Session 1 of 5

AI Foundations

Thank you for your
[Laisser tourner l'animation quelques secondes avant de parler. Elle montre déjà tout le message de la séance.] "Good morning everyone, and thank you for being here."

Your two speakers

Olivier Durand
Co-founder and Chief Technology Officer, ADALAN

Twenty years building technology companies.

I help companies make AI useful: in a pragmatic way, a secure way, and in a way where they keep control of their own data.

Far from the demos that shine, and the projects that get stuck.

Etienne Galland
Co-founder, ADALAN

A career in consulting, with financial services as home ground.

I know how these businesses actually run: the constraints, the vocabulary, the way a decision gets made.

Olivier knows what the technology can do. I know what your business will accept.

AI only creates a real advantage when teams change how they work. That is why you are in this room.

[Ne pas dérouler les CV. Trois lignes chacun suffisent : l'ancienneté, la posture, le scepticisme nommé avant eux. La dernière ligne est dite à deux voix, c'est la raison de leur présence.] [La troisième ligne d'Etienne est le moment du slide : elle nomme la complémentarité au lieu de la laisser deviner. La dire en se regardant, pas en regardant la salle.] [Garder la souveraineté et l'open source pour le bloc confidentialité, où ça vaudra beaucoup plus cher.] "We are French, so our English is not perfect. If we say something strange, or if we go too fast, stop us. Really. Raise your hand."

Five sessions

Session 1The foundations
Session 2Your own examples
Session 3Prompting in practice
Session 4Tools in your workflow
Session 5Building your habits

Everything we do in the next four sessions builds on today.

[Titres des séances 2 à 5 à ajuster une fois le programme complet arrêté.] "Let me tell you what these five sessions are about."

This is not a technical training

  • You will not write code.
  • You will not need to remember any formula.
  • You will get a mental model: a picture in your head of what happens when you type something into one of these tools.

Once you have that picture, you stop being a passenger.

"You will not write code. You will not need to remember any formula. What you will get is a mental model. Once you have that picture, you stop being a passenger. You start making good decisions on your own."
Over to you
Who used an AI tool this week?
For what?
[Attendre. Ne pas remplir le silence. Noter deux ou trois exemples au tableau, ils serviront de matière pendant toute la séance.] "Good. Keep those examples in mind. I will come back to them."

Two words you hear everywhere

The AI we had before

Sorted things.
This email is spam.
This transaction is fraud.

Generative AI

Creates something new.
Text, images, audio, code.

Today we focus on text. The tool that generates text is called an LLM: a Large Language Model.

"Generative AI simply means AI that produces content. That is the difference with the AI we had before."
The only technical idea to keep
An LLM predicts
the next word.
  • It is not looking anything up.
  • There is no database of right answers inside it.
  • It is not searching. It is predicting.
[Ralentir ici. C'est la phrase dont tout le reste découle. Renvoyer à l'animation du slide 1.] "You give it a beginning, and it computes what is most likely to come next. Then it does it again. Word by word."

Imagine a colleague

Who has read almost everything published on the internet, up to a certain date. Enormous general knowledge. Very confident. Always has an opinion.

  1. 1No access to your CRMIt knows nothing about your accounts.
  2. 2No access to your filesIt has never seen a single document of yours.
  3. 3Stopped reading at a dateIt does not know what happened last month.
"A brilliant generalist, with no access to your company, and no memory of recent events. Keep the mechanism in mind, because in ten minutes it will explain everything that goes wrong."

It does not read words.
It reads tokens.

A token is a piece of a word.
Photo: Declan Sun / Unsplash
[Enchaîner immédiatement sur le tokenizer en direct, avec le nom d'un vrai client de la salle. Voir un nom connu se casser en trois morceaux fait le travail à ta place.]

Why this matters to you

  1. 1MoneyThese tools are billed by the token, in and out. Long documents cost more. That is the whole story.
  2. 2Strange behaviourIt never saw the letters. So it is bad at counting them, or spelling backwards. Not low intelligence. Different eyes.
"When someone tells you an AI project costs a certain amount per month, the number of tokens is what drives that cost."
The context window
A desk.
Not a filing cabinet.
  • Whatever is on the desk, it can see.
  • Whatever is not on the desk does not exist.
  • The desk has a fixed size.
Photo: Andrej Lisakov / Unsplash
"When you have a long conversation and it seems to forget what you said at the beginning, that is the desk being full. The oldest things fall off."

Everything you give it has to fit

systemthe tool's own rules
toolswhat it is allowed to do
your setuprole, house style
the conversationevery message · every document you paste · every result
…room to work
0context limit →

The box is never empty when you start. And the answer still has to fit in what is left.

[Faire apparaître les blocs un par un. Le point qui compte est le dernier : la place restante n'est pas la taille annoncée de la fenêtre.] "Notice that a large part of the box is already taken before you type a single word. The system instructions, the tools, the way your company configured it. Then everything you paste. And the answer has to fit in what remains."
The context window is a box
And you are
the one packing it.
  • Everything you add pushes something else out.
  • Two pages of the right document beat forty pages of the wrong one.
  • When the subject changes, open a new conversation. A clean box.
[C'est le passage d'une métaphore passive - un bureau - à un geste actif. La salle doit sortir d'ici avec l'idée qu'elle décide du contenu de la boîte.] "Packing that box is a real skill, and it is the one that separates people who get good answers from people who do not. We will work on it in session 3."
If you want a good answer,
put the right things on the desk.

Your context. Your constraints. Your examples. The tone you want.

Most bad results come from an empty desk, not from a weak model.

"That is why the same tool gives one person mediocre answers and another person excellent ones. It is often not the model. It is what was placed on the desk."
The obvious next question
Where does
that desk live?
Photo: Alex Duffy / Unsplash
[Transition directe depuis la fenêtre de contexte. C'est la question numéro un d'une salle de commerciaux, autant la prendre de face.]

Three things people mix together

  1. 1TransitIs it encrypted while it travels? Yes. Same as your online banking. Not the problem.
  2. 2StorageYour conversation is kept. Some tools now remember across sessions. What you write today can come back next week.
  3. 3TrainingWill it be used to improve the model? This is the real question.
[Passer vite sur 1 et 2, ralentir sur 3.]
And here is the point
The answer does not depend
on the technology.
It depends on the contract.

The same AI does not follow the same rules depending on how you access it.

A personal account on your phone and the tool your company deployed can be the exact same model, with completely different terms.

"Consumer accounts and business accounts are not the same world."

Two risks that are yours

Client confidentiality

Pasting a contract or a specification into an unapproved tool can break the NDA you signed with that client.

The risk is contractual, not technical.

Personal data

A contact list. A meeting report with names. A CV.

That is GDPR, and the responsibility goes up to the company.

[Ne pas faire de cours RGPD. Deux risques concrets, formulés en langage commercial, suffisent.]
One sentence, ninety percent of situations
If you would not send it by email
to an outside supplier,
do not paste it into a tool
your company has not approved.
[La faire répéter, ou au moins marquer un silence de trois secondes après.]

So clean it first

Replace the client name, the amounts, the company names. Thirty seconds of work.

The model does not need the real names to help you with structure, wording, or argument.

The answer is not "you can do nothing". The answer is "anonymise".

[Démo en direct : extrait de proposition commerciale, anonymisation, relance, comparaison. C'est ce qui désamorce l'objection "alors on ne peut rien faire".]

A simple traffic light

GreenPublic or generic content. Rewriting, brainstorming, structuring an argument, drafting a generic email.
OrangeInternal content that is not sensitive. Use it, but anonymise first.
RedNamed client data, contracts, negotiated pricing, anything covered by an NDA. Not in an unapproved tool.
[Si la charte interne existe, la citer ici. Sinon : "that is a question worth asking this week".] [C'est ici que ton expertise open source et on-premise vaut le plus. Si quelqu'un demande "et si on ne veut rien envoyer dehors ?", tu peux répondre en une phrase : "There are models you can run inside your own walls. That is part of what we do at ADALAN, and it is a subject for another session." Ne pas développer, tu perdrais la salle.]
The most important part of today

Hallucination

Something that sounds completely credible, and is simply false.

Photo: Juli Kosolapova / Unsplash
"I want to change how you think about it."
A hallucination
is not a bug.

It is the normal behaviour of the system, applied to a question it does not have the answer to.

[Phrase pivot de la séance. La dire lentement, puis se taire.]

It was never trained to say
"I do not know"

It was trained to produce a plausible answer.

So when you ask something it does not know, it does not stop. It does what it always does. It produces the most plausible continuation.

Plausible is not the same thing as true.

"Remember what we said at the beginning: it predicts the most likely next word."
Live
Let me show you.

Confident. Well structured. Correct format. Completely invented.

A human who does not know usually hesitates. You hear it in their voice. The model never hesitates. The false answer and the true answer look exactly the same.

[Démo : demander une information vérifiable de leur métier qu'il ne peut pas connaître (chiffre interne, référence client, clause d'un contrat récent).] [Plan B si le modèle refuse : "Interesting, it refused. That is progress. But do not rely on it. It refuses sometimes, not always."]

Where the risk is highest

  • Precise facts. Numbers, dates, names, references, quotes.
  • Anything about your company. It has no access to it.
  • Anything recent. It stopped reading at a certain date.
  • Anything where you push it. Ask for a source and it will often produce one. A source is a plausible continuation too.
[Le quatrième point surprend toujours. Insister dessus.]

Three habits

  1. 1Separate the two types of taskReformulate, structure, summarise what you provided: low risk. Retrieve facts from its memory: high risk. Same tool, different risk.
  2. 2Verify anything a client will seeEvery number, every reference. Same check you would run on a junior colleague.
  3. 3Give it the information instead of asking for itIf the facts are on the desk, it does not have to invent them.
"The third one is the single most effective technique, and it costs you nothing."
It is not a source of truth.
It is a way of working with
information you provide.
[Ligne à retenir du bloc. Marquer une pause avant d'enchaîner sur le harness.]
A word you will never say out loud
The model is the engine.
The harness is the car.

An engine on a table is not transport.

Photo: Garett Mizunaka / Unsplash
"The model on its own takes text in and produces text out. That is all."

Same engine.
Different car.

Two products can use the exact same model and give you very different results.

So when someone says "we tested AI and it did not work", the question is rarely which model. It is what was built around it.

Do not ask what is inside. Ask what it can reach.

MODEL
interface
memory
web search
your files
guardrails
instructions
"This is where most of the value is created in a company project. Not in choosing the model, models change every few months anyway."
MCP
A USB-C port between
AI and your tools.

Until recently, every connection was custom. One bridge per tool. Expensive, slow, and it broke constantly.

One standard plug. Any tool that supports it connects quickly.

Photo: Mika Baumeister / Unsplash
"I do not need you to remember the acronym. I need you to understand what it changes."

Until now

You go to the AI. Open a window. Copy your text in. Copy the answer back.

What is happening now

The AI comes to you. Into your email. Your CRM. Your drive. Your meeting tool.

This year, not in five years.

AI CRM Mail Drive Chat
[Faire le lien explicite avec le bloc confidentialité : quand l'IA est branchée sur le CRM, la question n'est plus ce qu'on colle, mais ce que l'outil a le droit d'atteindre tout seul.]
Agents
A model, a harness,
tools, and a loop.

It does something, looks at the result, decides what to do next, and continues. Until the objective is reached.

Photo: Homa Appliances / Unsplash
"Take everything we have seen, and add one more thing: a loop."

Today, you give an instruction

"Write me a follow-up email for this client."

With an agent, you give an objective

"Follow up with the clients who did not answer this week."

You moved from asking for a task to delegating a goal.

plan act observe adjust
"It decides the steps. Check who did not answer. Read the history. Draft the right message for each one."

The honest version

  • This is the area with the biggest gap between the demos and reality.
  • They work well when the task is well defined and the tools are reliable.
  • They fail unpredictably when the task is vague.
  • Mistakes compound. A wrong step two becomes steps three, four and five.

Almost every serious deployment keeps a human at the validation point. The agent prepares. The human approves.

"The reason I am telling you this in session one is that it is the direction everything is moving."
One sentence ties it all together
A model on its own knows nothing
about your company, and will
invent things confidently.

What we build around it is what
makes it useful and safe.
[C'est la phrase que tu veux les entendre répéter dans le couloir. La dire, puis se taire.]

The words

Generative AICreates content, instead of sorting it.
LLMPredicts the next word.
TokenThe pieces it reads, and what you pay for.
Context windowIts desk. A box you pack.
HallucinationWhat happens when it does not know and answers anyway.
HarnessThe car around the engine.
MCPThe plug that connects it to your tools.
AgentAll of it, in a loop, working toward an objective.
[Distribuer le glossaire papier FR/EN ici, pas avant : sinon ils lisent au lieu d'écouter.]
Before next session

Test one real question
from your own work.

Either

A result that was useful. Tell me why.

Or

A hallucination you caught. Tell me how you spotted it.

We open the next session with your examples, not mine.

"That is when this becomes real."

Questions?

Thank you.

[Laisser ce slide affiché pendant toute la séquence de questions.]
40:00
NOTES
← → navigate  ·  O overview  ·  N notes  ·  T timer  ·  F fullscreen