Designing Conversations
The words around the model.
Two chatbots have the same model. People love one of them, and hate the other. What's the difference? Design. How it greets you, how it admits a mistake, and how it hands you to a person. Most language projects fail here, rather than in the model.
- Same model, different reactions
- The difference is design
- Most projects fail here
Principles of good conversation
Say what you can do, right at the start, so that people know what to ask. Keep replies short. Ask one question at a time. Confirm before doing anything important. When you're not sure, ask, rather than guess. And never trap someone. There should always be a clear way to reach a person.
A bad reply and a good reply
Compare these. The bad reply is long, it apologises three times, and it doesn't do anything. The good one says what it found, says what it did, and says what happens next. People don't want warmth instead of help. They want help, delivered warmly.
Bad
- We deeply apologise for any inconvenience
- Your satisfaction is very important
- Please be patient
Good
- Your parcel is at the Leeds depot
- I've asked them to prioritise it
- It should arrive by Friday
Failing gracefully
Every system will fail. So plan for that. When you don't understand, say so, and offer choices. If a person takes over, pass on the full history, so that the customer doesn't have to repeat it. And read real conversations each week. Nothing will teach you more about what's going wrong.
Summary
Good design is what makes a model feel smart. Next, when your customers don't all speak English.
Key takeaways- Set expectations first
- Help, delivered warmly
- Plan for failure