Conversation Design for Chatbots and Voice Agents

Conversation design is the discipline of deciding what a chatbot or voice agent says, what it asks, in what order, and what happens when things go sideways. It covers the greeting, the questions, the error recovery, the confirmations, and the exit. It is closer to service design than to writing copy, and it is the part of a build that most determines whether customers like the thing.

Why it matters when you are buying or building

Every platform I work with can produce a fluent-sounding agent within an hour. Fluency is not the bottleneck any more. What separates a bot people use from a bot people escape is design: does it ask for the right information, in an order that makes sense, without asking twice, and does it recover gracefully when it does not understand.

Voice and chat need genuinely different designs, and copying one to the other is the most common mistake I see. Chat can show a bulleted list of five options; voice cannot, because a caller will not hold five options in their head. Chat can send a long answer the reader skims; voice has to be two sentences or the caller interrupts. Voice needs turn-taking and barge-in behavior that chat never thinks about. If a client wants both channels, plan on two scripts sharing one knowledge base, not one script on two channels.

A production example

Across the receptionist agents I run, the changes that improved results most were design changes, not model changes. Shortening the greeting so callers stop talking over it. Asking for the phone number early, because a call that drops after thirty seconds is still a lead if you have the number. Confirming the number back digit by digit rather than as one long string. Offering "or press it on your keypad" for anything numeric. None of that is AI work. It is call-flow work, and it moves the numbers.

The other lesson is about honesty in the script. An agent should state what it can do and, explicitly, what it cannot. On one receptionist that told callers it had blocked a number, an ability it never had, the fix lived in the script rather than the model: a plain list of things the agent must never claim to do.

A checklist for a first pass

• Write the greeting to be under five seconds spoken.
• Decide the one piece of information you cannot afford to lose and capture it first.
• Write the failure branch for every question, not just the success branch.
• Confirm anything you will act on: name spelling, phone digits, appointment time.
• Give an obvious route to a human handoff at any point.
• Read the whole thing out loud. If you run out of breath, so will the agent.