AI Lexicon for executives: twenty words explained simply

This glossary does not seek to make you an engineer. It gives you the correct meaning of twenty words, organized by theme, so that you can understand an offer and ask the right question at the right time. Each definition fits in one or two sentences and is accompanied by a concrete scenario. If you are looking for the difference between an agent, a chatbot, and an automation, the dedicated article covers it in detail.

IThe basics: model, LLM, token, prompt

The model is the computing brain; the product you use dresses it in an interface, rules and tools. Two products can use the same model and deliver very different results, depending on the instructions and documents they provide to it.

A token is the unit of text that the model reads and writes. It is neither a letter nor quite a word: a short word often fits in one token, a long word in several. A plumber's reminder email, with the invoice attached, represents a few thousand tokens. Providers bill per token, which is why we talk about it so much.

The prompt is the text you send to the model to ask it something. "Write a polite email to request invoice 2024-118, which is three weeks overdue" is a prompt. The more it specifies the purpose, tone and useful information, the better the response.

IIAn agent's work: mission, automation, dry run, validation

An agent is a program that uses a model to complete a task from start to finish: it reads, decides on the next step, uses tools and delivers a result. Where a chatbot answers a question, the agent processes a file, for example all overdue invoices from Monday.

The mission is the lasting instruction of an agent: what it does, for whom, at what pace and with what limits. It is a prompt that does not change from one time to the next. For a bakery: "Each Friday, prepare the flour order from the sales table, never exceeding the indicated budget."

Automation refers to a sequence of fixed actions triggered without human intervention: when an invoice is paid, archive it in such a folder. It does not think. An agent, on the other hand, can handle an unexpected case, which requires more oversight.

The dry run is a complete execution where nothing is sent: the agent prepares its emails, quotes or lists, and you read them before any sending. Validation is the moment when you approve a result before it is sent. At ABYS, you validate the first sendings by hand; after three validations, you can switch to automatic validation, and the risks of this choice are written in black and white.

IIIWhat the agent reads: context, RAG, fine-tuning, data, multimodal

Context is all the text the model has before its eyes at the moment of responding: the mission, the conversation, attached documents. Its size is limited and is counted in tokens. What is not in the context does not exist for it: a model does not remember your last exchange, unless the software gives it back to it.

RAG (retrieval-augmented generation) consists of searching, before each response, for useful passages in your documents, then placing them in the context. A real estate agency gains an agent who answers about its own listings rather than generalities. Fine-tuning is something else: you continue training the model on your examples to durably modify its style or format. It is costly, and poorly suited to information that changes often, such as prices.

Structured data is arranged in fixed boxes: a spreadsheet with a client column, an amount column, a due date column. An email or scanned PDF is unstructured data. An agent reads both, but makes far fewer mistakes with the former. A multimodal model, finally, processes several kinds of content: text, images, sometimes sound. It can read the photo of a crumpled delivery slip.

IVQuality and risks: hallucination, safeguard, log

A hallucination is a false answer stated with confidence: a law article that does not exist, an invented date, a rounded price. It comes from the very functioning of the model, which composes a plausible continuation. You limit it by providing sources, by asking the agent to say when it does not know, and by reviewing.

A safeguard is a rule that prevents the agent from crossing a limit: never grant a discount, never write to a client in dispute, stop and ask as soon as an amount exceeds a threshold. The log is the dated trace of what the agent did: when it passed, what it read, what it prepared or sent. It is what you open when a client says they received nothing.

VCosts and technique: API, latency, cost per call

An API (application programming interface) is a door through which one software uses another. When an agent sends text to a model, or creates a contact in billing software, it goes through an API. You do not have to manipulate it, but its presence tells whether two tools can talk to each other.

Latency is the delay between the request and the response, from less than a second to several minutes for a long file. For an agent that works each night, it matters little; for a phone response, it matters a lot. The cost per call is the price of an exchange with the model. It depends on the tokens sent and the tokens produced, the latter often being billed more expensively. Reviewing twenty pages therefore costs more than sorting ten lines.

VIThe path of a request, with all these words

These twenty words actually describe a single journey. Here is what happens when the agent of an accounting firm prepares, on Tuesday morning, reminders for missing documents.