Chatbot and AI agent glossary: 19 terms for choosing without confusion

Market labels change faster than capabilities. The same solution may be called an AI chatbot, virtual assistant, conversational agent or AI agent, while two products using the same label can do entirely different jobs.

This glossary does not decide which name a vendor is allowed to use. It translates every term into a question you can test in a demonstration. For a direct comparison, the guide to chatbots versus AI agents applies the same test to a complete process.

Chatbot

A chatbot is a software interface that exchanges messages with a person. It may work with buttons, rules, search, a language model or a combination. It may also have integrations; the word chatbot alone does not reveal whether it queries systems or takes actions.

How to test it: ask where each answer comes from and what happens when the required fact is not in its configured content.

AI chatbot

An AI chatbot uses one or more models to interpret or produce messages. That may make a conversation more flexible, but it does not prove that the system knows a current fact or can act on it.

How to test it: request an answer that changes, such as availability or order status, and watch whether it queries the real source or writes from static context.

Conversational AI

Conversational AI is a broad category for systems that interact through natural language, by text or voice. It describes the conversational experience, not the operational scope behind it.

How to test it: separate dialogue quality from completed work. A fluent answer and a correct CRM update are two different tests.

Large language model or LLM

A large language model (LLM) receives context and produces an output, usually text or a structured decision. By itself it does not know your inventory status, have permission to change your CRM or confirm that a payment occurred. It needs context and tools to work with those systems.

How to test it: ask which information enters the model, which information does not and which component carries out the next action.

AI agent

An AI agent receives a goal, chooses one or more steps and uses tools within defined permissions. It may hold a conversation, but it does not have to: an operational agent may run a scheduled task or react to an event without showing a chat.

How to test it: request the tool list, permissions, error handling and execution record.

Met

A Met is a Meteor AI agent. Conversational Mets serve, qualify, quote or schedule on WhatsApp, Instagram, Messenger and web chat. Operational Mets work across connected systems even when the task does not involve a conversation.

Met is the product name, not a general industry category. Its scope depends on the tools, credentials and limits configured for each implementation.

Agent tool

A tool is a specific function an agent can call to query or change something: search for a contact, check stock, create an appointment or prepare a payment link. It needs a name, defined inputs and a result the system can interpret.

How to test it: do not accept “it integrates with the CRM” as enough detail. Request the available actions and ask which are read-only or write actions.

Integration or connector

An integration connects the platform with another system through its API, events or another supported mechanism. Having a connector does not mean having access to the whole external product: scope depends on the credential and implemented operations.

How to test it: review the connector record and request both the successful path and the error path in a demonstration.

Model Context Protocol (MCP)

Model Context Protocol (MCP) is an open standard for connecting AI applications with context and capabilities. Its specification distinguishes resources, prompts and tools; a tool is a function the model can invoke to obtain information or take an action.

MCP standardises the connection, but it does not make every tool safe by itself. Permissions, consent, confirmations and logs remain implementation decisions.

RAG

Retrieval-Augmented Generation (RAG) searches a source for information and adds it to the context before producing an answer. It is useful for policies, manuals and knowledge bases. It does not replace a transactional query when the fact changes in real time.

How to test it: ask which source is searched, how it is updated, whether the answer identifies its source and what happens when two documents disagree.

Knowledge base

A knowledge base holds content the system can retrieve: frequently asked questions, documentation, policies or procedures. It is not the same as the ERP, store or CRM where changing operational states live.

How to test it: identify which answers come from documents and which must be queried in a transactional system.

AI second brain

An AI second brain is a system or method for preserving notes, meetings, decisions and other personal material, then retrieving it with AI. The label describes a way to organise and query external memory; by itself it does not demonstrate enterprise permissions, traceability or governance.

How to test it: ask what is preserved, who may view, correct or delete each record and whether an answer links back to its original source.

Company Brain

A Company Brain is a company’s shared operational context: it connects structured knowledge, contacts, conversations, roles, processes and tools. In Meteor, collections organise company knowledge, MCP queries live systems and Mets use and update context within their configured scope. The internal AI chatbot and Company Brain page explains how this operational architecture works.

How to test it: ask the same question under two different roles and inspect which source it uses, which part of the answer changes and how it blocks information that role cannot access.

Context, memory and history

Context is the information available during one execution. Memory is information the system preserves for later use. History is the record of what happened. They may overlap, but they are not interchangeable: saving a chat does not guarantee that the next commercial step was assigned.

How to test it: start a conversation, close it, return after navigating elsewhere and check what is preserved, for how long and where it can be corrected.

Conversational CRM

A conversational CRM connects the conversation with contacts, owners, opportunities, tasks and next steps. Its value is not displaying messages; it is turning them into a record the team can continue and measure.

How to test it: request the contact, owner, stage, note and task created from a real conversation.

Workflow and automation

A workflow defines steps, conditions and outcomes. An automation executes those steps when an event occurs or a rule matches. It may use AI for one part without making the entire journey autonomous.

How to test it: draw the trigger, each system queried, decisions, approval and final outcome. Anything missing from that journey has not yet been demonstrated as automated.

Human handoff or escalation

A human handoff transfers a conversation or task to a person when information is missing, an exception appears, the customer asks for help or a decision requires judgement. A button is not enough: the person needs the context and a clear pending action.

How to test it: cause a case the system cannot resolve and inspect who receives the task, with which history and expected response time.

Human approval

Human approval is a control point before a consequential action such as publishing, spending a budget or issuing a document. It is not the same as a handoff: the agent may prepare the work and wait for authorisation before continuing.

How to test it: ask which actions stop, who can approve them and whether the authorisation is recorded.

Hallucination

A hallucination is a plausible output that is not supported by the available data. RAG, tools and rules can reduce particular errors, but no category label eliminates the risk. A responsible design limits what the system can do when it lacks enough evidence.

How to test it: use a nonexistent fact, an ambiguous request and an external-system failure. The correct response may be to stop, request information or escalate.

The test that joins the whole glossary

To assess a chatbot or agent, take one real case and complete this sentence:

When this message or event arrives, the system queries this source, takes this action with this permission, records this outcome and gives this person what it cannot resolve.

An empty space does not mean the solution is bad. It means that part of the scope has not yet been demonstrated. That is a more useful basis for deciding than any category name.