A short glossary, tools, and references

Fundamentals, Key Applications and Types of AI

Term	Definition & Practical Context
AI (Artificial Intelligence)	The ability of a machine to imitate human intelligence, including reasoning, learning, and problem-solving. Context: Your company could use AI to automate inbound email classification, allowing your teams to focus on complex cases.
Machine Learning (ML)	A branch of AI where systems learn from data to make predictions, without explicit programming for each case. Example: ML allows predicting customers at risk of cancelling their subscription (churn) by analyzing their usage history.
Deep Learning	A subfield of ML that uses very complex neural networks. It is the basis for the current performance of the most advanced models. Example: It is Deep Learning that enables object recognition on a production line for automated quality control.
Generative AI	A type of AI that creates new content (text, images, code, video) instead of simply analyzing existing data. Application: You can quickly generate initial versions of marketing materials or blog post drafts.
LLM (Large Language Model)	Large Language Model, trained on massive amounts of text to understand and generate human language. Application: Used to power a conversational agent capable of answering your customers’ common questions 24/7.
Multimodal AI	An AI model capable of processing multiple types of data (text, image, sound) at the same time. Practical Case: A multimodal tool could analyze a photo sent by a customer (image of a broken product) and respond with a pre-filled claim text.
Computer Vision	The field of AI that enables computers to “see”, analyze, and interpret the content of images and videos. Practical Case: Used in automatic construction site analysis by drone to track work progress and detect anomalies.
NLP (Natural Language Processing)	Natural Language Processing, allowing machines to read, understand, and interpret human language. Application: Essential for automatically classifying customer support ticket content or analyzing sentiment in online reviews.
Foundation Model	A very large AI model, trained on general data to serve as a solid base and then be adapted to specific business tasks. Context: Instead of training an AI from scratch, you start with an already powerful Foundation Model to specialize it.

Interaction, Personalization, and Action of AI

Term	Definition & Practical Context
Prompt	The instruction, question, or initial text you provide to the AI to guide its response. Example: Asking the AI: “Write a value proposition for our new service, targeting SMEs and using a dynamic tone.”
Context	The prior information (chat history, documents, data) that the AI uses to generate a relevant and consistent response. Practical Case: The agent knows you are talking about your shower renovation project (even if you don’t specify it every time) because it keeps the context of the conversation.
RAG (Retrieval-Augmented Generation)	Retrieval-Augmented Generation. The model first searches for information in your private database before formulating a response. Customer Benefit: Allows an internal chatbot to answer employee questions based on the company’s HR or technical documents.
MCP (Model Context Protocol)	A standardized protocol allowing AI models to access tools and external data sources in real-time, a bit like a “USB port” for AI. Benefit: Allows your AI agent to directly query your up-to-date customer database to give you a precise sales figure, without an intermediate process.
Prompt Engineering	The art of designing precise and optimal prompts to improve the quality and reliability of AI results. Utility: Good prompt engineering is what takes AI from a simple brainstorming tool to a high-performing and consistent content writer.
Context Window	The maximum amount of text (measured in tokens) that the model can “see” at one time to generate its response. Importance: A large window is crucial if you need to summarize a 200-page financial report or analyze a long legal contract.
Fine-Tuning	The process of training a pre-existing LLM on a small dataset specific to your company. Result: The model learns to use your brand’s exact vocabulary, technical jargon, and tone.
Embedding	A numerical representation of a word or phrase that captures its semantic meaning. Utility: Allows internal search engines to find relevant documents even if you use synonyms in your query.
Zero-Shot Learning	The ability of a model to correctly answer a new task or question without having received a prior example for this specific task. Example: The model can translate a text into a language it has never seen during its training because it has generalized translation rules.
Hallucination	The tendency of AI models to generate information that seems factual and convincing, but is totally false or invented. Mitigation: Using RAG is the best defense against hallucinations, as the AI must reference verifiable sources.

Risks, Ethics, and Operations (MLOps)

Term	Definition & Practical Context
Algorithmic Bias	Systematic and unwanted error in the results of an AI system, often caused by social prejudices reflected in training data. Client Issue: If your recruitment model is biased against certain profiles because of your historical data, it will perpetuate discrimination.
Explainability (XAI – Explainable AI)	Field of AI aiming to develop systems whose decisions can be understood by humans. Importance: Crucial for credit or insurance decisions, where the law requires being able to explain why a customer was refused.
MLOps (Machine Learning Operations)	A set of practices aimed at deploying and maintaining AI models in production reliably and efficiently. Function: Ensures that the model making predictions is always up to date, high-performing, and does not introduce bugs into your system.
Turing Test	A test designed to evaluate if a machine can exhibit intelligent behavior indistinguishable from that of a human during a conversation. Relevance: It is the historical reference for judging how realistic a conversational agent is.
Generalization	The ability of an AI model to successfully apply learned knowledge to new and unseen data. Importance: A model that generalizes well is reliable and useful in the real world, unlike a model that does Overfitting.
Overfitting	Situation where the model has learned the training data too exactly, including specific errors. Consequence: If the model over-learns on your Christmas sales data, it will be unable to correctly predict Easter sales.
Data Pipeline	The sequence of automatic steps through which raw data is collected, cleaned, transformed, and routed to the AI model. Utility: Ensures that the AI always works with clean and up-to-date data, guaranteeing the quality of its predictions.
Transfer Learning	A technique where a model already trained on a task (like a generalist LLM) is reused and adapted to solve a different task. Benefit: Significantly reduces the time and resources needed to create a specialized AI for your business sector.

The AI Ecosystem (Key Players & Tools)

The "Brains" (LLM Models)

Name (Player/Tool)	Description & Strengths
GPT-4o (OpenAI)	The current leader, versatile and high-performing. Market benchmark.
Gemini (Google)	The direct competitor, very strong in multimodal (video/image) and integrated into Google Workspace.
Claude 3 (Anthropic)	Renowned for its large context window (analysis of large books) and its ethics.
Llama 3 (Meta)	The powerful Open Source model (free and modifiable).
Mistral (Mistral AI)	The French gem 🇫🇷. Excellent performance/price ratio, often used in enterprise.

The Creatives (Images & Video)

Name (Player/Tool)	Description & Strengths
Midjourney	The artistic reference for “wow” images.
DALL-E 3	The simple image generator integrated into ChatGPT.
Stable Diffusion	The infinitely customizable image generator (Open Source).
Sora (OpenAI)	Ultra-realistic video generation (currently being deployed).
Runway	Professional tool for AI video editing and generation.

The Voice & Sound

Name (Player/Tool)	Description & Strengths
ElevenLabs	The leader in voice cloning and ultra-realistic text-to-speech.
Whisper (OpenAI)	The best tool for transcribing audio to text (for meetings).
Synthesia	Creation of video avatars that speak your text (for e-learning).

The Infrastructure (The "Pipes" and the "Hardware")

Name (Player/Tool)	Description & Strengths
NVIDIA	The chip (GPU) seller without whom no AI works.
Microsoft Azure	The cloud hosting OpenAI (secure for pros).
AWS (Amazon)	The cloud giant with its “Bedrock” platform to access all models.
Google Cloud (Vertex AI)	Google’s pro platform for building AIs.
Hugging Face	The global “library” where researchers share Open Source models.

Productivity Tools (Applications)

Name (Player/Tool)	Description & Strengths
Microsoft Copilot	AI in Word, Excel, PowerPoint.
GitHub Copilot	The AI that writes code for developers.
Perplexity	The next-generation search engine that cites its sources.
Zapier	Automation tool that connects AI to 5000 other apps.
Notion AI	AI integrated into note-taking and project management.

Key AI Metrics: Measuring Proof of Value

Business Impact Metrics (The ROI of AI)

Business Metric	Definition & Practical Case	Client Argument
Average Handling Time (AHT)	The average time needed to complete a task from start to finish (e.g., answering a customer ticket, processing an invoice).	Proof of Efficiency: “Thanks to our support LLM agent, we reduced AHT by 40%, going from 3 hours to 1h48, freeing your teams for critical cases.”
Cost Per Interaction/Task	The total cost (salary, infrastructure, time) to process a unit (customer, transaction, document).	Proof of Profitability: “AI automation decreased the cost of verifying an invoice by €3, resulting in a saving of €60,000 per year for 20,000 invoices.”
AI Resolution Rate	The proportion of customer requests or internal tasks that the AI resolves alone, without human intervention.	Proof of Autonomy: “Our RAG-powered chatbot achieves an 85% Resolution Rate on Level 1 questions, allowing your experts to focus on Level 2.”
Human Error Rate (HER)	The percentage of critical tasks performed by AI, which were previously prone to human error.	Proof of Reliability: “Quality control by computer vision brought the HER on the assembly line from 3% to 0.2%, drastically improving the quality perceived by the final customer.”
False Alarm / Detection Rate	Measures the efficiency of AI in correctly identifying critical events (fraud, failure, churn).	Proof of Relevance: “Our predictive maintenance model increased the critical failure detection rate from 65% to 95%, eliminating unplanned production stops.”

Model Performance Metrics (Technical Reliability)

Technical Metric	Definition & Context	Why it matters for the client
Precision	Measures the accuracy of positive predictions: among everything the model said was positive (e.g., a fraud), what proportion actually was?	Argument: High Precision means the AI doesn’t generate “false positives”. Your sales team won’t waste time calling prospects who weren’t qualified by the AI.
Recall	Measures the model’s ability to find all relevant occurrences: among all actual positive cases (e.g., all real frauds), how many did the model identify?	Argument: High Recall is vital in fraud detection or medical fields. It guarantees the AI doesn’t let any threat or critical anomaly pass.
F1 Score	A weighted average of Precision and Recall. It is a more balanced indicator when both metrics are important.	Argument: It is the single reference figure to evaluate the balance between not missing critical cases (Recall) and not generating false alerts (Precision).
False Positive Rate (FPR)	The percentage of cases where the model sounded the alarm incorrectly (e.g., labeling a legitimate email as spam).	Consequence: If this rate is too high, your employees will lose confidence in the AI tool and stop using it.
False Negative Rate (FNR)	The percentage of cases where the model missed the critical event (e.g., letting a real fraud pass without detection).	Consequence: A high FNR is the most costly. It results in direct financial losses or security risks.
Latency	The time the AI model takes to generate a response after receiving a request.	Customer Experience: Low latency is crucial for a good user experience. Nobody waits 10 seconds for a chatbot response.