A chatbot is only as good as its conversations. Technology decides what a system can understand, but design decides whether people want to keep talking to it. Conversation design is the discipline that deals with this, and almost every serious book or guide on the subject returns to the same principles.
Begin with a purpose
Before writing a line, decide what the bot is for. A narrow, clearly stated purpose, such as helping someone track an order or find an opening time, produces a better experience than a vague aim to “answer anything”. List the five or ten most common requests and design for those first. Everything else can wait.
Know the user
Interview or observe the people who will use the bot. Notice the words they actually use, which are rarely the words an organisation uses internally. Collect real phrasings, because users will ask the same question in dozens of ways. This research shapes both the script and the training data.
Give the bot a voice
A voice is a consistent way of speaking: formal or relaxed, brief or warm. Write a short style note and share it with everyone who edits dialogue. Avoid pretending to be human. Users generally do better when the system is open about being software, and in many places honesty about this is expected.
Keep turns short
People read chat quickly. Break information into small pieces, ask one question at a time and put the most useful part of a reply first. Use buttons or suggested replies where choices are limited, which reduces typing and errors.
Plan for misunderstanding
Every bot will fail to understand something. Design the failure:
– Say plainly that you did not understand, and avoid blaming the user.
– Offer examples of what can be asked.
– Rephrase once, then provide a different route, such as a menu.
– Never trap someone in a loop.
Confirm before acting
When a conversation leads to an action, such as cancelling a service, repeat the key details and ask for confirmation. This small step prevents many costly mistakes.
Offer a human exit
Some topics are sensitive or complex. Make the route to a person easy to find, and pass the conversation history on so the user need not repeat themselves.
Write for the ear and the eye
If the bot also speaks, read every line aloud. Spoken replies need shorter sentences and fewer lists. Written replies can use formatting, but they should still be accessible to screen readers.
Handling tone and emotion
Users often arrive annoyed. A bot should acknowledge frustration briefly and then help, rather than offering cheerful scripts that sound dismissive. Humour can work but is risky across cultures and moods, so use it sparingly and only where the voice guide allows it.
Maintain the script
Dialogue is never finished. Review conversations regularly, remove answers that have gone out of date and add the questions people actually ask. Treat the script as content with an owner, a review date and a change log.
Test with real people
Even a paper script can be tested by having someone play the bot. Watching where people hesitate reveals problems no review will catch. Read on in rule-based and AI chatbots to see how these designs meet the technology, or return to the guide to choosing a book.