What is an AI agent? Definition, examples, and how it differs from a chatbot

The short answer

An AI agent is software that you give a goal, and that decides on its own, step by step, which actions to take to reach it: look up information, call a tool, write a reply, hand off to a human. It relies on a language model (the kind of model behind chat assistants) to reason, and on connected tools to act. That choice of actions, left to the model, is what separates it from a scripted chatbot and from a classic automation.

n8n’s documentation puts it this way: a chain follows a sequence fixed in advance, while an agent uses a language model to work out which actions to take (n8n documentation, “What agents do” (accessed November 11, 2026)).

One example: “Where’s my order?”

An online store gets messages every day asking “where’s my order?” Software can answer them in three ways.

• A scripted chatbot. It follows a tree written in advance: “Enter your order number,” then a link to the tracking page. If the customer writes something the script did not plan for, there is no branch for it.
• An automation. A fixed series of steps: for every message containing the word “order,” send the canned reply. The same steps, in the same sequence, whatever the message says.
• An AI agent. It gets instructions (“answer tracking questions; when in doubt, pass the message to a person”) and tools: read the order in the store, check the carrier’s tracking, send a message, open a ticket. On each message, the model chooses which tools to call and in what sequence. For a late package it can check tracking and reply. For a refund request it can open a ticket.

The agent can also get it wrong: misread the request, call the wrong tool, state something false. That is not a footnote. It is the price of the choice you leave to it.

What an agent is made of

• A language model, which reads the request and decides what comes next.
• Instructions: its role, its limits, the tone to use.
• Tools: the actions it is allowed to take. An agent with no tools only writes. The MCP protocol is a standard way to plug those tools in.
• Knowledge: the documents it searches for answers (FAQ, return policy, site pages).
• A trigger or a channel: a message on the website, an email, a schedule.
• Guardrails: what it cannot do alone, and when it hands off.

Agent vs chatbot vs automation

	Scripted chatbot	Automation	AI agent
Who chooses the steps	the script’s author	the workflow’s author	the model, within set limits
What it receives	expected answers	structured data	free-form language
Same input, same result	yes	yes	not guaranteed
Cost of one request	stable	stable	variable: depends on the tools called and the model

An automation is still the right choice when the series of steps never changes: it is predictable and bills in a stable way. The guide to no-code automation starts there.

AI agent examples among the tools covered here

• Customer support agent. Botpress is a platform for customer support AI agents with a visual builder. Tidio adds an AI agent, “Lyro,” to a live chat and ticketing tool for websites and online stores.
• Work assistant. Lindy is an autonomous assistant tied to the team’s tools: Slack, email, calendar, meetings.
• Custom agents. Relevance AI is for building agents and agent teams (“Workforces”) connected to apps.
• Agent inside a workflow. n8n has an AI Agent node that sits as one step of an automation.

Table scrolls horizontally →

What each platform bills for an agent

Criterion	BotpressAI agent	TidioAI agent	LindyAI agent	Relevance AIAI agent	n8nAutomation tool
Billing unit	conversationssource	billable conversationssource	creditssource	actionssource	executionssource
Free plan	Yessource	50 billable conversations per monthsource	Nosource	Nosource	Nosource
AI features	AI customer support agents; "AI usage" (model costs) partly included, then $10 credits sold at costsource	"Lyro" AI agent billed per conversation, through a separate monthly quota (50 to 1,000); "Flows" automationssource	Autonomous assistant with scheduled routines and over 40 "skills"; everything is billed in credits from the shared poolsource	Agents and agent teams ("Workforces"); models charged in "Vendor Credits" with no markup, or your own API keyssource	AI Agent node included (Cloud and Community); models paid with your own keys or Gateway creditssource
Vendor	Visit Botpress (opens in a new tab)	Visit Tidio (opens in a new tab)	Visit Lindy (opens in a new tab)	Visit Relevance AI (opens in a new tab)	Visit n8n (opens in a new tab)

What this changes when you pick a tool

An agent does not take the same number of actions from one request to the next. The bill depends on what the vendor counts. Back to the store: 200 requests a month, and for each one the agent calls three tools (read the order, check tracking, reply).

• Botpress counts conversations: any exchange with at least two messages from the end user in the month. The 200 requests make 200 conversations at most, however many tools are called. The Plus plan includes 250 a month. On top of that sits an “AI usage” meter, the cost of model calls, partly included.
• Tidio counts Lyro conversations separately: any interaction with at least one reply from the AI agent. The 200 requests make 200 Lyro conversations, on a separate monthly quota that runs from 50 to 1,000.
• Relevance AI counts actions: one each time an agent runs a tool, even if the run fails. 200 requests × 3 tools make 600 actions, against 2,500 actions a month on the Pro plan. Models are billed separately, in “Vendor Credits.”
• Lindy counts credits by how heavy the task is: 2 to 250 credits for an everyday request, according to the vendor. The 200 requests therefore use between 400 and 50,000 credits, against a seat that starts at 5,000 credits a month.
• n8n counts executions: one complete run of the workflow, however many steps it has. If each message starts the workflow once, that is 200 executions, out of 2,500 on the Starter plan. Models are paid separately, with your own keys or Gateway credits.

The same work comes out as 200 conversations, 600 actions, 200 executions, or a range of 400 to 50,000 credits. You cannot compare those numbers by eye. The guide to billing units explains how to bring them back to one scenario, and the cost calculator does the math for the automation part, AI steps excluded.

A few pointers:

• If the agent has to answer your customers on your website or a messaging app, look at the customer support platforms in the AI agent builders comparison.
• If you want an assistant for your own work (email, calendar, meeting notes), Lindy’s model, credits per user, is the one to examine.
• If you already run automations and want to add a step that decides, the workflow route is described in the guide on how to build an AI agent.
• If the series of steps never varies, you do not need an agent.