What Is a Chatbot? Definition, Types, Examples and How They Work

Most people meet a chatbot before they know the word: the bubble in the corner of a shop's website, the assistant that answers a bank question at midnight, or ChatGPT itself. Whether you typed "chatbot", "chat bot" or "what is a chatbot", you probably want a plain answer and enough detail to tell the different kinds apart.

This guide is for business owners, support leads and students. It covers what a chatbot is, where it came from, how scripted and AI chatbots answer the same question, how chatbots differ from AI agents, virtual assistants and live chat, and how to put one on your website.

What is a chatbot?

A chatbot is software that talks with people. You type or speak a question, and the program replies in plain language, as if you were chatting with a person. Some chatbots pick replies from a script someone wrote in advance; others write each reply on the spot with artificial intelligence (AI). Both kinds can sit on a website.

The useful question is what a chatbot answers from: a scripted bot answers from its script, and an AI chatbot from the text it was given plus patterns learned in training.

Is it "chatbot" or "chat bot"? What the word means and where it came from

Chatbot and chat bot mean the same thing, and the one-word spelling is the standard. Merriam-Webster's chatbot definition lists "chat bot" only as a less common variant, defines a chatbot as a bot designed to converse with human beings, and dates the first known use of the word to 1994.

The meaning of chatbot is literally the two halves: chat, plus bot. A bot is a program that does a task automatically, and the word is short for robot. In his 1994 paper for the AAAI conference, Michael Mauldin of Carnegie Mellon University explains that computer-controlled players in the game TinyMUD were called bots, short for robots, and describes the "Chatter Bot" he built to talk with human players. Merriam-Webster's first known use of chatbot comes from the same year.

The idea is older than the name. Alan Turing's paper "Computing Machinery and Intelligence", published in the journal Mind in October 1950, opens by asking whether machines can think and proposes the imitation game: a judge exchanges typed messages and tries to tell a machine from a person.

What is the history of chatbots, from ELIZA to ChatGPT?

The history of chatbots runs from ELIZA, a keyword-matching program published in January 1966, to ChatGPT, the AI chatbot OpenAI released on 30 November 2022. The table lists the milestones in between, each dated from its primary source.

ELIZA is the program most histories start with. Weizenbaum's paper, ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine, describes software running on MIT's time-sharing system: a user typed on a remote typewriter, ELIZA scanned the sentence for keywords, and a rule tied to the highest-ranked keyword rearranged the sentence into a reply. When no keyword matched, it fell back on stock phrases such as "Please go on". Its main script imitated a psychotherapist who turns statements back into questions.

Weizenbaum noted that some users were very hard to convince ELIZA was not human, and warned that the program showed how easily a machine can create an illusion of understanding. That warning still applies to every chatbot on this page.

The next wave was scripted business bots. When Facebook opened Messenger to bots in April 2016, its announcement described automated updates, receipts, shipping notifications and rich messages with calls-to-action: content a business designed in advance, not text a model generated.

The turning point came on 30 November 2022. OpenAI's launch post says ChatGPT was trained with reinforcement learning from human feedback (RLHF), was free to use during the research preview, and refers to it in passing as the chatbot.

How does a chatbot work?

A chatbot works by turning your message into a reply in one of two ways: matching it against rules someone wrote, or generating new text with a language model. Here both kinds get the same question.

The worked example: a visitor to an online shoe shop types "Can I return shoes I wore once? I bought them 40 days ago." The shop accepts returns of unworn shoes within 30 days, and covers defects with a six-month warranty.

How a rule-based chatbot answers

1. The visitor opens the chat and sees three buttons: Track my order, Returns, Talk to us.
2. The visitor types instead of tapping. The bot scans the message for keywords it was given and finds "return".
3. The Returns rule fires. The bot sends the reply written in advance: "You can return unworn items within 30 days. Start a return here."
4. The details are ignored. "Wore once" and "40 days" are not keywords, so the reply does not address them. If the visitor writes back "but I wore them once", nothing matches and the bot says "Sorry, I didn't understand. Please choose an option."
5. The visitor taps Talk to us, fills in a form and waits for a person.

The bot behaved as designed: predictable and cheap, but unable to apply the policy to this person's situation.

How an AI chatbot answers

1. The visitor types the same question. There may be no buttons at all.
2. The chatbot searches the shop's content. A support chatbot looks through its knowledge base, the collection of web pages, documents and help articles the business gave it, for passages close in meaning to the question. It finds the returns paragraph and the warranty paragraph. This search is the first half of retrieval-augmented generation (RAG): retrieve the relevant text first, then generate an answer from it.
3. It builds a prompt. The instructions (tone, what to refuse, when to hand off), the two passages and the conversation so far are combined into one block of text for the model.
4. The language model writes the reply. A large language model (LLM) is a program trained on a very large amount of text to predict the next piece of text, which lets it write fluent sentences. Here it might reply: "Our returns cover unworn shoes within 30 days of delivery, so a pair worn once and bought 40 days ago is outside the policy. If they are faulty, the six-month warranty may apply. Would you like me to pass this to the team?"
5. It can act, if allowed. With an order lookup switched on, the chatbot could ask for the order number and check the real delivery date instead of trusting "40 days".
6. It hands off when it should. If nothing relevant is found, or the visitor asks for a person, a good AI chatbot says so and passes the conversation, with the transcript, to a human.

The AI version handled the details, but it can also be wrong. If the returns page is out of date, the reply will be confidently out of date too. That is why steps 2 and 6 matter more than the model.

What are the different types of chatbots?

There are five common types of chatbots, and most real products combine two or more of them. The table sorts them by how they decide what to say.

One more split cuts across all five. A general assistant such as ChatGPT, Claude or Gemini answers questions on any subject from what it learned in training. A task bot on a business website answers about one company, from that company's content, and should refuse everything else.

The intent-based type is easy to miss. Bank of America says Erica uses natural language processing grounded in machine learning, not generative AI or large language models, and selects the most appropriate response from a pre-defined set of answers. It understands many wordings, but every answer was written by the bank.

What is an AI chatbot, and what is an AI bot?

An AI chatbot is a chatbot that uses artificial intelligence, today usually a large language model, to understand a message and write a reply that nobody scripted. That is what lets it answer the shoe question above in one message instead of a menu.

"AI bot" is looser: it usually means an AI chatbot, but it can also mean a crawler that never chats.

Chatbot vs AI agent vs virtual assistant vs live chat: what is the difference?

A chatbot talks, an AI agent talks and acts, a virtual assistant is a chatbot that runs personal tasks, and live chat is a person typing. The table separates them by who replies and what each can change.

An AI agent is software built on a language model that can decide to use a tool, such as an order lookup, read the result and then answer. The guide to AI support agents covers the customer-facing kind in depth. Live chat is the opposite end: every reply is typed by a person, which the live chat vs chatbot comparison weighs against automation.

What are some examples of chatbots?

Examples of chatbots range from general AI assistants to small bots that answer one shop's questions. Here are eight, grouped by job; each named product was checked on its maker's own page on 12 November 2026.

General AI assistants

• ChatGPT (OpenAI): the general assistant whose launch on 30 November 2022 started the current wave.
• Claude (Anthropic): introduced on 14 March 2023 as an AI assistant for summarizing, writing, Q&A and coding.
• Gemini (Google): the assistant formerly called Bard; Google's page offers a free plan with a Google Account.
• Microsoft Copilot: Microsoft calls it an AI-powered assistant, and it also works inside Word, Excel, PowerPoint and Outlook.

Customer service and booking bots

• Erica (Bank of America): an intent-based banking assistant in the bank's mobile app that can, among other things, lock a misplaced debit card or track a refund.
• A store's support bot: answers shipping and returns questions from the store's policy pages and looks up an order by its number.
• A booking bot: a clinic or salon chatbot that checks open slots and books an appointment. PepoChat's agent does this through Cal.com or PepoSmart, on every plan.

Internal helpdesk bots

• An IT or HR helpdesk bot: answers employees' "how do I reset my VPN password?" or "how many leave days do I have?" from internal documents, and opens a ticket when it can't.

What are chatbots good at, and where do they fail?

Chatbots are good at answering the same questions quickly, at any hour and in many languages, and bad at judgement, exceptions and anything they were never given.

Chatbots are good at:

• Answering repeat questions (hours, prices, shipping, returns) instantly, at 3 a.m. too.
• Replying in the visitor's own language and wording.
• Collecting details, such as an email or an order number, before a person takes over.
• Handling many conversations at once without a queue.

Chatbots fail when:

• The answer isn't in their content. A scripted bot dead-ends; an AI bot may guess. OpenAI's own launch post warned that ChatGPT sometimes writes "plausible-sounding but incorrect or nonsensical answers".
• The case needs judgement. A bot will not decide that a loyal customer deserves an exception.
• There is no way out. In a Gartner survey of 3,566 business and consumer customers run in February and March 2026, only 27% said they would try a chatbot again after a negative experience, and only 7% had used a chatbot or digital assistant in their most recent service interaction (Gartner, 2 September 2026).

A wrong answer is also the business's problem, not the bot's. In Moffatt v. Air Canada (2024 BCCRT 149), decided by British Columbia's Civil Resolution Tribunal on 14 February 2024, the airline's website chatbot told a customer that a bereavement fare could be claimed after travelling, which the airline's policy did not allow. The tribunal rejected Air Canada's argument that it could not be held liable for its chatbot, held that the airline is responsible for all the information on its website, static page or chatbot, and found the customer entitled to $650.88 in damages. This is one Canadian small-claims decision, not legal advice, but it shows why grounding and a human handoff matter.

How do you add a chatbot to your website?

There are three ways to add a chatbot to a website: build a scripted bot in a visual builder, connect an AI support agent to your own content, or build your own on an AI model's API. For a small business answering customer questions, the second usually fits.

Route 1: a scripted chatbot builder

You draw the conversation as a flowchart of buttons and replies, then paste the provider's code snippet into your site. It suits narrow jobs such as routing visitors or booking a call from a fixed menu. Every new question means editing the flow by hand.

Route 2: an AI support agent trained on your content

You point the tool at your website, upload documents, and the agent answers from them; a person takes over when it can't. Setup is mostly content work: the knowledge base chatbot guide explains what to feed it, and website chatbot best practices covers placement and tone.

Route 3: build on an LLM API

A developer calls a model provider's API, stores your content for retrieval, and builds the widget, inbox and handoff around it. You get full control and full responsibility. The ChatGPT for customer service guide explains why ChatGPT itself, or a custom GPT, can't be dropped onto your site.

How much does a chatbot cost?

A chatbot costs anywhere from nothing to thousands of dollars a month, and the pricing model matters more than the product. Business chatbots usually charge through a free plan with a small allowance, a flat monthly plan, a fee per seat, or a fee per resolved conversation. Building on an API means paying the model provider for usage, plus hosting and developer time.

For scale, PepoChat's Free plan is $0 for 100 AI replies a month, and Starter is $39 a month for 1,000; the full list is on the pricing page. The breakdown of what an AI chatbot costs compares the models at three volumes, and the review of free AI chatbots for websites shows what $0 plans include.

What are the most common misconceptions about chatbots?

The most common misconceptions are that a chatbot learns from every chat, knows your business, understands you, is either all scripted or all AI, and can replace your team.

"It learns from every chat." A chatbot's model does not change during your conversation. Whether chats train future versions depends on the product: OpenAI says it may train on conversations from its individual ChatGPT plans unless you opt out, but by default not on data sent through its API or business plans (OpenAI, checked 12 November 2026). A support chatbot's answers usually change only when someone edits its content.

"It knows my business." An AI chatbot knows what it learned in general training, which does not include your refund window. Without your content, it guesses or declines.

"It understands me." It predicts likely text. Weizenbaum's 1966 warning about the illusion of understanding is still the right instinct when a reply sounds sure.

"Chatbots are either all scripted or all AI." Many mix buttons, intent matching, generated text and human handoff in one window.

"A chatbot replaces the team." Someone still answers handoffs, fixes content gaps and decides what the bot may do.

Where PepoChat fits

PepoChat is the AI kind of chatbot, built for customer support: an AI support agent that answers website visitors from your own knowledge base. You build that from a website crawl (up to 200 pages per crawl), uploaded files such as PDF and DOCX, Notion, Google Drive, Q&A pairs and Help Center articles. The agent searches it before every reply, as described in how answers work, and does not answer from general knowledge about your business.

When a question needs live data, the agent can use 22 prebuilt actions across 13 providers, such as Shopify order status by order number, Stripe subscription checks and help-desk tickets, plus custom actions to any HTTPS endpoint, and it can book meetings through Cal.com or PepoSmart. When it can't answer, or the visitor asks for a person, it hands the conversation to a built-in team inbox. Installation is one script tag.

The limits are real. There is no drag-and-drop flow builder; the closest thing is Procedures, playbooks of up to 15 steps that the AI follows. Channels are the web widget and, on paid plans, voice and inbound phone; WhatsApp and Messenger are in limited rollout. The "Powered by PepoChat" line shows on every plan. Answers change when your content changes, not from chats.

Every chat feature is on the free plan, with 100 AI replies a month and no card. You can create a free PepoChat workspace and test it on your own pages.

What to do next

• Compare the customer-facing kind in detail: What is an AI support agent?
• Decide whether you need a bot, a person or both: Live chat vs chatbot
• Check the plans and reply allowances on the pricing page.
• Build a support chatbot from your own website: start free on PepoChat, no credit card needed.
• Related explainers: What is an AI agent?, What is an LLM? and What is an AI hallucination?.