The AI and Agentic Commerce Glossary: Every Term Explained

AI terminology moves fast enough that even people working with these tools daily lose track of which term means what -- especially now that "AI" increasingly means an agent doing multi-step work, not just a chatbot answering questions. This glossary covers the terms that come up most often in our own reporting, grouped by category, defined in plain English.

Core AI & Language Model Terms

Large Language Model (LLM)

A machine learning model trained on large amounts of text to predict and generate language. LLMs are the underlying technology behind ChatGPT, Claude, Gemini, and most other AI chat products.

Foundation Model

A large, general-purpose model trained on broad data that can be adapted to many different tasks, as opposed to a model built and trained for one narrow job.

Token

The basic unit a language model reads and generates text in -- roughly a word or part of a word. Model pricing, context limits, and processing speed are all typically measured in tokens.

Context Window

The maximum amount of text, measured in tokens, a model can consider at one time, including everything in the current conversation, uploaded documents, and its own response. Content beyond that limit gets truncated or dropped.

Parameter

An internal numerical value a model learns during training and uses to produce its outputs. Parameter count is a rough (and increasingly unreliable) proxy for a model's scale and capability.

Training vs. Inference

Training is the process of building a model by having it learn patterns from data, which happens once (or periodically) and is extremely computationally expensive. Inference is the process of actually running the trained model to answer a prompt, which happens every time someone uses it.

Fine-Tuning

Taking an already-trained model and training it further on a smaller, specific dataset to specialize its behavior for a particular task, tone, or domain.

Prompt / Prompt Engineering

A prompt is the input given to an AI model to get a response. Prompt engineering is the practice of deliberately structuring that input -- wording, examples, formatting -- to get more reliable or accurate output.

System Prompt

Instructions given to a model before a conversation starts, usually invisible to the end user, that set its behavior, tone, or constraints for the whole session.

Hallucination

When a model generates information that sounds plausible and confident but is factually wrong or entirely fabricated. It's considered a structural property of how these models generate text, not simply a bug that gets patched out.

Retrieval-Augmented Generation (RAG)

A technique where a model looks up relevant information from an external source -- a document set, database, or live search -- before generating an answer, rather than relying only on what it learned during training. RAG generally reduces hallucination on fact-based questions.

Embedding / Vector Database

An embedding is a numerical representation of text, images, or other data that captures meaning in a form a computer can compare mathematically. A vector database stores and searches those embeddings, which is what makes RAG systems able to find relevant information quickly.

Multimodal

A model that can process and generate more than one type of content -- text, images, audio, video -- rather than being limited to text alone.

Reasoning Model / Test-Time Compute

A model trained and prompted to work through a problem in explicit, multi-step fashion before producing a final answer, rather than generating a response in one pass. That extra step-by-step computation happens at the time you ask the question -- hence "test-time compute" -- and generally improves accuracy on complex, multi-step problems at the cost of speed.

Open-Weight Model

A model whose trained parameters are published for anyone to download and run, as opposed to a closed model only accessible through a company's own product or API.

Mixture of Experts (MoE)

A model architecture that routes each input to a subset of specialized internal sub-networks ("experts") rather than running the entire model on every request, which can make a very large model cheaper to run.

Agentic AI & Protocol Terms

AI Agent / Agentic AI

An AI agent is a system that can take multi-step actions toward a goal with some degree of autonomy -- calling tools, browsing the web, writing and running code -- rather than just responding to a single prompt. "Agentic AI" describes this category of system generally.

Tool Calling / Function Calling

The mechanism that lets an AI model actually do something outside of generating text -- run a calculation, query a database, send an email -- by formally invoking a defined function or tool rather than just describing what it would do.

Orchestration / Multi-Agent System

Orchestration is the coordination logic that decides which tool, model, or sub-agent handles which part of a task. A multi-agent system is an architecture where several specialized agents work on parts of a problem together rather than one agent handling everything.

Model Context Protocol (MCP)

An open standard, originally released by Anthropic and now governed by the Agentic AI Foundation, for how an AI agent connects to external tools, files, and data sources in a consistent way, so developers don't have to build a custom integration for every tool an agent might use.

Agent2Agent Protocol (A2A)

An open protocol, now under the Linux Foundation, that standardizes how independent AI agents -- potentially built by different companies on different platforms -- discover each other's capabilities and communicate. See our full A2A guide for how this applies specifically to agents that shop and pay.

Agent Payments Protocol (AP2)

Google's open protocol for authorizing and auditing payments an AI agent makes on a person's behalf, built around cryptographically signed "mandates" that record what a user actually authorized. See our piece on what mandates do and don't secure.

x402

An open, stablecoin-based payment protocol, built around the long-dormant HTTP 402 "Payment Required" status code, that lets an AI agent (or any automated system) pay for an API call or resource directly over the web without a traditional checkout flow.

Sandbox

An isolated environment where an AI agent's actions are contained and can't affect real systems or data, used to test or run agents whose behavior isn't fully trusted yet.

Guardrails

Rules, filters, or checks -- whether built into a model's training or added as an external layer -- meant to keep an AI system's outputs or actions within acceptable bounds.

Mandate (Intent Mandate / Cart Mandate)

In AP2, a cryptographically signed record used to prove what a user authorized an agent to do. See the full explanation in our AI agent security article.

Know Your Agent (KYA)

A framework, led by card networks including Mastercard, for verifying an AI agent's identity and authorization before allowing it to complete a payment -- modeled conceptually on "Know Your Customer" rules in banking. See our coverage of Mastercard's rollout.

AI Safety & Trust Terms

Prompt Injection

An attack where malicious instructions are hidden inside content an AI agent processes -- a webpage, a document, an email -- so the agent follows the attacker's instructions instead of its actual user's. Considered one of the most serious unsolved problems in agentic AI security.

Jailbreak

A prompt or technique specifically designed to get a model to ignore its safety training and produce output it would normally refuse to generate.

Lethal Trifecta

A framework describing the specific combination that makes an AI agent dangerous to prompt injection: access to private data, exposure to untrusted content, and the ability to communicate externally. An agent with all three can potentially be turned into a data-exfiltration tool.

Alignment

The broad effort to make an AI system's behavior actually match human intentions and values, rather than technically satisfying its training objective in an unintended or harmful way.

Red Teaming

Deliberately trying to find an AI system's flaws, vulnerabilities, or failure modes -- including trying to make it misbehave -- before real users or attackers do.

Human-in-the-Loop

A design where a person must review or approve an AI agent's action before it takes effect, as opposed to letting the agent act fully autonomously.

AI Search & GEO Terms

Generative Engine Optimization (GEO)

The practice of structuring web content so AI systems -- chatbots, AI-generated search summaries, and autonomous agents -- can find, understand, and accurately cite it, as a complement to traditional search-engine optimization.

llms.txt

A proposed standard, analogous to robots.txt, for a plain-text file at a website's root that gives AI crawlers and agents a clean, structured summary of a site's content, meant to make that content easier to parse than crawling full HTML pages.

Structured Data / Schema Markup

Machine-readable code embedded in a webpage (commonly in JSON-LD format, using the schema.org vocabulary) that explicitly labels what a piece of content is -- an article, a product, a FAQ -- so search engines and AI systems don't have to infer it from the page's visible text alone.

AI Crawler / Bot

An automated program that visits websites specifically to gather content for training or powering an AI system, such as GPTBot (OpenAI) or ClaudeBot (Anthropic), distinct from the crawlers traditional search engines use.