Artificial intelligence glossary

Agent memory is the information an AI agent keeps so it can use it later: what has been said in the current conversation (short-term) and stored details from earlier conversations, such as a customer's preferences or history (long-term).

Agent orchestration is the coordination of several specialised AI agents so they complete a task together: one component decides which agent acts at each step, what information it receives and how the results are combined.

Agentic AI is AI that pursues a goal instead of just answering: it plans steps, uses tools, checks the results and decides what to do next with a degree of autonomy, within rules set by a person.

AGI (artificial general intelligence) is a hypothetical AI able to learn and perform almost any intellectual task a person can, with the same flexibility. There is no universally agreed definition and no system widely accepted as AGI.

An AI agent is a system that uses a language model to understand a request, decide which steps to take and carry them out with real tools, such as your calendar or CRM, until the task is done rather than just replying.

AI automation means using artificial intelligence models to handle tasks that used to need human judgement, such as reading a message, working out what it asks for and choosing the next step, rather than just repeating fixed steps defined in advance.

AI bias is a systematic error that leads an artificial intelligence system to treat similar people or cases differently, often unfairly. It usually stems from unbalanced training data or from design choices.

An AI copilot is an assistant built into a work tool that suggests, drafts or summarises while a person works, without acting on its own: the final decision and the send button always stay with the user.

An AI hallucination is a response from an AI model that sounds convincing but is false or not backed by any source: a figure, quote, price or reference the model invents while generating the most probable text.

An AI model is the program you get by training an algorithm on data: a set of learned parameters that, given a new input, produces an output such as a prediction, a label, a piece of text or an image.

An algorithm is a finite sequence of well-defined steps for solving a problem or completing a task. In AI, the algorithm is the method that learns from data; the model is what that method produces.

An API (application programming interface) is a set of rules that lets one program ask another for data or an action: you send a request in an agreed format and get a response back, without needing to know how the other system works inside.

Artificial intelligence (AI) is the branch of computing that builds systems able to do tasks we associate with human intelligence, such as understanding language, recognising images or making decisions, by learning patterns from data.

Automatic transcription is the process of turning an audio or video recording into written text without anyone typing it, using speech recognition. It usually adds punctuation, timestamps and labels showing who is speaking at each point.

An AI benchmark is a standardised test, with a set of questions or tasks and a fixed scoring method, used to measure and compare how different models perform at a specific skill such as reasoning, coding or answering questions.

A callbot is a voicebot that works inside phone calls: it answers inbound calls or places outbound ones, talks with the person by voice and handles tasks such as bookings, reminders or lead qualification, connected to your phone system and your tools.

Chain of thought is a technique where an AI model writes out the intermediate steps of its reasoning before giving the final answer. It improves results on calculations, chained rules and multi-step problems.

A chatbot is a program that holds text conversations with people, on a website or on WhatsApp, to answer questions or guide them through a task. It either follows fixed rules or uses a language model to understand what people write.

Computer vision is the branch of AI that lets a system interpret images and video: recognising objects, reading text, spotting defects or describing a scene. Today it mostly relies on neural networks trained on large numbers of images.

A context window is the maximum amount of text, measured in tokens, that a language model can take into account at once: instructions, conversation, documents and its own reply. Anything outside it, the model simply doesn't see.

A conversation flow is the predefined path of an automated conversation: the steps, questions and replies a bot follows depending on what the user answers, usually drawn as a branching diagram.

Conversational AI is the set of technologies that let a machine hold a conversation in natural language, by text or voice: understand what is said, keep track of the context and reply in a coherent way.

A dataset is an organised collection of examples, such as text, audio, images or records, used to train, fine-tune or evaluate an AI model. Its quality directly shapes what the model learns.

Deep learning is a type of machine learning that uses neural networks with many layers. Each layer picks out more abstract features of the data, which lets it handle text, speech and images without hand-written rules.

An embedding is a list of numbers that represents the meaning of a piece of text, an image or audio. Items with similar meaning have embeddings that sit close together, which lets you search by meaning rather than by exact words.

The EU AI Act is Regulation (EU) 2024/1689, the European Union's artificial intelligence law. It sorts AI systems by level of risk and sets obligations for those who build and use them, from banning some uses outright to transparency duties.

Few-shot prompting is a technique where you include a handful of worked examples in the instruction, so the AI model copies the format, tone or judgement you want without having to be retrained.

Fine-tuning means further training an AI model that's already been pre-trained, using a small set of your own examples, so it specialises in a particular task, format or style without starting from scratch.

Function calling is a language model's ability to reply, instead of with text, with a structured request to use a specific function, giving its name and arguments. Your system runs that function and hands the result back to the model.

Generative AI is artificial intelligence that creates new content, such as text, images, audio, video or code, from an instruction, using models trained on huge numbers of examples of that kind of content.

GPT (generative pre-trained transformer) is OpenAI's family of language models. The name sums up how they work: they generate text, are pre-trained on huge amounts of data and are built on the transformer architecture.

Grounding is the practice of making an AI model base its answers on specific, verifiable sources, such as your documents, a database or a search, rather than on what it remembers from training.

Guardrails are the safeguards placed around an AI model to keep it within agreed limits: checks on what goes in and comes out, limits on topics and actions, and rules for handing a case to a person when needed.

Human handoff is the point at which a chatbot or AI agent passes the conversation to a member of staff, together with the context it has gathered, because the case needs it, a rule requires it or the customer asks for it.

Inference is when an already-trained AI model is put to use: it receives a new input, such as a question, an audio clip or an image, and computes a response from what it learned, without changing its parameters.

An intent is the goal behind what a person writes or says to a chatbot or AI agent, such as booking, cancelling or asking for a price. Recognising it lets the system decide what to do with the message.

An AI jailbreak is an attempt to get a language model to bypass its safety rules or the instructions of whoever set it up, using messages designed to trick it or pull its attention elsewhere.

A knowledge base is a business's organised store of information (FAQs, terms, opening hours, product sheets, procedures) that an AI agent or a team draws on for answers.

A language model is an AI system that calculates which word, or token, is most likely to come next in a piece of text. By repeating that prediction it can complete sentences, answer questions, summarise or translate.

Latency is the time between making a request and starting to receive the response. In a voice agent, it's the silence from the moment the caller stops talking to the moment they hear the agent's first word, and it decides whether the call sounds natural.

Lead scoring is a way of giving each sales contact a score based on how close they are to buying, combining who they are (sector, size, location) with what they've done (questions, visits, replies), so the sales team deals with the most promising first.

An LLM (large language model) is an AI model trained on vast amounts of text to predict the next word. That single skill lets it draft, summarise, translate, classify and hold a conversation.

Machine learning is the branch of AI in which a system learns to do a task from examples, rather than following hand-written rules, and gets better as it sees more data.

MCP (Model Context Protocol) is an open standard that defines how an AI application connects to external tools and data sources. Each service is exposed once as an “MCP server”, and any compatible assistant can use it without a bespoke integration.

Mixture of experts (MoE) is an AI model architecture split into several sub-networks called experts. For each token, a router switches on only a few of them, so the model can be very large without using all of its compute at every step.

A model's parameters are its internal numeric values, mainly the weights of a neural network, which are adjusted during training and store what it has learned. Their count is used as a rough measure of model size.

Model training is the process in which an AI model learns from a dataset: it's shown examples, its errors are measured and its parameters are adjusted, over and over, until its outputs are good.

A multimodal model is an AI model that can take in or produce several types of data, such as text, images, audio or video, and connect them. For example, it can read a photo of an invoice and answer questions about it.

Natural language processing (NLP) is the branch of artificial intelligence that lets machines work with human language: read, analyse, classify, translate and produce text or speech the way people use it.

Natural language understanding (NLU) is the part of NLP that deals with the meaning of a message: what the person wants (their intent) and which specific details they give (entities such as dates, names or products).

An artificial neural network is a mathematical model made of layers of connected units, called neurons, that turn a numeric input into an output. It learns by adjusting the weight of each connection based on examples.

OCR (optical character recognition) is the technology that turns the text in an image, a scanned PDF or a photo into digital text you can edit, search and process. With AI, it also understands the document's layout and structure.

Open-source AI is AI whose model is released so anyone can use, study, modify and share it. In practice, many so-called open models publish only their weights, not the training data or all of the training code.

A prompt is the instruction or text you give an AI model to get something done: a question, a command, a document to summarise or a mix of all three. Much of what the model gives back depends on the prompt.

Prompt engineering is the work of designing, testing and refining the instructions given to an AI model so that it answers well consistently, not just once.

Prompt injection is an attack in which someone slips instructions into text an AI model reads, such as a message, an email or a web page, to make it do something other than what whoever set it up intended.

Quantization is a technique that stores an AI model's parameters as lower-precision numbers, for example 8 or 4 bits instead of 16. The model takes up less memory and often runs faster, in exchange for a loss of quality that can be small.

RAG (retrieval-augmented generation) is a technique in which a language model first retrieves the relevant passages from your own documents and then writes its answer from them, instead of relying on what it memorised in training.

A reasoning model is a language model trained to produce a chain of intermediate steps before giving its final answer. It does better on logic, maths and planning problems, at the cost of slower replies and more tokens used.

Sentiment analysis is the AI technique that detects the emotional tone of a text (positive, negative or neutral, and sometimes emotions such as anger or urgency) so you know how the writer feels without reading everything by hand.

A small language model (SLM) is a language model with far fewer parameters than an LLM. It knows less about everything, but it replies faster, costs less and can run on a modest server, a laptop or even a phone.

Speech recognition is the branch of artificial intelligence that lets a machine identify the words a person says and turn them into text or commands. Its technical name is ASR, short for automatic speech recognition.

Speech synthesis is the artificial production of human speech by computer. It's the discipline behind text to speech, voice cloning and talking agents, and today it relies mainly on neural networks trained on real recordings.

Speech to text (STT) is the technology that turns spoken audio into written text. It's the software component, usually a model reached through an API, that does the listening in a voicebot, a dictation tool or a transcription service, either live or on recordings.

A system prompt is the set of fixed instructions an AI model receives before any conversation starts. It sets the model's role, its tone, what it knows about the business and what it isn't allowed to do.

Temperature is a language model setting that controls how much randomness goes into choosing each token. At low values the model nearly always picks the most likely option and replies predictably; at high values it varies more and comes across as more creative.

Text classification is the AI task of automatically assigning one or more categories to a piece of text, such as an email or a message: type of request, topic, priority or language. It's the basis for sorting, prioritising and routing work.

Text to speech (TTS) is the technology that turns written text into spoken audio. It's the component, usually a model reached through an API, that gives a voice to a phone agent, a screen reader or any audio generated from text.

A token is the chunk of text a language model works with: a word, part of a word or a punctuation mark. Models read, count and bill in tokens, not in words or characters.

Tokenization is the process of splitting text into tokens, the pieces a language model understands, and mapping each one to a number in its vocabulary. It's the first step before a model can read or write anything.

A transformer is a neural network architecture that processes every word in a text at once and uses an attention mechanism to work out which parts relate to each other. It's the foundation of almost every language model in use today.

A vector database stores embeddings, lists of numbers that represent the meaning of content, and finds the ones closest to a query very quickly. It's the search component in most RAG systems.

A virtual assistant is AI software that helps a person, or a company's customers, with a range of tasks by text or voice: it answers questions, keeps track of the conversation and handles simple jobs such as booking an appointment.

A voicebot is a program that holds spoken conversations: it listens to what a person says, works out what they want and answers out loud. It can run on a phone line, a website, an app or a smart speaker, and with generative AI it handles open conversation, not just menus.

A webhook is an automatic notification that one system sends to another system's web address when an event happens, such as a new booking or an incoming message, so the second system can react straight away without having to ask.

Zero-shot is when an AI model completes a task from the instruction alone, without any prior examples of how to do it. It manages thanks to what it already learned during training.