The word chatbot covers any program that converses in writing. The most frequently cited predecessor is ELIZA, written in 1966 by Joseph Weizenbaum at MIT, which imitated a therapist by rephrasing what the other person said. Website chatbots in the 2010s followed the same rule-based principle: a decision tree, buttons, keywords and answers written in advance.
The main types of chatbot
- The scripted chatbot. Every path is written by hand: “Track my order”, “Book an appointment”, “Speak to an adviser”. It works as long as the visitor stays on the planned path and gets stuck as soon as they ask an unexpected question
- The language model chatbot. It uses a language model: it understands the question as asked and writes an answer. When it answers from specific documents and cites its sources, it is more often called an AI assistant
- The hybrid chatbot. It combines button-based paths for simple requests with a language model for everything else, which is common on customer service websites
When a scripted chatbot is enough
For a short path that is identical for everyone, a script remains simple and predictable: choosing a time slot, finding the right department, tracking a parcel by its number. It only says what has been written, which makes it predictable, although a prewritten answer can also be outdated. Its limit appears as soon as questions vary: an unexpected question often ends with “I didn’t understand” or a link to a form.
What changes with a language model
Visitors no longer have to guess the right keywords. In return, answers are no longer written in advance: the tool needs boundaries so it answers from selected documents, says when it does not know and hands sensitive topics over to a person. Without these boundaries, a modern chatbot can produce a convincing but false answer, a hallucination.
Chatbot, chat widget, messaging: different things
The chat widget is the window at the bottom of the page; the chatbot is what answers inside it. The same window can also be used by a team member replying live.
KernChat assistants belong to the language model family, with boundaries: they answer using the organization’s documents and hand over to the team when they do not know (see the features).