The AI Glossary

Artificial Intelligence, AI

Software that can perform tasks that usually require human thinking, such as writing, summarizing, analyzing, classifying, planning, or answering questions.

Generative AI

AI that creates new content, such as text, images, audio, code, summaries, emails, reports, or plans.

LLM, Large Language Model

An AI model trained to understand and generate language. Examples include the models used inside tools like ChatGPT, Claude, and Gemini.

Model

The AI system that processes input and generates output. Different models have different strengths, costs, speeds, and limitations.

Model Parameters

The internal values a model learns during training. The number of parameters is often used as a rough measure of a model’s size and capability (for example, a “70B” model).

Foundation Model

A large general-purpose AI model that can be adapted or used for many different types of tasks.

Reasoning Model

A model designed to spend more effort working through complex problems, planning, logic, math, coding, or multi-step tasks.

Chain-of-Thought

When a model works through a problem step by step before giving its final answer, which often improves accuracy on complex or multi-step questions.

Model Communication Style

The tone, length, and formatting a model tends to use in its responses. Different models “sound” different, and the style can be guided with prompts and system instructions.

Multimodal AI

AI that can work with more than one type of input or output, such as text, images, audio, video, or files.

Training Data

The large collection of examples a model learns from. The quality and range of this data shape what a model is good — and not good — at.

Fine-Tuning

Further training a model on your own examples so it performs better on a specific task, tone, or domain. Different from giving a model information through prompts or retrieval.

Distillation

Creating a smaller, faster, cheaper model that mimics a larger one — keeping much of the quality while lowering cost and speed requirements.

Quantization

Compressing a model so it uses less memory and runs faster and cheaper, usually with only a small trade-off in quality.

Inference

The moment a model actually runs to produce an answer. Most day-to-day AI cost comes from inference, not from training.

Temperature

A setting that controls how predictable or creative a model’s responses are. Lower is more focused and consistent; higher is more varied.

Open-Source Model

A model whose code or weights are openly available, so it can be run, customized, or self-hosted instead of only accessed through a provider’s service. Examples include Llama and Mistral.

Knowledge Cutoff

The date a model’s training data ends. Anything that happened after it isn’t “known” to the model unless you provide it through tools or retrieval.

Prompt

The instruction, question, or request given to an AI model.

Prompt Engineering

The practice of designing better instructions so AI gives more useful and consistent results.

System Prompt

A higher-level instruction that tells the AI how it should behave, what role it should play, and what rules it should follow.

Meta Prompt

A prompt that helps generate, improve, or control other prompts.

Few-Shot / Zero-Shot Prompting

Giving a model a few examples of what you want (few-shot) versus none at all (zero-shot). A few good examples often improve accuracy and consistency.

Context

The information the AI has available when answering, such as the user’s question, previous messages, documents, examples, or instructions.

Context Window

The amount of information a model can consider at one time.

Token

A small piece of text processed by an AI model. Tokens are used to measure input, output, cost, and usage.

Tokenization

The process of breaking text into tokens — the small pieces a model actually reads and counts. How text is tokenized affects both cost and how much fits in the context window.

Output

The response or result generated by the AI.

Hallucination

When AI produces something that sounds confident but is wrong, unsupported, or made up.

Grounding

Giving AI reliable source material to base its answer on.

Bias

When an AI system produces unfair, incomplete, or skewed outputs because of data, design, or user instructions.

ChatGPT

OpenAI’s AI assistant product.

OpenAI

The company behind ChatGPT and many GPT models.

Claude

Anthropic’s AI assistant and model family.

Anthropic

The company behind Claude.

Gemini

Google’s AI model and assistant ecosystem.

GPT

A type of language model associated with OpenAI. In everyday usage, people may also say “a GPT” to mean a custom ChatGPT assistant.

Custom GPT

A customized assistant inside ChatGPT with specific instructions, knowledge, and capabilities.

Gem

Google’s version of a customized Gemini assistant.

Skill

A reusable capability, instruction set, or workflow component that helps an AI system perform a specific task consistently.

Agent

An AI system that can work toward a goal, use tools, make decisions, and take actions across multiple steps.

Agentic

Describes AI that behaves more like an active worker by planning, using tools, taking actions, checking results, and continuing across steps.

Agentic Development

Building AI systems that can perform multi-step work with less human input.

Multi-Agent System

Several AI agents working together, each handling part of a job and coordinated to complete more complex work than one agent could alone.

Memory

An agent’s ability to retain useful information across steps or sessions — for example, remembering a client’s preferences or an earlier instruction.

Subagent

A smaller, specialized agent that another agent creates or delegates to for part of a job — useful for breaking big tasks into focused pieces.

Agent Loop

The cycle an agent repeats to get work done: plan a step, take an action or use a tool, check the result, then decide what to do next.

Computer Use

An AI agent operating a computer or browser the way a person would — clicking, typing, and reading screens — to handle tasks in tools that don’t offer a direct connection.

Prompt Engineer Agent

An agent designed to create, improve, test, or manage prompts.

Onboarding Agent

An agent that helps manage or automate parts of the client onboarding process.

Requirement Agent

An agent that helps collect, clarify, or structure client requirements.

Chatbot

An AI or rules-based tool that interacts through conversation. A chatbot usually answers or responds, while an agent may take action.

Automation

A system that completes a task or workflow with little or no manual effort.

Trigger

The event that starts an automation — for example a new email, a form submission, a schedule, or a button press.

Cron, Scheduled Job

An automation set to run on a fixed schedule — for example every morning, every hour, or on the first of the month.

Workflow

A sequence of steps used to complete a business process.

AI Workflow

A workflow where AI performs one or more steps, such as summarizing, drafting, deciding, routing, or checking.

Orchestration

Coordinating multiple agents, tools, and steps so they work together in the right order to complete a larger task. It’s the backbone of multi-agent systems.

Pipeline

A series of connected steps where the output of one becomes the input of the next, moving work from start to finish automatically.

Queue

A line of tasks waiting their turn to be processed, so work is handled in order and nothing gets dropped when things get busy.

Idempotency

A safety property meaning that running the same step twice has the same result as running it once — so retries or duplicates don’t cause double charges, duplicate records, or other mistakes.

Human-in-the-Loop

A setup where AI does part of the work, but a human reviews, approves, or handles exceptions.

Guardrails

Rules, limits, or checks that guide what AI is allowed to do.

QA Loop, Quality Assurance Loop

A process for checking whether AI output meets the required standard.

Eval

A structured test used to measure how well a model, prompt, or workflow performs against a set standard. Evals help teams compare options and catch quality problems before clients do.

Learning Loop

A system where AI output is tested, measured, improved, and tested again.

Auto Research

A workflow where AI generates ideas, tests them, measures results, and improves based on feedback.

Iteration

Repeating a process with improvements each time.

Self-Improvement

When an AI system or workflow uses results, feedback, or scoring to improve future performance. This does not mean the AI is becoming conscious. It means the workflow is designed to improve based on measured feedback.

Deployment

Putting a tool, agent, or automation into active use.

Agent Deployment

Launching an agent so it can perform its intended work in a real environment.

Sandbox

A safe, isolated environment for testing AI tools, prompts, or automations without affecting real systems, data, or clients.

Dry Run

Testing an automation without letting it make real changes, so you can confirm it behaves correctly before going live.

Staging

A testing or review environment before something goes live.

Production

The live environment where real users or clients use the system.

Rollout

The process of introducing a new tool, agent, or workflow to users.

Canary Release, Canary Deployment

Rolling out a change to a small group first to see how it performs before releasing it to everyone. If something goes wrong, only a few are affected — like an early warning.

Rollback

Reverting to a previous, known-good version when a new change causes problems — the safety net behind any release.

Blue-Green Deployment

Running two copies of a system — the current one and the new one — and switching over only once the new version is confirmed healthy, so there’s no downtime.

Adoption

Whether people actually use the new tool or workflow successfully.

Change Management

Helping people understand, accept, and use a new way of working.

API

A way for software systems to talk to each other. APIs allow AI tools to connect with platforms like CRMs, Google Workspace, websites, databases, or other business systems.

API Key

A secure credential that lets one system access another system’s API. It should be treated like sensitive access information.

Endpoint

A specific address (URL) on an API that performs one action, such as “create a contact” or “send a message.”

SDK, Software Development Kit

A ready-made toolkit that makes it easier and faster to build with a particular service or API.

Authentication, OAuth

How one system proves who it is to securely connect to another. OAuth is a common standard that lets tools connect without sharing your password.

Console

A provider dashboard where technical users manage API keys, billing, usage, models, and settings.

Cloud Console

A dashboard for managing cloud services, such as Google Cloud resources, APIs, billing, and usage.

CLI, Command Line Interface

A text-based way for technical users to interact with software.

Repo, Repository

A place where code, prompts, files, and system assets are stored, often in GitHub.

GitHub

A platform commonly used to store, version, and collaborate on code or technical assets.

API Limit

A restriction on how much a tool or model can be used within a certain time period.

Rate Limit

A limit on how many requests can be sent to a system in a certain time window.

Session Limit

A limit on usage within a user session or time period.

Usage Limit

A general cap on how much of a tool or service can be used.

Credits

Prepaid or allocated usage value for a tool or model.

Token Usage

The number of tokens used by prompts and outputs. Token usage can affect cost and capacity.

Token Optimization

Designing prompts, inputs, and workflows to use fewer tokens — lowering cost and helping work fit within a model’s context window, while keeping quality.

Load Balancing

Distributing work across multiple accounts, systems, or resources to avoid overload.

Concurrent Connections

Multiple tasks, users, or agents running at the same time.

Throttling

Slowing down requests because a system is hitting usage or rate limits.

Latency

Delay between making a request and getting a response.

Streaming

Delivering a model’s response piece by piece as it’s generated, so text appears to “type out” in real time instead of arriving all at once.

Cache, Caching

Temporarily storing results so they can be reused instantly instead of being recomputed. In AI, caching repeated context (often called prompt caching) can noticeably lower cost and speed up responses.

Prompt Caching

Reusing parts of a prompt the model has already processed (such as a long set of instructions or documents) so repeated requests cost less and respond faster.

Logging, Observability

Keeping records of what a system did (logs) and being able to see how it’s performing (observability) — essential for troubleshooting and trust.

Cost per Run

The cost each time an AI task or workflow runs.

Cost Optimization

Designing AI workflows to control cost while maintaining quality.

Compute

The processing power needed to train or run AI systems.

Self-Hosted, On-Premise

Running software or AI on your own server or infrastructure rather than only through a provider’s cloud. Often chosen for privacy, control, and data ownership.

Model Tier

Different versions of models with different cost, speed, and capability levels. Examples may include mini, nano, pro, or advanced versions depending on the provider.

Model Cost

How much it costs to use a particular model. Cost is usually driven by token usage and the model’s price per token, so more capable models often cost more per use.

Knowledge Base

A collection of information the AI can reference, such as documents, FAQs, transcripts, client materials, policies, or internal resources.

Digital Twin

A living digital model of a business, process, or system — built from its real data and knowledge — that AI can use to answer questions, test ideas, and take action. Sometimes called a “company brain.”

Retrieval

The process of finding relevant information from a knowledge source before answering.

RAG, Retrieval-Augmented Generation

A method where AI retrieves relevant information from external sources and uses it to generate a better answer.

Chunking

Splitting long documents into smaller pieces so AI can search them and pull just the relevant parts. How content is chunked has a big effect on retrieval quality.

Reranking

Taking the results a search returns and reordering them by how relevant they really are, so the AI works from the best material first.

Indexing

Organizing data ahead of time so it can be searched and retrieved quickly — much like the index at the back of a book.

Semantic Search

Search based on meaning, not just exact keywords.

Embedding

A numerical representation of text, images, or other content that helps AI compare meaning.

Vector

A set of numbers representing meaning or features.

Vector Database, Vector Store

A database designed to store embeddings and retrieve similar information quickly.

Knowledge Graph

A way of storing information as connected facts and relationships — people, places, things, and how they relate — helping AI reason about how things fit together.

Data Source

Any system or file the AI may use, such as a CRM, Google Drive, website, spreadsheet, database, or client document.

Data Ingestion

Bringing data into a system so it can be searched, processed, or used by AI.

Metadata

Information about other information — for example a file’s author, date, type, or tags — used to organize, filter, and find content.

Data Cleaning

Fixing messy, duplicate, outdated, or inconsistent information.

Source of Truth

The official, trusted place where correct information lives.

Permissions

Rules that determine who or what can access specific information.

Privacy

Protecting sensitive personal or business information.

PII, Personally Identifiable Information

Information that can identify a specific person, such as a name, email, phone number, or address. It must be handled carefully for privacy and compliance.

Security

Protecting systems, data, and access from misuse or unauthorized access.

Prompt Injection

A security risk where hidden or malicious instructions are slipped into content an AI reads, trying to make it ignore its rules or reveal information. A key reason guardrails and human review matter.

Encryption

Scrambling data so only authorized people or systems can read it, protecting it both while stored and while being sent.

RBAC, Role-Based Access Control

Granting access based on a person’s role, so each user only sees and does what their job requires — a common way to keep data secure.

Compliance

Meeting legal, regulatory, contractual, or internal policy requirements.

Integration

Connecting two or more tools so they can share data or trigger actions.

Connector

A prebuilt connection between systems.

Translation Layer

A part of a system that converts information from one format or language into another, so two tools, models, or systems can understand each other.

Middleware

Software that sits between two systems and helps them work together — passing, translating, or coordinating information as it moves between them.

Tool Calling

When AI uses an external tool to complete a task, such as searching, sending a message, creating a task, or reading a file.

Function Calling

A technical method that lets AI call specific software functions in a controlled way.

Scaffolding

The supporting structure built around an AI model — instructions, tools, steps, memory, and checks — that turns a general model into a reliable system or agent for a specific job.

MCP, Model Context Protocol

A standard that helps AI tools connect to external systems and context sources.

Webhook

A trigger that sends information from one system to another when something happens.

REST API

The most common style of web API, where systems exchange information through simple, standard web requests. Most tools you’ll integrate with offer one.

Zapier

A no-code automation platform that connects apps and automates workflows.

Make

A no-code or low-code automation platform for connecting apps and building workflows.

n8n

An automation platform often used for more flexible or technical workflows.

CRM

Customer Relationship Management software used to manage contacts, deals, clients, and sales activity.

PRD, Product Requirements Document

A document that explains what needs to be built and why.

Tech Scope

A technical plan that outlines how a system or workflow will be built.

XML

A text format that uses tags to label and structure information, for example <task>…</task>. It is sometimes used to organize prompts or data so AI can read them more reliably.

JSON

A common text format that organizes data as labeled fields (for example, name, email, date). Widely used to pass structured information between systems and to get clean, structured output from AI.

Structured Output

When AI returns its answer in a fixed, organized format such as JSON, so other systems can read and use it automatically instead of parsing free text.

AI Readiness

How prepared a team or organization is to adopt AI successfully.

AI Literacy

A person or team’s ability to understand and use AI responsibly.

AI Jargon

The specialized or technical vocabulary used when talking about AI. Jargon can make AI feel more complicated than it is — translating it into plain language is exactly what keeps clients and teams confident.

AI Governance

Rules, roles, and processes for how AI is used safely and effectively.

Responsible AI

Using AI in a way that is safe, fair, transparent, and aligned with human values.

Jailbreak

An attempt to trick a model into ignoring its safety rules or restrictions. Closely related to prompt injection, and a reason guardrails and testing matter.

Red Teaming

Deliberately stress-testing an AI system to find weaknesses, risks, or harmful outputs before real users — or bad actors — do.

AI Policy

Internal rules for how employees can and cannot use AI.

Data Residency

Where your data is physically stored and processed — for example, kept within a specific country. It often matters for privacy laws and client requirements.

Audit Log, Audit Trail

A time-stamped record of who did what and when, so actions can be reviewed for accountability, security, and compliance.

Shadow AI

Employees using AI tools without company approval or oversight.

Singularity

A theoretical point where AI becomes more capable than humans in a broad, transformative way. This term is more future-focused and is not usually needed in everyday client delivery conversations.