What is an AI agent in plain language?

An AI agent is a system that performs work within set boundaries: it runs a process from request to result, picks tools and the next step, escalates exceptions to a human, and leaves a trail. It is not the same as a chatbot, which only answers questions. It is also not a digital employee or a party responsible for business decisions.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions from a knowledge base: FAQs, statuses, opening hours. An agent performs work: it runs a case within boundaries, escalates exceptions, and leaves a trail. If customers mostly ask simple FAQs, a chatbot is enough at a fraction of the price of an agent that runs a process, which starts from €6,000 net.

What are MCP and function calling?

Function calling is a mechanism where the model, instead of writing text, returns a structured tool call (a name and arguments) that the application executes. MCP (Model Context Protocol) is an open standard describing how a model connects to tools and data so you do not build every integration from scratch. Neither is the agent itself; both are ways of connecting tools.

← Back to blog

GlossaryAI agent glossary 2026

AI Agent Glossary: 25 Terms Sellers Conflate, in Plain Language (2026)

An AI agent, in plain language, is a system that performs work within boundaries, not one that only answers questions like a chatbot. A glossary of 25 terms sellers conflate: agent vs chatbot vs copilot, agent washing, MCP, RAG, human-in-the-loop, tokens. One difference sets the budget: an agent that runs a process from €6,000 net, a chatbot a fraction of that.

SyntalithPublished July 6, 2026Updated July 6, 202611 min read

An AI agent, in plain language, is a system that performs work in a company within set boundaries, not one that only answers questions like a chatbot. This glossary explains 25 terms sellers most often conflate. One difference sets the budget: a chatbot costs a fraction of an agent that runs a process (from €6,000 net).

Quick answer: agent, chatbot, copilot, assistant - the one difference that sets the budget

The price does not depend on the label, but on how much work the system actually performs. These seven terms are the ones most often conflated in offers, so we put them side by side. The "what it means" column sets the budget, not the label.

Term	What it means	What it is NOT
Chatbot	Answers questions from instructions and a knowledge base: FAQs, statuses, opening hours.	Not something that performs work in your systems or runs a case to the end.
Copilot	Suggests to a person and drafts versions; speeds up one person.	Does not act on its own; leaves the decision and the send to the human.
Assistant (operator)	Works beside you: tidies documents, inbox, repeatable pieces of work under supervision.	Does not run a whole business process without your approval.
Automation	Runs a fixed sequence of steps by rules: "if A, then B, then C".	Does not interpret ambiguous data or pick the next step on its own.
RPA	Clicks through interfaces like a human: rekeys data between apps that lack an API.	Does not understand content; usually breaks when a screen changes.
AI agent	Runs a process within boundaries, picks tools and the next step, escalates exceptions, leaves a trail.	Not a digital employee, nor a party responsible for decisions.
Agentic AI	A build pattern: the model helps choose the next step and calls tools within a task.	Not the technology itself, nor a guarantee of autonomy; it is a way of designing.

The most expensive mistake is paying for a level of autonomy the process does not need. If a copilot with a human at the end would do, and someone sells you an "agent", you are paying for a word, not a system.

Agent washing: why "agent" in an offer often means chatbot

Agent washing is selling a chatbot or a simple automation under the name "AI agent" because the word sounds more expensive than the product. You spot it by one thing: the offer says what the system "can do", but not what exactly it will do on its own and how you will know it did.

In June 2025 Gartner estimated that of thousands of vendors only about 130 offer genuinely agentic systems, and predicted that over 40% of agentic AI projects would be cancelled by the end of 2027, mainly due to rising costs and unclear value. Do not pay agent prices if a chatbot will handle the process.

One question exposes agent washing before you sign: "what exactly will this system do on its own, and how will I know it did it?". How to tell a real agent from a repainted chatbot step by step, we explain in the guide on what an AI agent is, which lays out the seven-criteria detector.

How an agent works under the hood: technical terms without jargon

These words show up in offers as proof of "sophistication". On their own they are not an answer; what matters is how they are constrained.

Language model (LLM). A program trained on text that predicts the next words and so generates answers. What it is not: it has no access to your systems and no knowledge of your company until you connect them, and it performs no action on its own.
Context window (context). The amount of text the model considers at once: the question, instructions, case history, attached documents. What it is not: it is not persistent memory; once the limit is exceeded, older information drops out unless the application supplies it again.
Function calling (tool calling). A mechanism where the model, instead of writing text, returns a structured call: a tool name and arguments. What it is not: the model executes nothing itself; the application runs the function and can validate, limit, or reject it.
Tools. Actions the application exposes to the model: read from CRM, create a ticket, draft a reply, send after approval. What they are not: they are not the model's "skills"; they are the company's code with a description, input validation, and a log.
MCP (Model Context Protocol). An open standard (proposed by Anthropic) describing how a model connects to tools and data sources, so you do not build every integration from scratch. What it is not: it is not an agent or "intelligence"; it is a way to plug in, like a connector.
RAG (retrieval-augmented generation). A technique where the system first retrieves fragments of documents and only then generates an answer grounded in sources. What it is not: it is not a guarantee of correctness or a grant of permissions; the agent may get the wrong fragment, and finding a refund procedure is not permission to run it.
Orchestration. The layer that governs the order of steps: what the model does first, when to call a tool, when to stop for a human. What it is not: it is not the model itself; it is the process logic around the model, and it usually decides safety.
Multi-agent system. An arrangement of several specialized agents that split a task and hand work to each other. What it is not: it is not better by default; it makes it harder to trace which component made which decision, so it is rarely a good first step.

The terms that decide whether an agent can be trusted

These are not decorations. Without these mechanisms you have a demo, not a production system.

Boundaries. Rules the agent cannot cross, even if the user asks: amount limits, locked fields, a list of allowed tools. What they are not: they are not a suggestion in the prompt text; they are hard limits enforced by the application.
Human-in-the-loop. A pattern where the agent stops before a sensitive action and waits for a human decision: sending a disputed reply, refunding money, changing a contract. What it is not: it is not "someone glances at the logs sometimes"; the approval point is built into the process.
Escalation. The moment the agent hands a case to a human because confidence is low or the case falls outside the boundaries. What it is not: it is not a system failure; a narrow, accurate escalation stream is a sign the agent works correctly rather than guessing.
Audit trail. A log of what the agent saw, which tool it called, what it changed, who approved, and how much the query cost. What it is not: it is not an optional add-on; if a week later you cannot reconstruct what the system did and why, it is not an agent, just roulette with a nice interface.
Observation mode (and draft mode). A production start where the agent classifies and prepares drafts, but nothing goes out without approval. What it is not: it is not a demo build; it is the order that turns a demo into a system before auto-send is switched on at all.
Autonomy. The extent to which the agent acts without a human: from "only drafts" to "acts within boundaries and escalates exceptions". What it is not: it is not a scale where more means better; the most expensive mistake is paying for a level of autonomy the process does not need.

Risks and costs: what you must understand before you buy

This is where good offers speak plainly and agent washing stays quiet.

Hallucination. An answer that sounds confident but is inconsistent with the facts or sources, because the model generates plausible text, not verified truth. What it is not: it is not a rare exception; that is why a production agent cites sources and, in their absence, says "I don't know" or escalates.
Prompt injection. An attack where malicious content hidden in an email, a document, or a web page tries to hijack the agent and make it do something outside intent: leak data, send money. What it is not: it is not theory; it is the primary risk for agents with access to tools. More in the piece on prompt injection in AI agents.
Tokens. The units models split text into and by which providers bill cost: you count input and output. What they are not: they are not a subscription; they are a variable cost that at typical volumes runs to cents per case, but has to be calculated on the company's real traffic:

Monthly model cost =
  number of cases per month
  x average tokens per case
  x the provider's token price

You check current token rates directly with the model provider, because they change faster than a service price list. The quote must name the daily limit and how the system behaves once it is exceeded.

When you do NOT need an AI agent

Knowing these terms has one practical purpose: not overpaying for a word. There are situations where an agent is a bad purchase, however fashionable it is.

The process is mostly simple FAQs and statuses. A chatbot at a fraction of the price is enough, and an agent is overkill.
The rules change every week and live in someone's head. Write the process down on paper first. That is 80% of the work before AI even enters the picture.
The process happens rarely. The cost of building and maintaining it will not pay back even in an optimistic scenario. Manual handling can be cheaper.
A ready SaaS or a simple automation without a model is enough. Do not build a dedicated agent where an off-the-shelf product already works.

If any of these fits your situation, we will say so plainly at the scan, before you spend anything.

How to start

Take one process and name what the system should do on its own and what it should only prepare for approval.
Test the offer with one question: "what exactly will this system do on its own, and how will I know it did it?". A direct answer, or the lack of one, tells you more than the whole pitch.
Book a free process scan: 30 minutes with an engineer and a written takeaway in two business days, with a recommendation on whether you need an agent, an automation, or nothing yet.

Book a free process scan | What an AI agent is: seven criteria | See pricing