122 essential AI terms for work and everyday life — in plain English, with why each one matters and what it looks like in practice.
The core ideas everything else is built on.
Computer systems that perform tasks normally requiring human intelligence — understanding language, recognizing patterns, making decisions, and generating content.
A branch of AI in which models learn patterns from data to make predictions or decisions, rather than relying only on hand-written rules.
A type of machine learning that uses many-layered neural networks to learn complex patterns from large amounts of data.
A model loosely inspired by the brain, made of interconnected “nodes” that pass signals to recognize patterns.
A set of step-by-step instructions a computer follows to solve a problem or complete a task.
The trained product of a machine-learning process — a file that takes an input and produces an output (text, a prediction, an image).
The examples used to teach a model. Quality and relevance of this data largely determine how well the model performs.
The act of running a trained model to get an answer — i.e., actually using the AI, as opposed to training it.
A proposed form of AI with broad, adaptable capabilities across many kinds of intellectual work. There is no universally agreed definition or test for AGI.
Machine learning that learns from examples paired with the correct labels or answers.
Machine learning that finds patterns or groups in data without being given correct answers for each example.
Training a system to choose actions using feedback expressed as rewards or penalties.
When a model learns its training examples too closely and performs poorly on unfamiliar data.
The numerical values a model learns during training that shape how it processes inputs and produces outputs.
The technology behind chatbots, copilots, and content tools.
AI that creates new content — text, images, audio, video, or code — rather than just classifying or predicting.
A model trained on vast amounts of text that can understand and generate human-like language.
A large, general-purpose model trained on broad data that can be adapted to many tasks.
AI that works across more than one type of input or output — text, images, audio, and video together.
A unit a model processes, such as a word fragment, punctuation mark, or part of an image or audio representation. Tokenization depends on the model and input type.
The maximum number of tokens a model can handle in one request, subject to its input and output limits.
A numerical representation of text (or images) that captures meaning, so a computer can measure how similar two things are.
The field focused on helping computers understand and work with human language.
Open-source AI makes specified components available under licenses that allow use, study, modification, and sharing. Proprietary models restrict access or reuse; models with downloadable weights are not automatically open source.
A model near the leading edge of demonstrated AI capabilities at a given time, often across a broad range of tasks.
A compact, efficient model that runs faster and cheaper than a frontier model, sometimes even on a phone.
A neural-network architecture that uses attention to process relationships between parts of an input.
A mechanism that lets a model assign different importance to different parts of its input while processing information.
A model whose trained numerical weights are available to download, subject to its license.
Training a smaller model to reproduce useful behavior from a larger model using examples or other training signals.
Representing model weights or calculations with lower-precision numbers to reduce memory and computation needs.
How to actually talk to AI and get reliable results.
The instruction or question you give an AI to get a response.
The practice of crafting and refining prompts to reliably get high-quality, useful outputs.
Behind-the-scenes instructions that set an AI's role, tone, and rules for an entire conversation.
A sampling setting that influences randomness when a model selects output tokens. Lower values generally produce less varied responses; higher values permit more variation.
A sequence of intermediate reasoning steps used while solving a problem. Some prompting methods elicit steps, while some models keep their internal reasoning hidden.
A newer class of model that spends extra time “thinking” through a problem before answering, trading speed for accuracy.
Giving a model a handful of examples inside the prompt so it follows the pattern you want — no retraining required.
Asking AI to do a task with no examples — just a plain instruction — and trusting it to figure it out.
Structured tests that measure how well a model or AI feature performs on the tasks you actually care about.
Designing the instructions, documents, tool results, and conversation history an AI receives for a task.
An AI response constrained to a specified format or schema, such as fields in a JSON object.
Shortening conversation history or tool results to keep essential information within a model's context window.
Information an application stores and retrieves across interactions to help an AI maintain continuity.
A standardized set of tasks used to compare model or system performance.
Building software by describing desired behavior to an AI and iterating on the generated code.
When AI stops chatting and starts doing the work.
An AI system that can take actions to accomplish a goal — browsing the web, running code, calling other software, and completing multi-step tasks.
AI systems that work autonomously toward goals, making decisions and chaining steps together rather than waiting for each instruction.
An AI's ability to use external tools — search, a calculator, your CRM, a calendar — by calling them as functions.
An open protocol that standardizes how AI applications connect to servers offering tools, resources, and prompts.
Coordinating multiple AI steps, models, or agents so they work together in a reliable sequence.
A setup where AI does the work but pauses for a person to review or approve before key actions happen.
An AI that works alongside a person inside their existing tools, suggesting and helping rather than acting fully on its own.
Using software (increasingly AI-driven) to run repetitive, multi-step business processes with little manual effort.
Several specialized AI agents working together, each handling part of a job and passing results to the next.
AI that operates a real computer or web browser — clicking, typing, and navigating apps like a person would.
The software around a model that manages its tools, instructions, state, permissions, and execution loop.
Passing a task and relevant context from one agent or workflow stage to another.
A reusable package of instructions, and sometimes scripts or resources, that guides an agent through a specific type of task.
An isolated environment that restricts what software or an AI agent can access or change.
A required authorization step before a workflow performs a specified action.
A property that lets an operation be repeated without applying its effect more than once.
Always-on agents are AI systems that remain available over time to carry out standing instructions, monitor events, or resume work without a fresh prompt for every step. They typically combine persistent state with schedules or event triggers.
The execution environment that runs an agent, invokes models and tools, and manages its lifecycle. A runtime may provide scheduling, isolation, state storage, and recovery.
The repeated cycle in which an agent receives context, chooses a next action, uses a tool or produces output, observes the result, and decides whether to continue.
An AI agent designed to work on an extended task across many steps, sessions, or context windows while maintaining progress.
An AI agent that performs work outside the user's immediate interactive session and reports progress or results later.
An AI agent activated at a specified time or recurring interval to perform an assignment.
An AI agent activated by an event such as a new message, changed record, uploaded file, or monitoring alert.
An execution approach that persists workflow progress so work can resume after interruptions or failures, rather than restarting the entire process.
The information an agent or its application maintains about its current task, such as goals, completed steps, tool results, and pending approvals.
A saved representation of agent or workflow progress that can be used to inspect, resume, or recover a task.
The Agent2Agent protocol is an open standard that lets agent applications discover capabilities, exchange messages, and coordinate tasks across systems.
The practice of designing, testing, and improving the software around an AI model that controls agent behavior, tool use, context, state, permissions, and recovery.
Making AI answer from your documents, not the open web.
A technique that lets an AI look up relevant information from your own documents before answering, instead of relying only on its training.
Further training an existing model on your specific examples so it specializes in your tasks, tone, or domain.
A database that stores embeddings so AI can find information by meaning, not just exact keywords.
The curated collection of documents and facts an AI draws on to answer questions about your business.
Connecting an AI response to external evidence, such as retrieved documents, database results, or tool outputs.
A standard way for software systems to talk to each other and exchange data or services.
A program that converses with users in natural language, today usually powered by an LLM.
Structured data is neatly organized (spreadsheets, databases); unstructured data is free-form (emails, PDFs, images, calls).
Artificially generated records or examples used to simulate data for training, testing, or analysis.
Search that retrieves information based on meaning rather than only matching exact words.
Search that combines keyword matching with meaning-based retrieval.
Splitting documents into smaller pieces for indexing, retrieval, or processing by an AI system.
Reordering retrieved search results using an additional relevance-scoring step.
Technology that converts text in images or scanned documents into machine-readable text.
Converting spoken audio into written text, also called automatic speech recognition.
Generating spoken audio from written text.
Where AI delivers value — and how to use it safely.
AI that analyzes historical data to forecast future outcomes.
Using AI to detect the emotional tone behind text — positive, negative, or neutral.
AI that interprets images and video — recognizing objects, reading text, or detecting defects.
When an AI states false information confidently as if it were fact.
Systematic unfairness in AI outputs, usually inherited from skewed or unrepresentative training data.
Instructions, checks, filters, and access restrictions intended to keep an AI system within approved boundaries.
The policies and oversight that ensure AI is used responsibly, legally, and in line with company values.
Protecting personal and confidential information, including how it's handled when sent to AI tools.
An attack where hidden malicious instructions trick an AI into ignoring its rules or leaking information.
Employees using unapproved AI tools at work without oversight.
AI-generated fake video, audio, or images that convincingly imitate a real person.
Laws and regulatory requirements that govern how AI systems are developed, supplied, or used. Requirements vary by jurisdiction, sector, and use case.
How clearly an AI system can show why it produced a particular answer or decision.
Methods for making an AI system's behavior better match intended goals, instructions, and human preferences.
An attempt to make a model bypass its intended behavioral or safety restrictions.
Deliberately probing a system for failures, misuse, and security weaknesses.
Unintended exposure of sensitive information, or use of information in model evaluation that should have been withheld.
Information that can identify a person, either on its own or when combined with other data.
How long a service keeps inputs, outputs, logs, and other stored data before deletion or archival.
Rules that determine which users or systems may view data or perform actions.
Collecting traces, metrics, and logs to understand how an AI system behaves in operation.
Running AI well: speed, deployment, usage, and cost.
AI inference that runs on your own device or infrastructure rather than on a remote model service.
AI capabilities delivered through remote infrastructure operated by a service provider.
AI processing close to where data is produced, such as on a camera, phone, or factory sensor.
Selecting which model handles a request based on the task, cost, speed, privacy, or quality requirements.
Reusing previously processed portions of a prompt when a provider or inference system supports it.
Processing a collection of model requests together, often asynchronously.
The time between making a request and receiving a response or completing an operation.
The amount of work a system completes in a given period, such as requests per minute.
A restriction on how many requests or tokens a service allows within a time period.
A limit on the tokens allocated to a request, task, or workflow.
A change in a model's performance as real-world data, conditions, or usage patterns change.
The total cost of completing one useful unit of work, including model calls, tools, retries, and relevant human review.
Download the glossary as a PDF to share internally — or talk to us about turning these concepts into real workflows for your business.