What Is a Chatbot? A Plain-English Explanation

The definition, broken into its parts

A one-sentence definition is easy to read and hard to use. It is more useful to take the idea apart, because each piece of it is a decision somebody made when they built the thing, and those decisions are what make one chatbot good and another one useless.

Chatbot

A chatbot is a software program you talk to in ordinary language, by typing or speaking, that answers or acts without a human being involved in each individual reply.

The word is a contraction of chat and robot, and it has been used loosely enough that it now covers everything from a four-button menu to a system that reads a 200 page manual before answering. What all of them share is the interface. You express what you want in a sentence, rather than by hunting through a menu, and something writes back.

That is the whole category. Everything else, including whether it is any good, comes down to three questions you can ask about any chatbot you meet.

The three questions that describe any chatbot

• Where does the answer come from? A list a person wrote earlier, or text composed on the spot?
• What is it allowed to know? Your documents, the whole internet, or a handful of scripted facts?
• What happens when it does not know? It says so, it guesses, or it fetches a human.

Examples: Website support widgets, Phone menus, ChatGPT, In-app onboarding assistants

Natural language

Natural language means ordinary human sentences, as opposed to keywords, form fields or menu options that a computer has been designed to expect.

A search box wants keywords. A form wants your details in the boxes provided. A chatbot is supposed to accept the way you would actually say it: "my order hasn't turned up and I'm going away on Friday". Handling that sentence means recognising a delivery problem, a missing order and a deadline all at once, from a phrasing nobody wrote down in advance.

This is the part older chatbots were worst at, and it is the reason so many people learned to distrust them. If a system can only recognise the phrasings its author anticipated, then every slightly unusual way of asking produces the same reply: I did not understand that. The frustration people associate with the word chatbot is almost entirely a memory of this specific failure.

Automation

The bot half of the word means the reply is produced by software on its own, in the moment, with nobody sitting at a keyboard typing it.

This is what separates a chatbot from live chat. Live chat is the same window with a person behind it, so the answer quality is whatever your staff can manage and the response time is whatever their queue allows. A chatbot answers immediately and at any hour, and it answers the hundredth person as fast as the first. The trade is that it answers with whatever competence it was built with, which does not improve because the question is urgent.

Most real deployments are not either-or. The software handles the questions it can, and hands the rest to a person along with the transcript so nobody has to repeat themselves. The comparison between the two is worked through in chatbot vs live chat.

Knowledge source

A chatbot's knowledge source is the material it is allowed to answer from, and it is the difference between a bot that knows your refund policy and one that invents a plausible-sounding version of it.

Nothing about being an AI chatbot means it knows anything about your business. A language model trained on the public internet has never seen your prices, your opening hours or your returns window. If you point it at a question about them with no source material, it will produce something fluent and wrong, which is the origin of nearly every hallucination story you have read.

So a business chatbot is built by giving it your material to read: your website, your help articles, your PDFs, your policy documents. The good ones then show you which page each answer came from, so a wrong answer points at a document you can fix rather than at a mystery. When you are comparing tools, this is the single most useful thing to ask about.

A short, dated history of the chatbot

Chatbots are older than the personal computer. The idea has been worked on continuously since the mid-1960s, and knowing the sequence makes the current moment much easier to understand, because almost everything before 2020 belongs to one approach and everything after it belongs to another.

Read down that column and you will notice the break. For the first fifty-odd years, from ELIZA to the Messenger rush, every one of these systems worked by matching your input against patterns a human had written and returning text a human had written. They got more elaborate, never different in kind. What changed around 2020 is that a model could compose a sentence that had never existed before, appropriate to a question nobody anticipated. That is the line between the old chatbots people learned to hate and the current ones.

The six kinds of chatbot you will actually meet

Every vendor and every article carves the category up slightly differently, but six groupings cover essentially everything in use. They are listed here roughly from simplest to most capable.

1. Rule-based chatbots

Somebody drew the conversation as a flowchart. If the visitor says this, reply with that, and offer these next steps. Every sentence the bot can utter already exists in its configuration. Predictable, cheap to run, and completely stuck the moment a question falls outside the chart.

2. Keyword-based chatbots

A variation that scans your message for trigger words and picks the reply attached to whichever it finds. It copes slightly better with varied phrasing than a strict flowchart, and it fails in a distinctive way: mention two topics in one sentence and it latches onto the wrong one.

3. Menu or button-based chatbots

The bot does not try to read your sentence at all. It offers buttons and you pick one. This is honest software. It never misunderstands you, because it never attempts to. It is also the version most likely to leave you clicking through four levels to discover your question is not on the list.

4. Contextual chatbots

These use machine learning to work out what you meant, which is called intent recognition, and they remember earlier turns in the conversation. Ask about a product, then say "how much is it", and the bot knows what "it" refers to. They still choose from responses somebody prepared, but they choose far better.

5. Generative AI chatbots

These send your question, along with relevant passages from a set of documents, to a large language model such as Claude, GPT or Gemini, and the model writes an answer that did not exist until you asked. This is what most people now mean by AI chatbot. It handles phrasings nobody planned for and it can combine facts from two different pages into one reply, which no earlier type can do.

6. Hybrid chatbots

Generation for open questions, fixed rules for the paths where the exact wording or the exact action matters: identity checks, legal disclosures, collecting a phone number, handing over to a person. Nearly every serious business deployment ends up here, whatever it is marketed as.

The boundaries between these are blurring, and a lot of products sit in two categories at once. If you want each type pulled apart properly, with what each costs and where each still wins, the six types of chatbots goes through them one at a time. The narrower question of scripted versus generated is covered in rule-based vs AI chatbots.

When a chatbot is not the answer

Knowing what something is includes knowing when not to buy it. Businesses tend to arrive at this question from one of four positions, and the honest advice is different at each.

One: answering by hand is genuinely fine

A few enquiries a week, each one slightly different, and you recognise most of the people sending them. A reply you type yourself is faster than anything you could set up, and it is better, because you can tell from the wording that this customer is anxious about the sizing and say something reassuring about it. There is nothing here to automate. A chatbot at this stage is one more thing you have to check on.

Two: you notice you are repeating yourself

The same question, three times this week, answered each time by hunting down an old reply and pasting it. That repetition is the first genuine signal. It is still not a reason to buy software, though, because the answer you keep pasting lives in your sent folder and nowhere else, and a chatbot can only answer from material it has been given to read. Put those answers on your website as ordinary pages and watch what happens. Quite often the questions stop, and you have fixed it for nothing.

Three: doing it by hand starts costing you

Two colleagues give two different answers about the same returns window, and a customer quotes the wrong one back at you. Messages arrive on a Sunday evening and get read on Tuesday. Nobody can say which question comes up most, because the evidence is spread across several inboxes and one person's memory. This is where the software becomes worth the setup, and it is worth noticing what had to be true first. The answers existed in writing. And there was a person the bot could hand the hard ones to.

Four: the conversation that should never reach a bot

Some conversations are the wrong shape for this technology however well it is built. A bereavement. A question whose answer is legally required to be worded in a particular way. In cases like those an instant, fluent reply is not neutral, it actively makes things worse, and the right design is to recognise the situation and fetch a person on the first message. That is something somebody has to configure deliberately. It does not happen by default, and no vendor will raise it for you.

How a modern chatbot produces an answer

Here is what happens between you pressing enter and words appearing, for the generative kind, with no mathematics and no acronyms left undefined.

Step one: it works out what you are asking

Your sentence is compared against the kinds of thing people ask this particular bot. Is this a question about pricing, a complaint, a request to speak to somebody? This is called intent classification, and it decides what the system does next rather than what it says.

Step two: it goes and finds relevant material

The system searches the documents it was given for passages that relate to your question. Not a keyword search: it compares meanings, so a question about "getting my money back" will surface a page titled "Refunds". This search step is the reason a well-built chatbot can answer about your business at all.

Step three: a language model writes the reply

Your question and those retrieved passages go to a large language model, which is a system trained on enormous amounts of text to predict what words should come next. It reads the passages and writes an answer grounded in them. Combining search with generation like this is called retrieval-augmented generation, and it is the standard design for business chatbots.

Step four: it decides whether to hand over

If the retrieved material did not actually answer the question, a well-configured system says so and offers a person, rather than producing something confident and invented. Whether a product does this properly is one of the few genuinely important differences between them.

That is the whole loop, repeated for every message, usually in a second or two. A longer walkthrough with the moving parts named is in how do chatbots work. If you are thinking about building one rather than buying one, how to build a chatbot is the practical version.

Where chatbots actually turn up

The category is much wider than the bubble in the corner of a website, and the reasons organisations deploy them differ a lot by setting.

• Company websites, as a widget that answers product and support questions and collects contact details from people who are interested but not ready to buy.
• Messaging apps and team chat, so a customer or a colleague can ask in the place they already have open rather than opening a support portal.
• Phone systems, where a voice chatbot handles the front of the queue before a human picks up.
• Inside software products, walking a new user through setup or answering questions about a feature without them leaving the screen.
• Internal help desks, answering staff questions about expense policy, holiday allowance or which form to use, which is a surprisingly large share of real deployments.
• General-purpose assistants such as ChatGPT and Claude, which are chatbots in the same sense but are pointed at everything rather than at one organisation's material.

Two forces explain most business adoption. The first is that a question asked at the wrong hour usually goes unanswered, and online shopping is full of moments where a small uncertainty ends the visit. Baymard Institute's meta-analysis of published studies puts the average documented online shopping cart abandonment rate at 70.19%, which is why so many shops now put an answer box on the product page rather than a contact form on a separate page.

The second is the cost of the alternative channels. ContactBabel's 2026 US Contact Center Decision-Makers' Guide puts the average cost of an inbound call at $7.20, which it reports is 47% more than handling an email and 23% more than a web chat. Anything that answers a routine question in the cheaper channel, instead of it becoming a phone call, is worth real money. What share of questions a chatbot can take is highly specific to your business, and any vendor quoting you a universal deflection percentage is quoting a number that does not exist. Work it out with your own volumes in the chatbot ROI calculator, or read the sector-specific version in ecommerce chatbots.

What chatbots genuinely cannot do

Sales material for this category is unusually loose, so it is worth being concrete about the limits. None of the following is fixed by picking a better vendor.

Things they cannot do at all

• Know anything you have not told them. A chatbot with no access to your order system cannot tell a customer where their parcel is, no matter how advanced the model behind it.
• Guarantee the exact wording of an answer. Generated text varies between runs, which is a real problem for legal disclosures, medical instructions and regulated financial language. That work still belongs to fixed scripts.
• Take responsibility. When an answer is wrong, the consequences land on your organisation, so somebody has to be reading the transcripts.
• Feel anything. A chatbot can produce sympathetic wording, and that is not the same as recognising when a situation has become serious enough to need a person.

Things they do badly enough that you should not rely on them

• Answering from thin material. Ask about a policy that exists only in somebody's head and you get an invented answer delivered in the same confident tone as a correct one.
• Multi-step tasks with real consequences, such as processing a refund or rebooking a flight. Some products do this. Test it hard before you switch it on.
• Anything ambiguous or emotionally charged, where a customer needs to feel heard rather than informed. This is the clearest case for handing over quickly.

There is also a naming problem worth knowing about before you sit through a sales call. A chatbot is not an AI agent, though the words are used interchangeably in marketing. Anthropic's engineering write-up "Building effective agents" draws the line usefully: agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks", while workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Most products sold as agents are the second thing. Gartner named this pattern "agent washing" in a June 2025 press release, estimating that only around 130 of the thousands of vendors claiming agentic AI were genuine.

This is not a reason to avoid the category. It is a reason to ask what a product does rather than what it is called, and the difference is unpicked properly in AI agent vs chatbot.

What getting this wrong costs

The monthly fee is the part you can see. The costs that decide whether this was a good idea never turn up on an invoice at all.

Begin with the customer who believed a wrong answer and acted on it. They were told the jacket would arrive before the weekend, so they did not order the alternative from somewhere else. A well-built bot cites the page it used, which is a real advantage, and it has a side effect nobody mentions: an answer with a source attached looks checked, so people question it less than they would question a stranger on live chat. The apology is the small part. The work is the escalation that follows, and the awkwardness of your own software having put the mistake in writing.

Then there is trust, which is spent once and does not refill. Most people who get a bad answer do not tell you. They close the window and stop asking, and every question they might have asked afterwards never arrives. In the reporting, that conversation looks fine: it ended, nobody escalated, no complaint was logged. So the number meant to show the thing working has quietly recorded a failure as a success, which is why somebody needs to read actual conversations in the first weeks rather than watch a chart.

The third cost arrives late. A chatbot that answers from your content is only ever as good as that content, and every question it cannot answer is an instruction to go and write something. Somebody has to do that writing, week after week, long after the launch has stopped being interesting. Usually nobody was given the job, because on launch day the thing worked. What you are left with is a fluent, well-cited description of a business you no longer run. So before you compare products, answer a different question: who, by name, reads what people asked last week, and who is allowed to change the page they were asking about?

Which kind of chatbot fits which reader

People arrive at this question from three quite different places, and the useful next step is different for each.

If you are in the middle column, the practical comparison is between the big incumbent suites and the smaller focused tools. Start with Intercom pricing and the Zendesk chatbot pricing breakdown to see how the per-seat and per-resolution models differ, then look at small business chatbots for the flat-rate end of the market.

Frequently asked questions

Sources

• Joseph Weizenbaum, "ELIZA", Communications of the ACM (1966) - ELIZA's 1966 date, authorship at MIT, and its keyword-matching design.
• Baymard Institute, cart abandonment rate research - The average documented online shopping cart abandonment rate of 70.19%, from a meta-analysis of published studies.
• ContactBabel, "2026 US Contact Center Decision-Makers' Guide" - The average inbound call cost of $7.20, reported as 47% more than an email and 23% more than a web chat.
• Anthropic, "Building effective agents" (19 December 2024) - The distinction between agents, which dynamically direct their own processes and tool usage, and workflows, which run through predefined code paths.
• Gartner newsroom, press release 25 June 2025 - The term "agent washing" and the estimate that only around 130 of thousands of vendors claiming agentic AI were genuine.