The AI glossary, in plain words.
Every AI term your vendor throws at you, defined in one honest paragraph and explained like you are smart but busy. 100+ entries, no jargon, one real example each.
Agent Economy
The agent economy is the emerging market in which AI agents transact directly with businesses and with each other: discovering services, negotiating terms, placing orders and paying for outcomes, with contracts and reputations attached.
ReadAgent Loop
The agent loop is the cycle an AI agent repeats while pursuing a goal: observe the current state, plan the next action, act through a tool, check the result, then continue, re-plan or stop.
ReadAgent Memory
Agent memory is the store that lets an AI agent remember what it has done and learned across steps and sessions: recent conversation, facts it has saved, and outcomes of past tasks, so it does not start every interaction from a blank page.
ReadAgentic Workflow
An agentic workflow is a business process whose steps are executed by AI agents: they plan the work, call tools, read results and decide the next move, instead of following a fixed script that breaks the first time reality differs.
ReadAI Accelerator
An AI accelerator is a processor built for one job: the matrix multiplications that neural networks are made of. GPUs, TPUs and NPUs are all accelerators; they trade the flexibility of a CPU for far higher throughput per watt on AI workloads.
ReadAI Agent
An AI agent is software that pursues a goal on your behalf: it reads the situation, plans the next step, acts through tools (email, calendars, databases, APIs) and checks the result before continuing, instead of answering one question and stopping.
ReadAI Audit
An AI audit is a structured review of how an organisation actually uses AI: every tool, model and workflow is inventoried, then tested for accuracy, data exposure, cost and compliance, with findings ranked by risk and a remediation plan attached.
ReadAI Bias
AI bias is a systematic skew in what a model predicts or decides, caused by skewed training data or skewed design, so it treats similar people or cases unfairly and consistently.
ReadAI Governance
AI governance is the set of policies, roles and controls that decide how an organisation may build, buy and use AI: who approves a use case, what data a model may see, how incidents are handled, and who is accountable when something goes wrong.
ReadAI Maturity
AI maturity measures how far an organisation has progressed from isolated AI experiments to AI as a managed part of everyday operations, usually assessed across dimensions like strategy, data, people, technology and governance.
ReadAI Model
An AI model is the finished artefact that comes out of training: a file of tuned numbers that maps inputs to predictions. You build it once and run it as many times as you like; running it is called inference.
ReadAI Readiness
AI readiness measures how prepared an organisation's data, processes, people and governance are to run AI in production, and it is assessed before any model is deployed.
ReadAI ROI
AI ROI is the money an AI system returns compared with everything it costs to build and run it: savings and new revenue weighed against licences, integration, data work and human oversight.
ReadAI Strategy
An AI strategy is the plan that turns AI from scattered experiments into a coordinated capability: which processes you automate, in what order, with what data, budget and governance, and when you stop.
ReadAI Watermarking
AI watermarking is the practice of embedding a machine-readable mark inside generated content (text, images, audio or video) so its origin can be verified later: either a visible label or an invisible signal baked into the pixels, the word choices or the sound frequencies.
ReadAlgorithm
An algorithm is a finite, unambiguous sequence of steps that solves a class of problems: a recipe that takes an input, applies rules, and always produces an output. Every piece of software, AI included, is algorithms all the way down.
ReadAlignment
Alignment is the work of making an AI system do what you actually meant, not merely what you literally typed: its goals, limits and refusals are shaped during training and deployment so its behaviour stays within what operators and society intend.
ReadArtificial General Intelligence
Artificial General Intelligence (AGI) is a hypothetical AI system with human-level competence across every cognitive domain: it could learn any new task from instructions or experience, transfer skills between unrelated fields, and plan long projects without being retrained or rebuilt for each one. No such system exists today; current models are general in what they accept but narrow in what they reliably do.
ReadArtificial Intelligence
Artificial intelligence is software that performs tasks which normally need human judgement: understanding language, recognising images, planning, deciding. Modern AI is built by training models on data rather than hand-writing rules, which is why it copes with messy input a rigid program cannot.
ReadAttention Mechanism
Attention is the mechanism inside a transformer that lets each token weigh how much every other token matters before being processed, so the model can track what 'it' refers to or which date belongs to which clause.
ReadAutomation
Automation is software that executes a defined task the same way every time, following explicit rules or a trained model, without a human doing each step by hand.
ReadAutonomous Agent
An autonomous agent is an AI agent that pursues a long-term goal with minimal supervision: it decides its own next actions, recovers from errors and keeps working until the goal is reached or it hits a hard boundary, without asking for approval at every step.
ReadBenchmark
A benchmark is a standardised test that ranks models on a fixed task, dataset and scoring method, so different systems can be compared on the same footing instead of on vendor claims.
ReadBuild vs Buy
Build vs buy is the decision between custom AI software shaped around your process and an off-the-shelf product you adopt and configure, weighed on cost, fit, speed and control.
ReadChain of Thought
Chain of thought is a prompting technique where you ask the model to reason step by step before answering, writing its intermediate steps out loud, which sharply improves accuracy on multi-step maths, logic and planning problems.
ReadChange Management
Change management is the discipline of moving people to a new way of working, so a new system gets used in production instead of being installed and quietly ignored.
ReadChunking
Chunking is the way long documents are cut into pieces before indexing: the cuts decide what the search can find, so a good chunk is a self-contained passage that keeps its meaning when read alone.
ReadCloud Compute
Cloud compute is renting processing power, storage and AI accelerators from a provider's data centre over the internet, paying per second, per gigabyte or per token instead of buying and running the machines yourself.
ReadComputer Use
Computer use is the capability of an AI agent to operate a computer like a person: it reads the screen through screenshots, moves the cursor, clicks, types and scrolls in any software, including programs that expose no API.
ReadContent Moderation
Content moderation is the pipeline that decides what user-submitted content (posts, images, comments, reviews) gets published, flagged or removed: automated classifiers screen everything at machine speed, clear cases are decided instantly, and the grey zone goes to human reviewers.
ReadContext Window
The context window is the model's working memory, measured in tokens: everything in a single prompt (instructions, documents, the conversation so far) must fit inside it, and anything beyond the limit is simply invisible to the model.
ReadCopilot
An AI copilot is software that assists a human inside the tools they already use, drafting and suggesting while the person keeps the final decision.
ReadData Pipeline
A data pipeline is the connected chain of steps that carries data from where it is produced to where it is consumed: ingest, validate, transform, store and serve, each stage handing clean, well-formed data to the next.
ReadDataset
A dataset is the organised collection of examples an AI learns from or is judged on: thousands or millions of rows of text, images or sensor readings, cleaned, labelled and split into training, validation and test parts.
ReadDeep Learning
Deep learning is machine learning with neural networks stacked in many layers, so the system learns the features by itself instead of a human hand-picking them. It is the technique behind today's language, image and speech models.
ReadDiffusion Model
A diffusion model is a generative model trained to reverse a gradual noising process: it learns to remove noise step by step, so at generation time it can start from pure static and refine it into an image, sound or other output that matches your prompt.
ReadDigital Workforce
A digital workforce is the set of AI agents and automations an organisation employs alongside its human staff, each with a defined role, tools and supervision, measured against business outcomes rather than licences bought.
ReadDistillation
Distillation is a training technique where a small 'student' model learns to reproduce the answers of a large, accurate 'teacher' model, so you can deploy the small one and keep most of the accuracy at a fraction of the running cost.
ReadEdge AI
Edge AI is running AI models on the device where the data is born: a camera, a sensor, a phone, a factory gateway, instead of shipping everything to a cloud data centre and waiting for the answer.
ReadEmbedding
An embedding is a list of numbers that captures the meaning of a piece of text, image or audio, so that content with similar meaning gets similar numbers and the computer can measure closeness with plain arithmetic.
ReadETL
ETL (extract, transform, load) is the process of pulling data out of source systems, reshaping it to fit a target schema, and loading it into a warehouse or database where reports and AI can rely on it.
ReadEU AI Act
The EU AI Act is the European regulation that sorts AI systems into risk levels: unacceptable practices are banned outright, high-risk systems must meet documentation, data-quality, logging and human-oversight requirements before they can be sold or deployed, and the rest face lighter transparency duties.
ReadExplainability
Explainability is the property of an AI system (or of a layer added to it) that lets you see why it gave a specific answer: which inputs mattered, which rules fired, and what would have changed the outcome.
ReadFine-tuning
Fine-tuning is additional training that teaches a general model your specific patterns: you show it hundreds or thousands of examples of your tasks, your tone and your formats, and it adjusts its internal weights until that behaviour becomes its default.
ReadFunction Calling
Function calling is the structured form of tool use: the model emits a machine-readable request naming a function and its arguments, written in JSON, which your code validates and executes, then feeds the return value back into the conversation.
ReadGAN
A GAN (Generative Adversarial Network) is a pair of neural networks trained against each other: a generator creates synthetic samples while a discriminator judges them against real data, and each improves by trying to beat the other.
ReadGenerative AI
Generative AI is software that produces new content, text, images, audio, code, by sampling patterns learned from vast training data, rather than only classifying or retrieving what already exists.
ReadGPU
A GPU is a processor built to run thousands of small calculations at the same time. Designed for graphics, it turned out to be the perfect engine for AI, because running a neural network is mostly one enormous pile of matrix multiplications.
ReadGuardrails
Guardrails are the checks and limits wrapped around an AI model so its behaviour stays inside your rules: they screen what goes in, constrain what the model can do, and verify what comes out before it reaches a person or a system.
ReadHallucination
A hallucination is content an AI model states with full confidence but that is not true: invented facts, fabricated citations, wrong numbers, delivered in exactly the same tone as its correct answers.
ReadHuman-in-the-Loop
Human-in-the-loop is a design pattern where an AI system pauses at checkpoints you define and asks a person to approve, edit or reject its proposed action before continuing, keeping consequential decisions under human control.
ReadImage Generation
Image generation is the process of creating new pictures from a text description or another image: a generative model, usually a diffusion model, turns a prompt into pixels step by step rather than copying existing photos.
ReadInference
Inference is using a trained model: you send input in, the model applies what training encoded in its parameters, and you get an answer out, one token at a time.
ReadInference Optimization
Inference optimization is the engineering discipline of making a trained model cheaper and faster to run: measure where time and compute go on each request, then apply levers like caching, quantization, batching and prompt trimming to cut both.
ReadInterpretability
Interpretability is the degree to which you can look inside an AI model and understand what its internal parts represent and how they combine to produce an answer, as opposed to explainability, which constructs a plausible justification after the decision is made.
ReadJailbreak
A jailbreak is an attempt to make an AI model ignore its safety training through a cleverly disguised prompt, so it produces content it was built and tuned to refuse.
ReadKnowledge Graph
A knowledge graph is a database of things and relationships: entities (people, parts, sites, contracts) are nodes, typed relationships are edges, and questions are answered by following paths through the graph.
ReadLarge Language Model
A large language model (LLM) is an AI system trained on massive collections of text to predict the next token, which gives it a general ability to write, summarise, translate and reason without being programmed for each task.
ReadLatency
Latency is the delay between a request and its answer, measured in milliseconds. In AI systems it adds up across queueing, prompt processing, token-by-token generation and network hops, and it decides whether a tool feels instant or unusable.
ReadLLMOps
LLMOps is the discipline of running large language models in production: managing prompts, context, costs, latency and outputs with the same rigour MLOps applies to predictive models, adapted to models that answer in words and can say almost anything.
ReadMachine Learning
Machine learning is a way of building software where the algorithm tunes its own internal numbers from examples, instead of a programmer writing every rule. You supply data and correct answers; the finished product is a model that predicts on inputs it has never seen.
ReadMLOps
MLOps is the set of practices, tools and pipelines that take a machine learning model from a data scientist's notebook into production and keep it running reliably: versioning data and code, training, evaluating, deploying and monitoring the model like any other business-critical software.
ReadModel Context Protocol
Model Context Protocol (MCP) is an open standard that lets an AI agent connect to your tools and data (files, databases, CRMs, APIs) through one common plug, instead of a custom integration for every tool.
ReadModel Drift
Model drift is the silent loss of accuracy that happens when live data stops resembling the data a model was trained on: the world moves, the model stays still.
ReadModel Evaluation
Model evaluation is the process of measuring how well a model performs on data it has never seen, using metrics and test sets agreed in advance, before you trust it with real decisions.
ReadMulti-Agent System
A multi-agent system is a team of specialised AI agents working on one goal: each agent has its own role and tools, and a coordinator divides the work, collects the results and resolves conflicts, instead of one model trying to do everything alone.
ReadMultimodal Model
A multimodal model is one AI model that handles more than one kind of input and output, text, images, audio, video, reasoning across them in a shared internal representation instead of requiring a separate tool per format.
ReadMusic Generation
Music generation is the use of a generative model to compose new music from a text prompt or a reference track: the model learns the statistical patterns of melody, harmony and rhythm from training data and renders new audio, a full mix or editable stems.
ReadNeural Network
A neural network is a function built from layers of simple units. Each unit multiplies its inputs by weights, adds them, passes the sum through a small nonlinearity, and forwards the result. Training adjusts the weights until the whole stack produces useful outputs.
ReadNPU
An NPU is a neural processing unit: a low-power AI accelerator built into laptops, phones and industrial devices, so that inference runs on the device itself with no round trip to the cloud.
ReadObservability
Observability is the practice of recording what a system actually did in production (every input, output, step, timing and cost), so you can see its real behaviour, diagnose failures and prove what happened, instead of inferring from averages.
ReadOCR
OCR (optical character recognition) is software that turns images of text, such as scanned pages, photographed delivery notes or exported PDFs, into machine-readable text that other programs can search, count and edit.
ReadOn-Premise AI
On-premise AI is running AI models on servers you own, inside your building or your private data centre, instead of a cloud provider's machines. Your data stays on your hardware and you control the entire stack, from the GPU up to the model.
ReadOrchestration
Orchestration is the coordination layer of a multi-agent system: one orchestrator receives a goal, splits it into tasks, assigns them to specialised agents, collects their outputs, resolves conflicts and assembles the final result.
ReadOverfitting
Overfitting is when a model memorises its training data instead of learning the underlying pattern: near-perfect on examples it has seen, unreliable on anything new.
ReadParameter
A parameter is one of the internal numbers a trained model stores: billions of them, each tuned during training, and together they encode everything the model learned.
ReadPlanner
A planner is the component of an AI agent that turns a goal into an ordered sequence of steps before any action is taken, deciding what to do in which order instead of improvising one move at a time.
ReadPrompt
A prompt is everything you hand to a language model in one call: instructions, context and the question itself. The model answers by continuing exactly that text, and nothing else.
ReadPrompt Engineering
Prompt engineering is the practice of writing prompts so a language model performs a task reliably: clear instructions, the right context, worked examples of the desired output, and tests against real cases before you trust it.
ReadPrompt Injection
Prompt injection is an attack where hidden instructions are planted inside content an AI reads (an email, a web page, a document), so the model executes a stranger's commands as if they were yours.
ReadPrompt Versioning
Prompt versioning is treating prompts as code: every change is numbered, stored in a repository, tested against a fixed set of examples, and rolled back when a new version regresses.
ReadQuantization
Quantization is a compression technique that rewrites a model's weights from high-precision decimals (32-bit floats) to smaller numbers (8-bit or 4-bit integers), so the model takes up less memory and runs faster on cheaper hardware, at a small, measurable cost in accuracy.
ReadRed Teaming
Red teaming is the authorised practice of attacking your own AI system to expose weaknesses before a hostile actor does: a dedicated team fires adversarial prompts, jailbreak attempts and abuse scenarios at it, records what gets through, and feeds every finding back into a fix.
ReadReranking
Reranking is a second, slower pass over search results: a cross-encoder model reads each candidate together with the query and rescores it for true relevance, so the top results are the actually-right ones, not just the roughly-similar ones.
ReadRetrieval-Augmented Generation
RAG is an architecture where the AI retrieves relevant documents from your knowledge base before writing an answer, so it quotes your data instead of guessing from memory.
ReadRLHF
RLHF, reinforcement learning from human feedback, is the technique that tunes a language model's behaviour with human taste: people rank sample answers, a reward model learns to predict those preferences, and the language model is trained to score higher.
ReadSemantic Search
Semantic search finds content by meaning instead of exact words: the query and the documents are converted into embeddings, and the closest matches are returned even when no keyword overlaps.
ReadStructured Output
Structured output is a mode where the model is forced to answer in a machine-readable format (JSON, a table, a fixed schema) instead of free text, so its answer can be validated, stored and passed to other software without a human retyping it.
ReadSynthetic Data
Synthetic data is artificial data engineered to reproduce the statistical patterns of real data: a model or a set of rules generates new records that behave like the originals without containing any real person's or customer's information.
ReadSystem Prompt
The system prompt is the standing instruction a language model receives before any user message: it defines the model's role, tone, limits, output format and the tools it may use, and it shapes every answer in the conversation.
ReadTemperature
Temperature is a setting on a language model that controls how random its word choices are: low values make it pick the most likely next word every time, high values let less likely, more varied options win.
ReadText-to-Image
Text-to-image is the task of producing an image from a written description: a text encoder converts your prompt into a numeric representation, and an image model uses it as guidance to generate pixels that match the description.
ReadToken
A token is the chunk of text a language model actually reads and writes: on average about three quarters of an English word, converted to a number from a fixed vocabulary before the model can process it.
ReadTokenization
Tokenization is the step that turns text into tokens, the small units a language model actually reads: the input is split according to a fixed vocabulary and each token is mapped to a number, so the model processes IDs instead of words.
ReadTool Use
Tool use is the capability of an AI model to do more than write text: it can call external tools, such as searching a database, sending an email or running a calculation, by emitting a structured request that software executes for real.
ReadTPU
A TPU is Google's custom chip built for a single job: the matrix multiplications at the heart of neural networks. You do not buy it; you rent it by the hour inside Google Cloud, where it often beats general-purpose GPUs on training cost.
ReadTraining
Training is the process that turns an untrained model into a working AI: you feed it thousands or millions of examples, compare its guesses with the correct answers, and nudge its internal settings until the guesses get good.
ReadTransformer
The transformer is the neural network architecture, published by Google researchers in 2017, behind every modern language model: it processes all tokens of a text at once and uses attention to let each token weigh the relevance of every other token.
ReadUnstructured Data
Unstructured data is information with no fixed schema: documents, emails, images, audio, chat logs, PDFs. Most of what a company knows lives in this form, but software must first parse, transcribe or embed it before it can be searched or counted.
ReadVector Database
A vector database is a database built to store numeric fingerprints of content (embeddings) and find the closest ones to any query in milliseconds, powering semantic search, recommendations and AI retrieval.
ReadVideo Generation
Video generation is the process of producing moving footage from a text prompt, script or storyboard: a generative model creates frames that stay consistent with each other, adds motion between them, and the audio is synthesised separately.
ReadVoice Cloning
Voice cloning is the process of building a synthetic copy of a specific person's voice from audio recordings: a model learns that speaker's timbre, rhythm and accent, and can then read any new text aloud in their voice.
ReadVRAM
VRAM is the dedicated memory on a graphics card or AI accelerator that holds the model's weights and the intermediate numbers of each calculation, so the processors doing the AI work never wait on slower system RAM.
Read§ 02
Terminology is the easy part. Shipping is ours.
We build the agents and systems behind these words. It starts with a 30-minute scoping call.