Artificial intelligence glossary

AI Act

EU Artificial Intelligence Act

Regulation (EU) 2024/1689 governs AI systems according to their level of risk. It bans certain uses, sets strict obligations for high-risk systems and transparency rules for chatbots and generated content. Its obligations apply in stages between 2025 and 2027.

AI agent

An AI agent is a program that is given a goal, decides on the steps itself and uses tools (software, APIs, documents) to reach it. Unlike a basic chatbot, it takes action: it can create a customer record, send an email or update a spreadsheet.

AI answer engine

An AI answer engine replies to a question directly with written text and a few cited sources, instead of a list of links. ChatGPT with search, Perplexity, Gemini and AI Overviews are examples.

AI assistant

An AI assistant is a general-purpose conversational app, such as ChatGPT, Claude, Gemini or Copilot, that helps people write, summarize, analyze or search. In a company, it comes with usage rules and, where useful, a connection to internal documents.

AI chatbot

An AI chatbot is a conversational assistant on a website, WhatsApp or Messenger that answers customer questions from the company's own content, around the clock. It can qualify a request, offer an appointment and hand over to a person.

AI literacy

AI literacy is the obligation, under Article 4 of the AI Act and applicable since February 2025, for organizations that use AI to ensure a sufficient level of AI skills among their staff. Training and clear usage rules are how it is met.

AI Overview

An AI Overview is the AI-generated summary that Google shows above the regular results for some searches. It cites a handful of source pages, so being one of them is now a visibility goal in its own right, like ranking first.

AI voice agent

callbot, AI receptionist

An AI voice agent answers the phone in natural language, understands what the caller wants and acts on it: booking appointments, answering common questions, routing the call to the right person. It combines speech recognition, a language model and speech synthesis.

API

application programming interface

An API is a standard doorway that lets two pieces of software exchange data. APIs are what allow an AI agent or chatbot to read a calendar, create a quote in a CRM or check stock levels.

Automation (workflow)

An automated workflow chains actions across several tools with no human input: a form submission creates a contact, sends an email and alerts the team. With AI, the workflow can also read, sort or write.

Callbot

Callbot is another name for an AI voice agent: a conversational bot that answers phone calls. Modern callbots understand free-form sentences, unlike older press-a-key phone menus.

Context window

The context window is the amount of text, measured in tokens, that a language model can take into account at once: the question, the conversation history and any documents provided. Beyond that, information has to be summarized or selected.

Document extraction

OCR, IDP

Document extraction automatically reads invoices, purchase orders or forms, including scanned ones, and turns them into structured data: amounts, dates, references. It removes most manual data entry.

E-E-A-T

experience, expertise, authoritativeness, trust

E-E-A-T refers to the quality criteria Google uses to assess content: the author's first-hand experience, their expertise, the site's authority and the reliability of the information. AI answer engines also favor identifiable, credible sources.

Embeddings

An embedding is a list of numbers that represents the meaning of a piece of text. Two sentences that mean the same thing have similar embeddings, even with different words, which is what makes search by meaning possible.

Evaluation (evals)

Evaluation means testing an AI system against a set of real questions and cases to measure answer accuracy, instruction-following and errors. It is repeated whenever the model, prompt or content changes.

Fine-tuning

Fine-tuning retrains an existing model on a company's own examples to adjust its style or a very specific task. To answer from documents that change often, RAG is usually simpler and cheaper.

GDPR

General Data Protection Regulation

The GDPR governs the use of personal data in the European Union. For an AI project, it requires among other things a legal basis, informing the people concerned, data minimization and safeguards on processors and transfers outside the EU.

Generative AI

Generative AI refers to models that produce new content, whether text, images, audio, video or code, from a simple instruction. ChatGPT, Claude and Gemini are the best-known examples.

GEO

Generative Engine Optimization

GEO is the set of practices that help a business get cited and recommended by AI answer engines such as ChatGPT, Perplexity, Gemini or Google's AI Overviews. It builds on SEO: clear content, verifiable facts, structured data and a presence on trusted sources.

Guardrails

Guardrails are the rules and checks that keep an AI in bounds: allowed topics, refusal of out-of-scope requests, answer checks, filtering of personal data. They make a chatbot or agent predictable and brand-safe.

Hallucination

A hallucination is a wrong answer stated confidently by a language model, such as a price, reference or law that does not exist. It is reduced by grounding answers in sources (RAG), setting guardrails and reviewing sensitive content.

Human handover

Handover is the moment a chatbot or voice agent passes the conversation to a person, along with a summary of the request. It happens when a question falls outside the agreed scope, when the customer asks for it or when the situation is sensitive.

Human in the loop

Human in the loop is a design principle: a person approves the AI's important actions before they go out, such as a quote, a customer email or a payment. It combines the time saved by automation with human control.

Inference

Inference is using an already trained model to produce an answer. Providers bill it per use, usually by the number of tokens processed, and it determines response time.

Intent

Intent is what a person actually wants behind their words: booking an appointment, getting a price, tracking an order. Recognizing intent lets a chatbot or voice agent choose the right action.

IVR

interactive voice response

IVR is the classic "press 1, press 2" phone menu. It routes calls but cannot understand free-form requests; an AI voice agent replaces it with a real conversation.

Knowledge base

The knowledge base gathers the content a chatbot or AI agent relies on to answer: website pages, FAQs, product sheets, procedures, terms and conditions. Its quality directly determines the quality of the answers.

Latency

Latency is the delay between a question and the start of the answer. It matters little for a report but a lot on the phone: a voice agent needs to respond within roughly a second for the conversation to feel natural.

LLM

large language model

An LLM is an AI model trained on vast amounts of text to understand and generate language. It predicts the most likely continuation of a text, which lets it answer, summarize, translate or write. GPT, Claude and Gemini are LLMs.

llms.txt

llms.txt is a text file at the root of a website that introduces the business and its main pages to language models, in Markdown. It is a proposed standard: it makes a site easier for AI assistants to read, with no guarantee of being cited.

Machine learning

Machine learning covers the methods that let a program learn from examples rather than hand-written rules. Language models are one application; sales forecasting and fraud detection are others.

MCP

Model Context Protocol

MCP is an open protocol, launched by Anthropic in late 2024, that standardizes how AI assistants connect to tools and data (CRM, calendar, files, databases). The same MCP connector can serve several compatible assistants.

Multi-agent system

A multi-agent system splits work between several specialized AI agents, coordinated by a lead agent: one reads emails, another drafts the quote, a third checks it. This setup makes long processes more reliable.

Multimodal model

A multimodal model understands and produces several types of content: text, images, audio and sometimes video. It can, for example, read a photo of an invoice, describe a diagram or hold a spoken conversation.

NLP

natural language processing

NLP is the field of AI that lets machines understand and produce human language: classifying text, extracting information, translating, analyzing sentiment. LLMs are its most recent step.

No-code and low-code

No-code and low-code tools such as n8n, Make or Zapier let you build automations by assembling visual blocks rather than writing all the code. They speed projects up; custom code takes over when the logic gets complex.

Open-weight model

open-source model

An open-weight model is a downloadable AI model that can run on your own servers, such as Llama or Mistral. It offers more control over data, in exchange for heavier setup and maintenance.

Prompt

A prompt is the written instruction given to a generative AI. The more precise it is about context, goal, tone and expected format, the better the answer. Teams save time by sharing a library of proven prompts.

Prompt engineering

Prompt engineering is the practice of writing and testing instructions for an AI so that results are reliable and consistent: examples, steps, output format, rules to follow.

Prompt injection

Prompt injection is an attack that hides instructions in a message, document or web page to make an AI ignore its own rules. Agents that read external content must be designed to disregard them.

Proof of concept

PoC

A proof of concept is a first, limited prototype that checks on real data whether an AI project works and delivers a measurable gain, before investing in a full solution.

Pseudonymization

Pseudonymization replaces data that identifies a person (name, email, number) with codes before processing it, for example before sending it to an AI model. It lowers the risk, but the data remains personal data under the GDPR.

RAG

retrieval-augmented generation

RAG first retrieves the relevant passages from a company's documents, then gives them to the language model so it answers from those sources. The result: up-to-date, verifiable answers with fewer hallucinations.

Semantic search

Semantic search finds content that matches the meaning of a question, not just its exact words. "My parcel hasn't arrived" will find the "Track your delivery" page.

SEO

search engine optimization

SEO covers the techniques that improve a website's position in Google's and Bing's unpaid results: technical quality, useful content, inbound links. It remains the foundation that GEO builds on.

Shadow AI

Shadow AI is staff using AI tools without the company's framework or approval. It exposes confidential data to leaks; the answer is a usage policy, approved tools and training.

Speech recognition

STT, speech-to-text

Speech recognition turns spoken words into text in real time. In a voice agent it is the first step, and it has to cope with accents, background noise and proper names.

Speech synthesis

TTS, text-to-speech

Speech synthesis turns text into a voice. Today's neural voices sound natural and expressive and are available in English, French, Dutch and many other languages.

Structured data

schema.org, JSON-LD

Structured data is information added to a page's code, using the schema.org vocabulary, to describe a business, article, FAQ or product unambiguously. It helps Google and AI answer engines understand and cite the page.

System prompt

The system prompt is the permanent instruction, hidden from the user, that sets a chatbot's or agent's role, tone, limits and rules. It applies to every conversation.

Temperature

Temperature is a setting that controls how much randomness goes into a model's answers. Low values give consistent, predictable answers, ideal for customer service; higher values give more variety, useful for brainstorming.

Token

A token is the unit of text a language model processes, often part of a word: an English word averages a little over one token. Prompt and answer length, and therefore usage cost, are measured in tokens.

Tool calling

function calling

Tool calling is a language model's ability to request a specific function (look up availability, calculate a price, read a page) and then use the result in its answer. It is the core mechanism behind AI agents.

Vector database

A vector database stores text as embeddings and retrieves, in milliseconds, the passages closest in meaning to a question. It is the internal search engine of a RAG system.