# Session 1 - AI Foundations
## Full English script

> **Comment lire ce document**
> `▶ SLIDE n` : tu changes de slide.
> Chaque slide s'affiche d'un coup, en entier : la flèche droite passe simplement à la slide suivante.
> *[Texte en italique entre crochets]* : indication de mise en scène, à ne pas dire.
> Le reste est le texte à répéter.
>
> **Le texte est écrit pour un francophone.** Phrases courtes, mots transparents, pas de verbes à particule, pas de subordonnées longues. Chaque point est une respiration. Ne cherche pas à enchaîner : les silences font partie du texte.

**Blocs**

<!-- deck:session-1-foundations -->
| Bloc | Slides |
|---|---|
| Intro et présentation | 1-5 |
| Generative AI and LLMs | 6-7 |
| Tokens and context window | 8-10 |
| Confidentiality | 11-13 |
| Hallucinations | 14-21 |
| Harness | 22-24 |
| MCP | 25-27 |
| Agents | 28-31 |
| Wrap-up and homework | 32-35 |
<!-- /deck:session-1-foundations -->

---

## Opening

**▶ SLIDE 1 - AI Foundations**

*[Slide d'accueil, affichée pendant que la salle s'installe. Dire bonjour, remercier, puis laisser l'animation tourner quelques secondes sans parler : elle dit déjà l'essentiel de la séance.]*

Good morning everyone. Thank you for being here.

This is our subject today. And this small animation already tells you most of it. We will come back to it.

**▶ SLIDE 2 - Chapter: Introduction**

*[Slide de respiration. Une phrase, pas deux. Laisser la salle poser les yeux dessus.]*

Where these five sessions go, and who is in front of you.

**▶ SLIDE 3 - Five sessions**

Five sessions. Here is the plan.

Today is session one. The foundation. You will understand four points. What these tools are. How they work. Why they sometimes lie. And what you can safely put inside. The next four sessions all start from today.

**▶ SLIDE 4 - Your two speakers**

*[Olivier]* 

My name is Olivier Durand. I am co-founder and Chief Technology Officer at ADALAN. And today, I am not alone.

I build technology companies. For twenty years now.

Today, I help companies make AI useful. In a pragmatic way. In a secure way. And they keep the control of their data.

So, I am not here to sell you a demo. AI demos are brilliant. AI projects are difficult. And the reason is almost never the technology.

*[Etienne]* 

My name is Etienne Galland. I am co-founder at ADALAN.

### *@TODO*

**▶ SLIDE 5 - This is not a technical training**

One point, very clear. This is not a technical training.

You will not write code.

No formula to remember.

And maybe, for some of you, this is not new. You already use these tools. You already know some of these words. That is good.

Today is the broom wagon. Like in a cycling race. The car at the back. It picks up everybody. Nobody stays behind.

Because from the next session, we work. Real workshops, on your own cases. And a workshop only works if everybody starts from the same place.

You will get a mental model. A picture in your head. A picture of the machine, when you type your question.

And with this picture, you are not a passenger any more. You take good decisions alone.

*[Question à la salle, sans slide : la poser sur celle-ci, avant d'enchaîner sur le chapitre. Attendre les réponses, noter deux ou trois exemples au tableau, ils serviront pendant toute la séance.]*

So, a quick question to start. Who used an AI tool this week? For what?

*[Écouter. Rebondir avec :]*

Perfect. I keep your examples. I will use them later.

---

## Generative AI and LLMs [Etienne]

**▶ SLIDE 6 - Chapter: Generative AI and LLMs**

*[Slide de respiration. Une phrase, pas deux. Laisser la salle poser les yeux dessus.]*

Two words you hear everywhere. Let us see what is really behind them.

**▶ SLIDE 7 - It creates. And it predicts.**

Generative AI. And LLM.

Before, AI was a filter. It sorted data. This email is spam. This transaction is fraud. Yes, or no.

Generative AI is not a filter. It creates. Text, images, audio, code. It creates new content. That is the difference.

Today, we stay on text. Text is what you will use, every day. And the tool that creates text has a name. LLM. Large Language Model.

*[Ralentir ici. La phrase qui suit est celle dont tout le reste découle. Renvoyer à l'animation de la slide d'ouverture.]*

And now the only technical idea of the day. Just one. And it is this one.

An LLM predicts the next word.

That is all. You give it a start. It calculates the next word. Then the next word. Then the next. Word by word, to the end of the answer.

So it does not search. There is no database of answers inside. No Google, no library. It calculates. It predicts.

Now, a picture. Imagine a new colleague. This colleague has read the internet. Almost all of it, up to a date. It knows a lot about everything. Very confident. Always an opinion.

But this colleague has three limits. Big ones.

No access to your CRM. Nothing about your clients.

No access to your files. It has never seen your documents.

And it stopped reading at a date. Last month? It knows nothing.

So, that is your starting point. A brilliant generalist. No access to your company. No memory of recent events.

Now, a question. If this is true, why does it work so well? Why does it look so intelligent?

Here is the answer. To predict the next word, and to predict it well, the model must capture how ideas connect. It has no choice. And that looks a lot like intelligence.

So, remember this mechanism. Later today, it will explain every problem we will see.

---

## Tokens and context window [Olivier]

**▶ SLIDE 8 - Chapter: Tokens and the context window**

*[Deux démos en direct arrivent dans ce bloc. Annoncer le titre, ne rien expliquer ici.]*

Two more words. You will see them on every invoice, and on every product page.

**▶ SLIDE 9 - It reads tokens**

Token. And context window.

The model does not read letters. It does not read words. It reads tokens. A token is a piece of a word.

And this cutting is purely technical. A machine cut. For us, a token means nothing. It is not a syllable. It is not an idea. It is just a piece.

Each token have a weight. And the weight gives relation. This is a neuronal network. The token are created during the model training.

But for the model, it is everything. The model only sees tokens. And it predicts the next token. One token, then the next one, then the next one. This is all it does.

*[Démo 1 - montrer le découpage]*

Look. I type the name of a company. You all know this company. And look how the machine cuts it. Small pieces. For one single word.

*[Démo 2 - Avec la faute]*

Now, a second test. A normal word. A word we all write, every day.

*[management]*

One token. Just one. Because the machine knows this word. It is in its vocabulary.

Now the same word, with a small mistake. One letter missing. Look. Three tokens.
The machine does not know this word any more. So it cuts it into pieces.

*[managment]*

Same word, same idea, one letter. And you pay three times more.

| Mot correct | Tokens | Avec la faute | Tokens |
|---|---|---|---|
| management | 1 | managment | 3 |
| meeting | 1 | meetting | 3 |
| customer | 1 | costumer | 3 |
| response | 1 | responce | 3 |

*[Et si le temps le permet, la version francophone : les accents comptent aussi.]*

| Mot correct | Tokens | Sans accent | Tokens |
|---|---|---|---|
| réunion | 1 | reunion | 2 |
| problème | 1 | probleme | 3 |
| développement | 1 | developpement | 4 |

*[Démo 3 - Avec la faute]*

And now, the punctuation. 

*[Thanks]*

This word, one token.

*[Thanks !!!]*

I add three exclamation marks. Four tokens.

Every mark is a token. A comma, a point, a space. Everything counts.

So, one simple rule. Write clean. Correct words, normal punctuation, no decoration.

Not for the style but for two very concrete reasons : 

- First, you use less tokens, less machine compute. A dirty text costs more tokens. For exactly the same question.
- Second, you get a better answer. The model predicts on the tokens. Clean tokens, clean prediction.

And one last thing on tokens. Money.

You pay by the token. Your question, and the answer. Both are counted. A long document costs more. A long conversation costs more.

So the price of an AI project depends on one number: the number of tokens.

**▶ SLIDE 10 - Everything you give it has to fit**

Now, the context window. It is the amount of text the model can hold at one time. Your text and its answer share the same space.

Keep one image in mind: a box. And the box has a fixed volume.

Look inside. First surprise: the box is never empty when you arrive.

First, the system instructions. The rules of the vendor. The tone, the format, and the guardrails: what the model must refuse. How to build a bomb. How to attack a company. Part of that refusal is written here, in the box, before your first word. You never see this text, and you cannot remove it.

Second, the list of tools. Everything the model is allowed to call. Your files, your calendar, a search. Each one is described inside the box, even when you use none of it. And when your company connects a new tool, it arrives in this block. The word for that is MCP. We come back to it in twenty minutes.

Third, the configuration of your company. The role you gave it, the house style, your templates. And the skills: a way of working, written once and given to the model. Same block.

And then, the conversation. Every message. Every document you paste. Every result a tool sends back. All of it stays in the box, from the first word to the last.

The rest is your free space. The space to work, and to write the answer.

So, in a long conversation, the model forgets the beginning. Why? The box is full. The oldest pages fall out.

And the answer must enter in the space that remains. A vendor says: my context window is very large. Ask the real question. How much is still free, after all the rest?

The more you put in, the less room is left for the answer.

---

## Confidentiality

**▶ SLIDE 11 - Chapter: Confidentiality**

*[Le bloc que la salle attend. Annoncer tout de suite qu'on répondra franchement, et que la conclusion ne sera pas « on ne peut rien faire ».]*

Now, the question everybody in this room is waiting for. And we will answer it honestly.

**▶ SLIDE 12 - Three things people mix together**

People mix several questions about confidentiality. I want to separate them.

Number one. Transit. Is my data encrypted when it travels? Yes. Standard encryption. Same level as your bank. Not the problem. Next.

Number two. Storage. Your conversation is stored. There is a history. And some tools now have a memory. So your text of today can return next week. In another conversation.

Number three. Training. And here is the real question. Is my text used to improve the model?

**▶ SLIDE 13 - A simple traffic light**

And to close this part, one simple system. You already know it. It is a traffic light. Green, orange, red. So, before you paste a document, one question. Which colour?

Green. Public or generic content. You rewrite, you brainstorm, you structure an argument, you prepare a generic email.

Orange. Internal content, not sensitive. You can use it. But you anonymise first.

Red. Client names, contracts, negotiated prices, and every document under NDA. Never in a tool not approved. Never.

*[Si la charte interne existe, la citer ici. Sinon, dire :]*

You do not know which tools your company approved? Ask the question this week. Five minutes of work. And a very large impact.

*[Si quelqu'un demande "et si on ne veut rien envoyer dehors ?" : "There are models you can install inside your own walls. We do that at ADALAN. And it is a subject for another session." Ne pas développer.]*

---

## Hallucinations

**▶ SLIDE 14 - Chapter: Hallucinations**

*[Bloc central de la séance. Ralentir à partir d'ici.]*

This is the most important part of the day.

**▶ SLIDE 15 - Hallucination**

You all know the word hallucination. It means: the model produces an answer. The answer is perfectly credible. And the answer is false.

I want to change your vision of this word.

**▶ SLIDE 16 - A hallucination is not a bug**

A hallucination is not a bug.

It is the normal function of the system. Applied to a question with no answer.

**▶ SLIDE 17 - It was never trained to say "I do not know"**

Remember the beginning. The model predicts the next word. It is trained to produce a plausible answer. It is never trained to say: I do not know.

So, you ask a question. It does not know. Does it stop? No. It continues. And it produces the most plausible answer.

And plausible is not true.

**▶ SLIDE 18 - Let me show you**

*[La démo est sur la slide, entière : question, réponse et fabrications en rouge. Lire la question à voix haute.]*

Let me show you. Imagine. One of you asks this question.

*[Laisser la salle lire la réponse en silence.]*

Read it. It knows the company. It knows the format of a framework agreement. It even gives you a number of clause.

*[Pointer les passages en rouge, un par un.]*

Now, look at the red. The date. The amount. The duration. The number of clause. The notice period. The fee. The contract is real. All these facts are false. The model never saw this document. And it never said so.

*[Variante en direct, si tu préfères la jouer en vrai : même question dans une fenêtre de chat à côté, avec le nom d'un vrai client de la salle. La slide reste le filet de sécurité.]*

Look at this answer. Confident. Well structured. Correct format. And completely invented.

And this is the danger. A human who does not know, hesitates. You hear it in the voice. The model never hesitates. The false answer and the correct answer are identical.

*[Plan B, si le modèle refuse de répondre :]*

Interesting. It refused. It is progress, and modern tools are better. But do not count on it. It refuses sometimes. Not always.

**▶ SLIDE 19 - Where the risk is highest**

Precise facts. Numbers, dates, names, references. Every specific detail.

Your company. Because it has no access to your company.

Recent events. Because it stopped reading at a date.

And the situations where you insist. You say: are you sure? Find me a source. And it produces a source. Because a source is also a plausible answer.

**▶ SLIDE 20 - Three habits**

So, what do you do? Here are the habits.

Number one. Separate the two types of task. You give a text, and you ask to reformulate, to structure, to summarise. Low risk: the information comes from you. Now, you ask for a fact, from its memory. High risk. Same tool. Two very different levels of risk.

Number two. Verify every document that goes to a client. Every number. Every reference. You verify the work of a junior colleague? Then verify this one.

Number three. Give the information. Do not ask for it. The facts are in the box? Then it does not invent them. It is the most efficient technique. And it is free.

**▶ SLIDE 21 - Not a source of truth**

The sentence to remember.

It is not a source of truth. It is a way to work with the information you give.

---

## Harness

**▶ SLIDE 22 - Chapter: The harness**

*[Bloc court.]*

A new word now. You will probably never say it out loud. But you need the concept.

**▶ SLIDE 23 - The model is the engine**

The word is: the harness.

The situation is simple. Alone, the model does almost nothing. Text enters. Text exits. It is all.

The rest — the interface, the buttons, the memory, the access to the web, the capacity to read a file, a calculation — all of this is constructed around the model. And this layer has a name. The harness.

The model is the engine. The harness is the rest of the car. Steering, dashboard, brakes, fuel tank.

An engine on a table is not a car. You need the complete car.

**▶ SLIDE 24 - Same engine, different car**

Why is this important for you? Look at the screen. The same question, asked twice. The same model, both times. Two very different answers.

*[Montrer la voie du haut, puis la voie du bas, puis lire la ligne du bas.]*

On top, a public chatbot. Around the model, almost nothing. A chat window. The public web. And nothing else. The answer is fluent, it is generic, and some of the details are invented.

Below, your internal assistant. The same model. But around it: your drive and your CRM. Instructions written by the house. Guardrails. And the obligation to show its sources. The answer is the real contract, with a link to it.

Same engine. Different car. The difference is the harness.

So, a person says: we tested AI, and it does not work. The good question is not: which model? The good question is: what was constructed around the model?

And this is the origin of the value, in a company project. Not the choice of the model. The models change every three months. The value is in the harness. The correct access to your data. The correct instructions. The correct guardrails.

For you, in practice, one consequence. You evaluate a tool? Do not ask: which model is inside. Ask: what can this tool reach?

---

## MCP

**▶ SLIDE 25 - Chapter: MCP**

*[Bloc court. Ne pas expliquer l'acronyme, expliquer ce qu'il change.]*

Next question. How do you connect the model to the tools you already use?

**▶ SLIDE 26 - A USB-C port between AI and your tools**

A standard exists for this. MCP. Model Context Protocol. It is a USB-C port. Between the AI and your tools.

Before, every connection was specific. You want AI in your CRM? You construct a bridge. In your email? Another bridge. Expensive. Slow. And it breaks all the time.

Now, one standard plug. Every compatible tool connects. And it is fast.

The acronym is not important. The consequence is important.

**▶ SLIDE 27 - The AI comes to you**

Here is what that changes. The AI comes to you. In your email. In your CRM. In your shared drive. In your meeting tool.

Read the names on the screen. Google Drive, SharePoint, Notion, Box. Salesforce, HubSpot, Slack, Stripe. GitHub, Jira, Figma, Snowflake. Every one of them publishes an MCP server. Hundreds of other products do too. And an internal tool — one of yours — can publish its own.

*[Demander à la salle quel outil de la grille elle ouvre tous les jours.]*

So there is no separate window to open, and nothing to copy across. The assistant works where the work already is. And it arrives this year. Not in five years.

And there is a direct consequence for our question of confidentiality. The AI is connected to your CRM? Then the question changes. It is not: what do I paste? It is: what can this tool reach, alone? And who decided?

So, the good news and the alert are the same sentence. Soon, you will not copy and paste any more.

---

## Agents

**▶ SLIDE 28 - Chapter: Agents**

*[Dernier bloc de contenu. Annoncer la couleur : c'est le sujet où l'écart entre les démos et la réalité est le plus grand.]*

Last concept. And the one where the demos and the reality are the furthest apart.

**▶ SLIDE 29 - A model, a harness, tools, and a loop**

Agents.

Take all our elements. A model. A harness around the model. Tools connected with MCP. And now, you add one more element. A loop.

The system does not answer one time and stop. It acts. It looks at the result. It decides the next step. And it continues. Until the objective.

That is an agent.

**▶ SLIDE 30 - From instruction to objective**

Here is the practical difference. And it is very big.

Today, you give an instruction. Write an email for this client.

With an agent, you give an objective. Contact all the clients with no answer this week.

And the agent decides the steps. Find the clients with no answer. Read the history of each conversation. Prepare the correct message for each client. And present them to you.

You passed from a task to an objective. From execution, to delegation.

**▶ SLIDE 31 - The honest version**

Now, the honest version.

Today, agents are the domain with the biggest distance between the demo and the reality.

They work well when the task is precise, and the tools are reliable.

They fail when the task is vague. And you cannot predict the failure.

And the errors accumulate. An error at step two? And the machine constructs all the next steps on this error.

This is why every serious project keeps a human at the validation. The agent prepares. The human approves.

Why do I talk about agents in session one? You will maybe not use one tomorrow. But it is the direction of the market. And with four words — model, harness, tools, loop — you can read any announcement, and understand the real promise.

---

## Wrap-up and homework

**▶ SLIDE 32 - The one sentence**

One final sentence. It contains all the session.

Alone, a model knows nothing about your company. And it invents, with confidence. What we construct around the model makes it useful. And safe.

It is the complete session. The rest is detail.

**▶ SLIDE 33 - The words**

Quick recap.

Generative AI creates content. An LLM predicts the next word. Tokens are the pieces it reads, and the pieces you pay. The context window is the box: everything it can see at one time. Hallucination: it does not know, and it answers anyway. The harness is the car around the engine. MCP is the plug for your tools. A skill is a way of working, packaged once. And an agent is all of it, in a loop, for an objective.

*[Distribuer le glossaire FR/EN maintenant, pas avant : sinon ils lisent au lieu d'écouter.]*

I give you a glossary. One page. All these terms, in French and in English.

**▶ SLIDE 34 - Test one real question**

Homework. Very short, I promise.

Before the next session, test one real question. A question from your real work.

And bring me one of two results.

Either a useful answer. And you explain why.

Or a hallucination. And you explain how you detected it.

We open the next session with your examples. Not mine. And it is the moment when all of this becomes real.

**▶ SLIDE 35 - Questions?**

Thank you. Questions?

*[Laisser ce slide affiché pendant toute la séquence de questions.]*

---

## Vocabulaire à répéter à voix haute

Les seuls mots du script où la prononciation peut coincer. Le reste a été choisi pour être transparent.

| Mot | Se dit | Piège |
|---|---|
| **hallucination** | a-lou-si-NÉ-cheunn | pas « allucination » |
| **harness** | HAR-niss | le « h » se prononce, souffle un peu |
| **truth** | *trouth* | dis **« trouf »**, personne ne verra la différence |
| **predicts** | pri-DIKTS | dis **« pri-DIKS »**, laisse tomber le « t » |
| **colleague** | KO-liig | accent sur la 1re syllabe, « -gue » muet |
| **mechanism** | MÈ-ka-ni-zeum | accent sur la 1re syllabe |
| **generative** | DJÈ-ne-ra-tiv | accent sur la 1re syllabe |
| **language** | LANG-gwidj | |
| **audio** | Ô-di-o | pas « aoudio » |
| **plausible** | PLAW-zi-beul | |
| **credible** | KRÈ-di-beul | |
| **guardrails** | GARD-relz | |
| **acronym** | AK-ro-nim | |
| **anonymise** | a-NO-ni-maïz | |
| **clause** | KLOZ | |
| **vendor** | VÈN-deur | |
| **efficient** | i-FI-cheunnt | |
| **objective** | ob-DJÈK-tiv | accent sur la 2e syllabe |

**Le son « th ».** Il ne reste que les mots courts et inévitables : *the, this, that, they, with, three*. Personne dans la salle ne t'en voudra. Si tu bloques, dis un « z » ou un « f » et continue : le sens passe.

## Phrases de secours

À garder en tête si tu perds le fil ou si une question te bloque :

- "Let me say it again, in a different way."
- "Good question. I keep it for the end, I want to finish this point."
- "I do not know. I will check, and I will send you the answer."
- "Is it clear for everyone?"
- "Let me give you a concrete example."