RAG, fine-tuning, agents, automation: a decision-maker’s AI glossary

You do not need to know how to program to decide on an AI project. You need to understand a dozen words, because those words determine choices that cost money or make it: hosting on your own premises or renting an API, indexing your documents or retraining a model, buying a chatbot or building an agent. When the vocabulary is fuzzy, it is rarely the buyer who benefits.

So this glossary is a decision tool: each term is explained in business language, with the moment it is the right choice and the moment it is not. Many projects that go off the rails start from a misunderstanding about vocabulary, as we show in our analysis of why AI projects fail in SMEs.

A language model is software trained on vast bodies of text to produce the most plausible continuation of a sentence. It is a probability engine, not a database: it writes, summarizes, rephrases and reasons, but it “knows” nothing about your business until you show it your documents.

An open-weight model can be downloaded and run on your own server, like the ones in our comparison of models to host yourself. The parameters, the billions of numerical values that make up the model, are its “brain size”: the more there are, the more nuanced the model, and the more memory it demands. Quantization compresses those parameters so a model fits on reasonable hardware, with a marginal loss of quality in most business uses.

RAG (retrieval-augmented generation) has the model consult your documents at the moment it answers: it finds the relevant passages, then writes from them. Your files stay files, the model does not change. It is the default approach for querying a company’s knowledge, because it updates continuously and it can cite its sources.

Fine-tuning changes the model itself from chosen examples. It excels at internalizing a style, a format or a trade terminology; it is poor at learning facts, which go stale and are hard to correct. Full training means building a model from scratch: that is the business of large laboratories, not of an SME, and it is almost never the right answer to your need.

A chatbot converses: it answers questions in a chat window, usually within a narrow scope. An assistant helps a person with their task: it writes, summarizes, prepares, but does not act on its own. An agent chains actions together: it queries your systems, cross-references data, prepares a complete deliverable, then submits it for approval. That is the sense we give the word at Cogio: AI prepares, people decide.

Classic automation (RPA, robotic process automation) is nothing like intelligent, and that is its strength: it replays identical steps, fast and without tiring. When a process is perfectly repetitive, RPA is more reliable and cheaper than an agent. The agent becomes relevant when you have to understand a document, judge a particular case or combine several sources; its cost depends above all on how many systems have to be connected.

Models do not read words, they read tokens: fragments of text, often shorter than a word. API billing is counted per token, and the context window, meaning how much text the model can consider at once, is measured in tokens too. A large window lets it swallow a full tender with its appendices; a small one forces you to slice it up, at the risk of losing the thread.

Hallucination is the most discussed flaw: the model asserts something false with complete confidence. It happens mostly when you push it to answer beyond what you have shown it. You do not eliminate it, you contain it: RAG that cites its sources, well-defined scopes, and human sign-off on anything that commits the company.

Inference is the moment the model does the work: it reads your request and produces its answer. It can happen on your own server (local inference) or at a provider’s, through a cloud API. The fundamental difference is not raw quality: it is the journey your data takes.

Locally, your documents never leave the company: the Law 25 compliance analysis gets simpler and the US CLOUD Act drops out of the equation. Through an API, you reach the most powerful models without buying hardware, but you have to examine the contract, the data residency and what is done with it. Our comparison of local AI and ChatGPT for business helps you decide according to how sensitive your documents are.

Embeddings turn each passage of text into a sequence of numbers that captures its meaning. Two passages about the same thing get neighbouring numbers, even with no words in common. That is the engine behind semantic search: type “conveyor failure” and find the record titled “line 3 shutdown.”

It is also the hidden half of a RAG project: the quality of the answers depends as much on the model that searches as on the model that writes. The right choice when your employees lose time digging through the shared server; of limited interest if your entire document set fits in one filing cabinet.

Vocabulary only matters if it leads to a decision. Here is the grid we use at scoping: start from what you want to get, and the rest follows.

Pick a concrete irritant (quotes that take too long, knowledge sleeping in inboxes, repetitive questions) and restate it in the words of this glossary: do you want to search, write, converse or act? That one-line sentence beats a long requirements document, and it protects you from solutions sold before the problem.

At Cogio, that is how every scoping starts: the business need first, the vocabulary second, the technology last. We design custom agents hosted at the client’s site, a form of enterprise brain that prepares the work while your people decide. If you want to sanity-check how you have framed a need, write to us: a conversation is often enough to restate a problem in AI terms, and it commits you to nothing.