Introduction
In this article, you'll learn what to keep in mind when your chatbot goes live. Your chatbot isn't live yet? No problem! Here you'll learn how to put your melibo chatbot into action.
You did it! Your chatbot is live and ready for the world. Now what? Your chatbot is already busy answering your customers' questions and taking the load off your support team. But to get the most out of your chatbot, here are five tips & tricks that will improve its quality and performance even further.
1) The traffic light system
For your chatbot to answer as many requests as possible after going live and deliver the right answers, it needs a knowledge foundation. Chatbots also need training to get fit. The more data you feed the chatbot, the better and smarter it gets. To check your chatbot's intelligence, we introduced the traffic light system.
🔴 Red shows that this question has been poorly trained. There's a big need for training here, as the chatbot has little or no guidance for answering this question.
🟡 Yellow means the chatbot has already learned some sample answers but still needs a bit of training to be really sure of its answer.
🟢 Green, on the other hand, means this question is very well trained. There are many sample answers, so the chatbot can recognize user requests on this topic with a high degree of probability.
In your library, the traffic light system lets you see at any time how well your chatbot is trained on the various chat topics. Here you can edit and retrain poorly trained topics.
2) Bot Gym
After going live, more and more requests will reach the chatbot over time. Users' questions may differ more from the sample questions you created. You can use these messages to train your chatbot. That's what the Bot Gym 🦾 is for.
The Bot Gym also uses the traffic light system. Here you can see which chats were successful and where there's still room for improvement. Both the chats and the user's exact messages are highlighted in the respective colors.
- 🔴 Red means not recognized
- 🟡 Yellow means recognized, but not clearly, and
- 🟢 Green means well recognized
Now you can improve your chatbot with just two clicks. First click on the user's message that needs retraining. Then, on the right, choose the chat name the user's phrasing should be trained to. The Insert button adds this answer to the chatbot's repertoire so it can be answered successfully in the future.
The exact process is explained again in this video.
3) Insights
Insights provide a powerful analysis tool for your chatbot's go-live, letting you evaluate answer quality and the most frequently asked topics, among other things. Besides an overview of the number of requests in a given period, you can also see how many messages still need training or weren't understood by the chatbot.
You should check Insights regularly to see how your chatbot's quality is doing and how often it's used. A quick overview of your customers' usage behavior and times can deliver results for the whole company. Maybe one product is especially popular or one day of the week has particularly many requests. With this information, promotions can be used and steered more effectively, for example.
4) Confidence score
Every time a user enters a message after go-live, the chatbot analyzes how accurately the message was recognized. It checks which chat topic matches best. This accuracy is output as a number and called the confidence score or recognition rate. The higher this number, the better the match. If several answers could fit the message, the one with the highest confidence score is chosen.
There is a minimum value the confidence score must reach for the corresponding answer to be given. This limit is called the confidence threshold. In a later update, you'll be able to adjust this threshold for your chatbot yourself, flexibly.
5) Prevention instead of reaction
As you can see, there are many ways to analyze user requests at go-live and close gaps in the chatbot. It's even better, though, if the chatbot can answer as many questions as possible without a user having to ask first before the missing answer is noticed. To make sure that doesn't happen to you, we've prepared 100 questions to reduce standard requests in e-commerce.
