Resources
AI glossary: the terms worth understanding
AI has created its own vocabulary, and much of it is misused, overused, or thrown around to sound sophisticated without saying anything concrete. This glossary gathers the terms that come up most often in real conversations about AI at work and explains them plainly. We’ll keep adding to it over time.
The basics
Artificial intelligence (AI) An umbrella term for systems that can perform tasks that usually require human intelligence: recognizing speech, solving problems, learning from experience, and making decisions. It describes a broad field, not one specific technology. It covers everything from simple rule-based systems to the most sophisticated models on the market.
Machine learning A branch of AI in which a system learns patterns from data instead of following instructions programmed step by step. You don’t tell it “if this happens, do that.” You give it examples, and it learns to generalize.
Deep learning A type of machine learning that uses neural networks with many stacked (“deep”) layers, loosely inspired by how biological neurons work. It underpins most recent advances in computer vision, speech, and language.
Model The result of training an AI system on data. It is the file or set of parameters that takes an input and produces an output. When people talk about “using AI,” they almost always mean using a model that has already been trained.
Parameters The internal numerical values a model adjusts during training to represent patterns in language or data. More parameters don’t always mean better results, but they usually mean higher costs and slower performance.
Rule-based automation A system that follows explicit logic written by a person, such as “if this happens, do that.” It is transparent and easy to audit, but it does exactly what it was programmed to do. It is not AI in the strict sense, although it is often marketed as if it were.
Language models
LLM (large language model) A model trained on enormous amounts of text to predict which word is most likely to follow another. That is what gives it the ability to write, summarize, translate, and hold a conversation.
Foundation model A model trained on broad, varied data and designed to serve as the basis for many different applications, rather than to solve just one task. Most commercial LLMs are foundation models.
Frontier model A term for the most capable and recent models available at a given time, usually trained by a small group of labs with the capital and infrastructure to do so. It is a relative label: a frontier model in 2024 is a mid-range model today. The question that matters is not whether a model is “frontier,” but whether its extra capability is needed for the task at hand.
Context window The amount of text a model can “keep in mind” at once when generating a response, measured in tokens. A large context window lets it work with long documents without “forgetting” the beginning.
Token The smallest unit a language model uses to process text. It is not always a whole word: it can be a syllable, a punctuation mark, or part of a word. The cost of using an LLM is almost always measured in the tokens it processes.
Hallucination When a model produces information that sounds convincing but is wrong or made up. It is not a random stroke of “bad luck.” It is often linked to the context it was given or to asking it something it cannot answer reliably.
Fine-tuning Taking a trained model and training it further on data specific to a field or company, so it can specialize in a particular task without starting from scratch.
Prompt The instruction or question given to an AI model. A prompt’s quality, including its context, requested format, and examples, directly affects the quality of the answer.
RAG (retrieval-augmented generation) A technique that lets a model answer questions using a company’s own documents, even if it was not trained on them. The system first searches a private database for relevant information, then gives it to the model along with the question so it can generate an answer.
Embedding A numerical representation of a text’s meaning that lets a system compare how similar two pieces of content are. It is what allows a RAG system to find the right document among thousands.
Agents and automation
Agentic AI An AI system that does more than generate text. It pursues a goal on its own: plans steps, uses real tools such as web search, an API, or a file editor, observes the result, and adjusts its next action. It repeats this process until it completes the task. Agency is a spectrum, not a yes-or-no label: an assistant that only searches the web is slightly agentic, while a system that manages a full sales process without supervision is highly agentic.
Agent A specific instance of an agentic AI system, configured for a particular task or role, such as customer support or classifying incoming invoices.
Agent orchestration Coordinating several specialized agents as they work together on a complex task, each handling one part: one retrieves information, another analyzes it, and a third writes the response.
Human in the loop A design in which a person reviews or approves an AI system’s decisions before they are carried out, especially for high-risk tasks. It is the opposite of full autonomy.
MCP (Model Context Protocol) An open standard that lets an AI model connect in an organized way to external tools such as calendars, databases, and email, without building a separate integration for each one.
Voice and conversation
STT (speech to text) Technology that turns spoken audio into written text. It is the first step in any voice assistant.
TTS (text to speech) The reverse process: turning written text into generated speech. It determines how natural a voice agent sounds when it speaks.
Latency The time a system takes to respond after receiving an input. On a voice call, high latency feels like an awkward silence before the agent answers, and it is one reason many voice agents sound artificial.
Data and governance
Bias When a model reproduces or amplifies unfair patterns in the data it was trained on. It is not a flaw unique to AI: it reflects the source data, which in turn reflects earlier human decisions.
Explainability The ability to understand why a model made a particular decision. Rule-based systems are easy to explain by design; many modern AI models are harder to interpret, which creates friction in regulated industries.
AI governance The internal rules, responsibilities, and processes an organization establishes to decide which AI systems it uses, with what data, under what oversight, and within what limits.
AI washing A marketing practice that exaggerates or invents the role AI actually plays in a product or service to make it seem more innovative than it is.
We’ll keep adding new terms to this glossary. If you’re looking for an AI concept that isn’t here, tell us and we’ll add it.
Sources
Get your team ready to work with AI.