- One LLM is an engine that predicts text. Everything else is built on top.
- One agent is an LLM who has been given tools and the right to complete steps alone.
- The RAG allows the model to answer from your own documents instead of relying on its general training. This is what makes AI useful in a business setting.
- One hallucination is an invention delivered with confidence. This is not a temporary fault, it is an operating characteristic.
- One jeton is the billing unit. Your costs are calculated in tokens, not in questions asked.
- The fine tuning is almost always the wrong answer for an SME. The RAG solves the same problem for much less money.
Why this Guide
You're in contact with a provider. It follows: agent, orchestration, RAG, context window, fine setting, crazy guard. You're nodding your head. You cannot judge what he proposes.
This guide fixes this problem. Each term receives a simple definition, a concrete image, and above all the only thing that really interests you: what it changes for your business and your invoice.
Read it once. Keep it on hand for your next meeting.
Basic terms
Artificial intelligence
Umbrella term that covers almost everything, which makes it almost useless in a meeting. When a provider says "our solution uses AI," he said nothing. Ask him what model, to do what, on what data.
What it changes for you: Nothing until the next question is asked.
Machine learning
A system that improves its results from examples rather than from rules written by a human. It's the technical family. Your spam filter has been one of them for 20 years.
What it changes for you: It reminds you that AI did not appear in your company in 2023. She was already there, discreetly.
Large Language Model, or LLM
The engine behind ChatGPT, Claude, Gemini and company. He read an immense amount of text and became very good at predicting which sequence of words answers a query well.
The useful image: imagine a brilliant intern who has read everything, who writes quickly, who remembers nothing from one meeting to the next, and who never admits that he ignores an answer.
What it changes for you: it's excellent for reading, writing, classifying and retrieving. It's not a database, not a calculator, and not a source of truth about your business.
Model
A precise LLM, with its version and features. There are dozens of them, from very big and expensive to small and fast. The choice of model is an engineering decision, not a brand preference.
What it changes for you: a provider who uses only one model for everything, whatever the problem, works out of habit and not by analysis. A smaller and well framed model often beats a large model poorly integrated, for a fraction of the price.
Token
The text cutting unit. About three quarters of a word in French. Everything you send to the model and everything it responds to is counted in tokens, and that's exactly what is charged.
What it changes for you: this is the key to your usage costs. A system that sends 30 pages of context to each question costs a hundred times more than a system that sends the relevant paragraph. Two solutions that do the same thing may have completely different monthly bills. Ask the question.
Context window
The amount of text the model can have in front of them at one time. Beyond that, he no longer sees the beginning.
The image: a work desk. You can display a certain number of documents there. The rest stays in the binder.
What it changes for you: this limitation explains why we cannot simply “give all our documents to the model”. This is precisely the problem that the RAG solves.
Prompt, or prompt
Instructions given to the model. In a business system, the prompt is not a question typed by a user, it is a constructed, tested, versioned text, which contains the rules, the tone, the expected format and the prohibitions.
What it changes for you: it is a serious part of the implementation work. A vendor that treats the prompt as a detail will deliver an unstable system.
Hallucination
When the model produces false information, formulated with the same assurance as true information. An invented date, an article of law that does not exist, a plausible but false amount.
Important point, and often misunderstood: it is not a bug in the process of being corrected. It's the very way these systems work. The rate drops with the best models and good architecture. It doesn't fall to zero.
What it changes for you: We don't settle it by trust, we settle it by design. Human verification where error is expensive, citing sources, format constraints, and never an irreversible automatic decision.
Architectural terms
RAG, or retrieval-augmented generation
The method that allows the model to respond from your documents to you. The system first searches for relevant snippets in your documentation, then gives them to the model and asks it to respond with just that.
The image: instead of asking your intern what he remembers, you place the three correct files on his desk before asking the question.
What it changes for you: it's the technology behind internal enterprise search and the majority of useful assistants. This also reduces hallucinations the most, because the model can cite where its response comes from. For an SME, this is almost always the right approach, and almost always cheaper than the alternatives.
Vectorization and diving
The technique which allows you to find a document by the meaning rather than by the exact words. A search for “parental leave” brings up a document that talks about “absence at the birth of a child”.
What it changes for you: This is the internal mechanics of the RAG. You don't have to understand it in detail, but you should know that it explains why an AI-powered search finds things that your current search engine misses.
Agent
A LLM who has been given tools, as well as the right to sequence several steps without supervision between each. A classic assistant answers. An agent acts: he consults your system, calculates, writes to a database, sends an email.
What it changes for you: this is where the real potential lies, and also the real risk. An agent who makes a mistake doesn't just produce a bad sentence, he takes a bad action. Any conversation about agents must begin with a list of what they are allowed to do, and especially what they are not allowed to do alone.
In 2026, agents work well on delimited and well-defined tasks. Demonstrations of autonomous agents that manage complete processes are still, in the majority of cases, demonstrations.
Tool, or function call
What we plug into a model so that it can do something other than write: consult your inventory, create a customer file, calculate a tax.
What it changes for you: This is exactly where the integration work is located, and therefore a large part of the cost of a project. The pattern is the easy part. Connecting it correctly to your systems is the real job.
Orchestration
The coordination of several steps, models or tools in the same process. Extract, validate, classify, route, alert.
What it changes for you: A real process almost always requires orchestration. This is also what distinguishes a business project from a gadget connected to a single query.
Human in the loop
A mandatory stop point where a valid person before the system continues.
What it changes for you: it is your main risk control lever. My rule: human validation wherever an error affects a customer, commits the company, or modifies financial data. Full automation where the error is minor, detectable and correctable. This boundary is decided with you, explicitly, before conception.
Railings
The constraints that prevent the system from going out of its way: imposed formats, prohibited subjects, limits on amounts, automatic verification of responses.
What it changes for you: ask to see the guardrails provided. A provider who hasn't thought about it hasn't designed for production, they designed for demonstration.
Terms that are expensive
Fine tuning or fine tuning
Retrain a model on your own examples for a particular behaviour or style.
What it changes for you: it is often suggested, rarely necessary. Nine times out of ten, what an SME really wants is for the model to respond from its documents, and it's the RAG that does that, for much less money and without having to start all over again when your documents change. If a provider offers fine tuning from the outset, ask them why RAG would not be enough. His answer will tell you a lot.
Open model and closed model
An open template can be downloaded and hosted on your infrastructure. A closed model is used remotely at its provider.
What it changes for you: the open model gives maximum control over data, with real infrastructure and maintenance cost. The closed model is simpler and often more efficient, but your data passes elsewhere. The choice is made on your confidentiality constraints, not on an ideological preference. A comprehensive guide covers this decision.
Inference
Rotating the model to get a response. Every inference has a cost.
What it changes for you: This is your variable cost. It grows with your use. Require a twelve-month estimate at the actual expected volume.
Temperature
A setting that controls the degree of creativity of responses. Low for data extraction, higher for writing.
What it changes for you: technical detail, but a teller. If an invoice extraction system gives you different results for the same document, this setting is one of the wrong things.
The translation table in one line each
- A LLM predicts text.
- A token is your unit of invoice.
- The RAG gives him your documents.
- An agent gives him hands.
- The guardrails give it limits.
- The human in the loop gives him a boss.
- A hallucination is the permanent risk that justifies all the above.
| Term | Role | Business question |
|---|---|---|
| LLM | Text forecast | What task does it have to do? |
| RAG | Give your documents | Where does each answer come from? |
| Agent | Give hands | What actions can he take alone? |
| Human | Valid | Where does the company take over the decision? |
Build the chain behind an answer
Activate the components to see what they add to the system.
The model produces text and a person remains responsible for the result.
What I see on the ground
The word “agent” is becoming what the word “cloud” was in 2012. It gets stuck on just about anything.
My chat test, and it still works: ask the provider what their bot can do without human validation. If the answer is vague or enthusiastic rather than specific, it’s not a real agent. It’s just a form with fancy words.
Second test: Ask where the response comes from when the system responds. A good business system cites its sources. A system that responds without ever showing what it relies on requires a level of trust that nothing justifies.
When this vocabulary is not enough
This guide gives you what you need to judge a proposal. It doesn't make you a technical architect, and that's not the goal.
Three situations require a real technical look: when sensitive or regulated data are involved, when the system has to write in your production systems, and when a provider asks you for a multi-year commitment. In these cases, have someone look at the proposal. An hour of second thought costs less than a bad contract.
Frequently asked questions
01What is the difference between a chatbot and an AI agent?
A conversational robot responds with text. An agent uses tools and gestures in your systems. The difference is not cosmetic, it is of nature. It completely changes the level of risk and integration work.
02Is RAG better than fine-tuning?
For the vast majority of SMEs, yes. The RAG responds from your current documents, updates when your documents change, and costs a lot less. Fine tuning is mainly used to change behaviour or style, not to add knowledge.
03Does an AI model learn from our conversations?
Depends on the provider and the contract. On offers to companies, training on your data is usually excluded, but it is a clause to check, not to assume. The privacy guide covers the exact questions to ask.
04Why do AI answers change from one request to the next?
These systems are probabilistic in nature. For tasks that require consistency, such as data extraction, this variability is greatly reduced through adjustments and automatic checks.
05How much AI terminology do business leaders actually need to know?
Five are enough to be a good buyer: LLM, token, RAG, agent, human in the loop. The rest is comfort.
Background guide prepared by GXN (Digital Governance), a division of MD79. Revision planned according to the rhythm documented for this cluster.
References to verify for your situation
These references support the external rules and frameworks cited in this guide. They do not replace legal or professional advice tailored to your organization.