning engineers, who already know everything covered here.

What the book does not cover

Model training, deep learning mathematics, infrastructure choice, image and video generation, and a detailed analysis of the European AI regulation — the latter requires a lawyer, not a book.

Format

PDF, 24 chapters and 4 appendices. Immediate access after purchase.

Models evolve quickly: this book focuses on mechanical and methodological principles that survive across versions, never on the settings of a particular product. No quantified productivity gains are presented — those circulating are unverifiable. Legal points are provided for guidance and are subject to change.