What is a chatbot?
A chatbot is a software program that holds a text or voice conversation with a person, usually to answer questions or complete a simple task. The earliest chatbots matched keywords against scripted replies. The next generation used natural language understanding to classify each message into an intent and walk the user through a fixed decision tree. Current chatbots are built on large language models (LLMs), which lets them handle unexpected phrasing and compose answers rather than pick from a list.
Most business chatbots today combine an LLM with retrieval-augmented generation. The system searches a knowledge base of approved documents, hands the relevant passages to the model, and the model answers from those passages. This grounding keeps the answers tied to company policy instead of to whatever the model absorbed during training.
Chatbot compared with conversational AI
A chatbot is a single application with a narrow job. Conversational AI is the broader set of capabilities, including speech recognition, dialogue management, and integration with back-end systems, that a chatbot or a virtual assistant draws on. A chatbot that can only answer questions is a reference tool. One that can also look up an order, book an appointment, or open a ticket has crossed into being an AI agent, and it needs the controls that agents need.
Chatbot in practice
A county government runs a chatbot on its website for permit and licensing questions. Residents ask how to renew a business license or which documents a building permit needs. The bot retrieves the answer from the county’s own ordinances and fee schedules and cites the page. Questions it cannot answer, and anything involving a specific case, are handed to staff with a transcript attached. Call volume to the permits desk falls, and the staff who remain on the phones handle the complicated cases that needed them in the first place.
What to watch for
- Restrict the bot to an approved knowledge base and test what it says when the answer is not there, because a confident wrong answer about a fee or a deadline is worse than no answer.
- Give users an obvious way to reach a person, and pass the conversation along so they do not start over.
- Decide what the bot may do versus only say; any action that changes a record needs an approval step and an audit log.
- Log every conversation and review a sample weekly, since the failure modes of an LLM differ from those of a scripted bot.
- Check where conversations are processed and stored, especially when customers share account numbers or health details.