In plain terms
A chatbot is software that people talk with. The older kind works like a phone menu in writing: it recognises a few keywords, walks the user down a prepared path and gets stuck when the user leaves it. The newer kind, built on a language model, understands free text, keeps track of the conversation and writes its own replies, so it can handle questions nobody scripted. It can also be wrong in a perfectly confident tone.
Why it matters
The chatbot is how most people and most customers meet AI, and for many organisations it is the first project. Scripted bots are predictable and limited: they say only what was written for them. LLM chatbots cover far more questions and need far less scripting; in exchange they can invent an answer or be talked into saying something off-policy, and a company is usually held to what its bot tells a customer. Decide which sources it may answer from, what it must hand to a person, and how answers are tested before launch.
Example
A retailer's scripted bot offers three buttons: orders, returns, other. A customer types, “The kettle I bought as a gift arrived dented. Can I swap it without a receipt?” The bot replies that it did not understand. The retailer replaces it with an LLM chatbot connected to the returns policy, which answers the same question in two sentences with a link to the policy. Within three months, the share of conversations resolved without a handover rises from 23% to 62%.
Most often confused with
Chatbot vs. Agent
A chatbot's output is a message. An agent's output is a completed piece of work: a record changed, a refund paid. The same model can sit behind both. What differs is whether it has tools, permission to use them and a loop that keeps going until the goal is met. Many chat assistants now switch into an agent mode for some requests, and that is the moment when permissions and oversight start to matter.
Origin: The first widely known chatbot was ELIZA, written by Joseph Weizenbaum at MIT in 1966; the word is short for “chatterbot”, coined by Michael Mauldin in 1994.
Under the hood
Three generations are in use. Rule-based bots match keywords or patterns to scripted replies, as ELIZA did in 1966. Intent-based bots use a classifier to map a message to one of a fixed set of intents, extract entities such as dates or order numbers, and follow a dialogue flow designed in advance. LLM chatbots generate each reply from a system prompt, the conversation history and, usually, passages retrieved from a knowledge base. A production LLM chatbot is the model plus that prompt, retrieval, memory across turns, guardrails on input and output, a handover path to a person, and logging. Many deployments are hybrids, with fixed flows for regulated steps such as identity checks and generated answers elsewhere. Typical measures: resolution rate without handover, escalation rate, customer satisfaction and the groundedness of answers. Typical failures: invented policy, prompt injection and jailbreaks, and loss of context in long conversations.