AI glossary for executive leaders
55 terms to discuss AI in the executive committee, without jargon or complacency. Each entry gives the short definition, the concrete example and the angle that matters to a leader.
Fundamentals(10)
- EmbeddingAn embedding is the numerical representation of a word, a sentence, or a document, in the form of a vector in a space of several hundred or thousand dimensions. Two texts semantically close have geometrically close embeddings, which enables search by meaning rather than by keywords.
- Foundation modelA foundation model is a generalist AI model trained on massive, diversified data, which serves as a reusable base for dozens or even hundreds of different business applications, by adaptation (fine-tuning, RAG, prompt) rather than by specific training.
- Generative AIGenerative AI is a family of artificial intelligence models capable of producing new content (text, image, sound, video, code) from a natural-language instruction, as opposed to traditional AI which was limited to classifying, predicting, or detecting.
- InferenceInference is the usage phase of an AI model, during which the model computes a response from a given prompt. It is the operation billed by API providers, distinct from training which is an initial fixed cost.
- LLM (Large Language Model)An LLM (Large Language Model) is a type of artificial intelligence trained on text corpora of several hundred billion words, which produces natural language by predicting, word by word, the most probable continuation of a given text.
- MultimodalA multimodal AI model is a model capable of processing and producing several types of content simultaneously: text, image, audio, video, code. The same model can analyse a photo, understand a voice query, read a document, and reply in writing, without any intermediate pipeline.
- ParameterA parameter of an AI model is a numerical value learned during training that determines the strength of connections between the artificial neurons of the network. The parameter count measures the model's capacity to store associations between the elements of its training corpus.
- TokenA token is the elementary unit of text that an AI model manipulates, generally a word fragment equivalent to 3 or 4 characters in English. It is both the model's unit of computation and the AI providers' unit of billing, counted separately on input and output.
- TrainingTraining is the construction phase of an AI model, during which the model ingests a massive corpus and adjusts its billions of internal parameters to learn the statistical regularities of that corpus. It is the most expensive operation in a model's life cycle.
- TransformerThe transformer is the neural-network architecture, introduced by Google in 2017, that underpins nearly all current generative AI models. Its central innovation is the attention mechanism, which lets the model dynamically weigh the relative importance of words in a sequence.
Interaction and usage(5)
- Context windowThe context window is the maximum amount of text an AI model can process in a single query, measured in tokens. It includes the prompt sent and the response to produce. Beyond this limit, the model loses access to the oldest information of the exchange.
- Conversation rolesConversation roles (system, user, assistant) are the labels assigned to each message in an exchange with an LLM. The system role defines the model's general behaviour, the user role carries the requests, the assistant role the model's past responses in the conversational history.
- HallucinationA hallucination is a factually false statement but formulated with confidence by a generative AI model. It stems from the statistical mechanism of generation: the model predicts the most likely next word, with no native capacity to verify. It is the main operational and legal risk of an enterprise AI deployment.
- PromptA prompt is the natural-language instruction sent to an AI model to obtain a response. Its quality directly conditions the quality, relevance, and cost of the generated output. A good prompt is not long; it is precise, structured, and contextualised.
- Prompt engineeringPrompt engineering is the discipline of writing prompts methodically. It is not a rare profession but a transverse competence, which consists of formalising, testing, and standardising the instructions sent to an LLM to obtain reliable and reproducible outputs at the scale of an organisation.
Advanced architectures(9)
- AI APIAn AI API is a technical interface that lets a software application send requests to an AI model hosted by a provider, and retrieve its responses. It is the standard access mode to AI in enterprise, as opposed to local hosting of the model.
- DistillationDistillation is a technique that transfers the knowledge of a large AI model (teacher model) to a smaller model (student model), while preserving most of the performance. It enables the deployment of lightweight models with reduced inference cost, viable on more modest infrastructures.
- Fine-tuningFine-tuning is an adaptation technique for an already-trained AI model, which consists of continuing its training on a dataset specific to your use case. It modifies the model's internal parameters, in contrast to RAG, which simply injects context at query time.
- Function callingFunction calling is the ability of an AI model to invoke predefined functions or tools to execute actions in an external system. The model returns a structured object (JSON) rather than text, allowing the application to call the function and reinject the result into the conversation.
- MCP (Model Context Protocol)MCP (Model Context Protocol) is an open standard, introduced by Anthropic in November 2024, that lets an AI model connect to data sources and external tools in a uniform way. It avoids writing specific connectors for every model-application combination.
- MoE (Mixture of Experts)Mixture of Experts (MoE) is an AI model architecture that splits the network into specialised sub-models, called experts. For each token processed, a router dynamically selects a few experts, leaving the others inactive. The model has the capacity of a large model but the compute cost of a smaller one.
- Open-source modelAn open-source AI model is a foundation model whose weights and architecture are freely downloadable and usable under a permissive licence (Apache 2.0, MIT). It contrasts with the proprietary model (Claude, GPT, Gemini) accessible only via API. The choice engages sovereignty, cost, and long-term flexibility.
- RAG (Retrieval-Augmented Generation)RAG (Retrieval-Augmented Generation) is an AI architecture that pairs a search engine across your documents with a generative model. The model answers by relying on citable business data rather than on its training knowledge alone.
- Vector databaseA vector database is a database specialised in the storage and retrieval of vectors (embeddings). It allows, for a given query, finding the most semantically close content in a corpus, without exact lexical match. It is the typical search engine of a RAG system.
Agents and automation(7)
- Agent orchestrationAI agent orchestration is the coordination of several specialised agents, each responsible for part of a complex task, so they collaborate towards a common goal. A central orchestrator (or a supervisor agent) distributes the sub-tasks, supervises their execution, and aggregates the results.
- Agentic loopThe agentic loop is the iterative cycle of an AI agent that acts, observes the result, and decides on the next action, until it reaches its goal or a stopping criterion. Without limits or observability, it can diverge into an infinite loop and consume in a few minutes the equivalent of several days of normal use.
- AI agentAn AI agent is an autonomous system that combines a language model with external tools to execute multi-step tasks: it analyses a request, plans an action sequence, executes, observes, and adapts its strategy. It differs from an assistant that merely responds.
- AI agent platformAn AI agent platform is an environment that allows enterprises to create, deploy, and operate autonomous AI agents. Three families exist in 2026: developer SDKs (Claude Agent SDK, LangGraph, AutoGen), no-code platforms (n8n, Make, Microsoft Copilot Studio), and self-hosted open-source integrated products (OpenClaw, OpenHands).
- AI workflowAn AI workflow is the structured sequence of several AI inference steps to accomplish a complex task, usually with external tools interleaved. It can be pre-defined (linear chain) or dynamic (orchestrated by an agent), and is the dominant operational mode of mature AI deployments in enterprise.
- Augmented automationAugmented automation is the use of AI to assist a human worker in performing a task, rather than to replace them. The human remains decision-maker and responsible, the AI accelerates, prepares, or validates. It is the dominant mode in 2026 for tasks with legal, creative, or relational stakes.
- CopilotAn AI copilot is an assistant that suggests, proposes, and completes actions to the user, who keeps control of the final decision and execution. It is distinct from the AI agent, which executes autonomously. The copilot is a prudent mode of use, the agent a delegated mode of use.
Risks and governance(9)
- AI ActThe AI Act is European Regulation 2024/1689, the world's first legal framework to classify AI systems by risk level and to impose differentiated obligations: prohibition for unacceptable uses, strict requirements for high-risk systems, transparency for the rest.
- AI carbon footprintThe AI carbon footprint is the set of greenhouse gas emissions related to its life cycle: model training, component manufacturing, data centre operation, and inference at usage. It is becoming a significant consumption item on a global scale.
- Algorithmic biasAn algorithmic bias is a systematic deviation in an AI model, inherited from training data, design choices, or deployment context, that produces unfair or erroneous decisions to the detriment of a population group or a type of case.
- AlignmentAlignment is the set of techniques that aim to steer the behaviour of an AI model towards the goals and human values of its user or publisher. It turns a raw model, capable of producing anything, into a useful, honest assistant that refuses requests contrary to the rules set.
- Digital sovereigntyDigital sovereignty is the capacity of a company or a state to control its data, its infrastructure, and its AI tools, independently of foreign jurisdictions. It conditions confidentiality, regulatory compliance, and resilience to access disruptions dictated by external political decisions.
- GDPR and AIGDPR governs all personal-data processing by an AI system operated in Europe or concerning Europeans. Four obligations apply as priorities: explicit purpose, data minimisation, access and opposition rights, and impact assessment for high-risk processing.
- GuardrailsGuardrails are technical control layers, added at the input or output of an AI model, that detect and block undesirable behaviours: malicious prompts, data leaks, forbidden content. They are distinct from alignment, which acts on the model's default behaviour itself.
- Red teamingRed teaming, borrowed from cybersecurity, is the practice of testing an AI system by simulating adverse usage attempts: rule bypassing, sensitive-data extraction, forbidden-content generation. It aims to identify vulnerabilities before a malicious actor exploits them in production.
- Training dataTraining data is the set of texts, images, code, and other content used to train an AI model. Its composition determines what the model knows, what it ignores, its biases, and its legal risks. A major part of current AI litigation concerns their provenance and lawfulness.
Business and adoption(8)
- AI POCAn AI POC (proof of concept) is a short pilot project aimed at testing the technical feasibility and business value of an AI application before any industrial investment. According to RAND 2025, 80.3% of enterprise AI POCs fail to deliver their expected business value.
- AI ROIAI ROI, modelled on classic financial ROI, is a deceptively appealing metric for evaluating an AI deployment. Three structural confusions invalidate it: underestimation of hidden costs, lack of robust pre-AI baseline, and concentration on operational efficiency to the detriment of the other two dimensions of value.
- AI TCOAI TCO (total cost of ownership) is the real total cost of an AI deployment in enterprise, aggregating five items: inference model, infrastructure, application integration, operational supervision, and team cost. According to RAND 2025, production overcosts average 380% of pilot estimates.
- AI use caseAn AI use case is a concrete application of artificial intelligence to an identified business process, with a measurable objective and a defined perimeter. The three most deployed use cases in enterprise in 2026 are software development assistance, customer service, and data analysis.
- Augmented productivityAI-augmented productivity refers to measurable production gains (volume processed, speed, quality) that a worker achieves when they benefit from generative AI assistance. Real-world studies show an average gain of 14 to 15%, strongly concentrated on the least experienced profiles.
- Cost per tokenCost per token is the elementary economic unit of an AI deployment. Providers bill input tokens (your prompt) and output tokens (the model's response) separately, with a typical ratio of 1 to 5. Mastering this cost requires distinguishing unit prices, consumed volumes, and optimisation levers.
- Shadow AIShadow AI refers to the use of generative AI models by employees without validation or framing from their company. Widespread in 2025-2026, it creates legal, financial (leakage of business data to public models), and quality (unaudited outputs integrated into client deliverables) risks.
- Usage disciplineAI usage discipline is the set of rules, methods, and habits by which a company orients the daily use of AI by its employees: which use cases to activate, how to formulate requests, which models to use, how to audit outputs. It is the dominant lever of the real return of an AI deployment.
SEO and GEO(7)
- AEO (Answer Engine Optimization)AEO (Answer Engine Optimization) is a sub-discipline of GEO, focused on producing direct answers to user questions. It targets answer engines (Google AI Overviews, ChatGPT, Perplexity) that require a concise, sourced response, restituable as is, rather than a page to browse.
- CitabilityCitability is the property of a piece of content of being extracted, restituted, and attributed as is by an AI engine in its response to a user. It is built through writing (self-contained sentences), markup (Schema.org), and topical authority (consistency and editorial density on a domain). It is the central objective of GEO.
- Conversational indexingConversational indexing is the process by which AI engines build their responses from web content: ingestion at training (dated corpus), retrieval at inference (RAG via web search), or both combined. It replaces the keyword indexing of classic engines with indexing by questions and intentions.
- Generative referenceA generative reference is the mention of a source in an AI engine's response to a user: textual citation, explicit attribution, or hyperlink to the original content. It progressively replaces the classic organic link as a signal of editorial and brand visibility on the web.
- GEO (Generative Engine Optimization)GEO (Generative Engine Optimization) is the set of practices aimed at making content identifiable, trustworthy, and citable by generative answer engines (ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews). It succeeds traditional SEO when users seek answers, not links.
- llms.txtllms.txt is a proposed web standard introduced in September 2024 by Answer.AI, providing AI models with a structured map of a website's content. Placed at the root of the domain, it is not a formal standard and its adoption by major LLMs remains uneven in 2026.
- Schema.org for AISchema.org is a standard structured semantic markup vocabulary (most often JSON-LD) that AI engines use to understand the nature of a web page: defined term, FAQ, article, product, event. Properly implemented, it multiplies by two to four the probability of being cited by generative answer engines.