Our client, a European insurance company operating in Switzerland, has a highly frequented customer support chat. The incoming load of customer requests grows day by day. The pressure on the chat-agents increased and it was hard to ensure a good quality of the answers. The goal of the client was to support their chat-agents and make them more effective. We helped them to integrate an AI solution which proposes answers to an incoming customer question into the existing interface.
1. Pre-processing the chats:
This step consisted in cleaning the available chats and identifying the questions that were being asked by the customers and the answers provided by the insurance agents. Using Pattern Mining and Information Extraction techniques, we structured the dialogues into the turns of speech by the customer and agent, recognized irrelevant turns such as greetings and obtained a high-quality set of question and answer pairs.
2. Train chatbots
In order to leverage the extracted information, we used various approaches including state-of-the-art Deep Learning technology to train different chatbots on the extracted question and answer pairs. The resulting bots were able to answer new questions with a high degree of accuracy.
3. Propose answers to agents
Once the chatbots were trained we switched to real-time. For each new incoming customer question we used a mixture of the different bots to propose the agents three suitable answers which he can easily integrate and modify if needed. Since there are always questions which are very customer specific and hard to answer automatically, we introduced a scoring system for the answers, which gives the agent a quick overview about the accuracy of the answers.
4. Self-learning loop
We carefully store all the reactions and behaviors of the agents towards the proposed answers. We used this data as the basis for the implementation of a self-learning loop in the AI system. This allows the chatbot to quickly adapt and improve itself even further.
Thanks to the cutting-edge Artificial Intelligence solution provided by Open Web Technology, the Client was able to increase the number of treated chat-requests per agent. At the same time, he could ensure a high quality of the answers of the agents. Of all the chat questions asked, in 60% of the questions, the intention and topic were correctly identified, which forms a good basis for an Entity Intent Chatbot.