What is a chatbot, and what is conversational AI?
A chatbot is any program that holds a conversation. The early ones were decision trees with a text box: press 1 for orders, type "refund" for refunds, and anything off the script gets "Sorry, I didn't understand." Conversational AI is the newer generation, built on large language models that understand loosely worded questions, keep track of context and answer in fluent sentences.
Most brand chatbots today sit somewhere between the two. They use a language model to understand what the customer wants, and a set of approved sources and actions to decide what to say and do: look up an order, book an appointment, recommend a product, hand over to a person.
Why it matters
Why chatbots matter
Customers expect an answer now, at any hour, in the channel they are already using. A well built chatbot handles the repetitive questions (shipping times, store hours, returns) instantly and frees human staff for the conversations that need them. On a product page it can work like a knowledgeable shop assistant, asking two questions and narrowing a catalog to the right three items.
A bad one does the opposite. It traps people in loops, invents answers and turns a simple question into a complaint. And because it speaks in the brand's name, what it says is the brand's responsibility: in 2024 a Canadian tribunal held an airline liable for refund advice its chatbot had made up.
In practice
How chatbots work in practice
- 01
The build starts with the questions, not the model. A team pulls a few months of real customer messages, groups them, and decides which ones the bot should answer, which it should route and which it should never touch. The model is then connected to approved content (help pages, product data, policies) and told to answer only from it, a setup usually called retrieval-augmented generation.
- 02
Then come the voice and the exit. The bot is given the brand's tone, a clear statement that it is a bot, and an easy route to a human whenever it is unsure or the customer asks. Before launch it is tested against awkward, angry and adversarial questions, and after launch the conversations are reviewed every week to find the gaps.
Watch out
What people get wrong about chatbots
Using them to avoid customers. A chatbot that exists to deflect contact, with no path to a person, saves money on paper and costs loyalty in practice. The other mistake we see is letting a language model answer from its general knowledge rather than from the brand's own facts, which is how a bot ends up confidently promising a discount that does not exist.
A third mistake is launching and walking away. Products change, policies change and customers find new ways to ask the same thing, so a bot that was accurate in the spring is quietly wrong by the fall. We treat a chatbot like any other piece of brand content: it has an owner, a review cycle and a standard it has to meet before anything it says goes out in the brand's name.
