Chat data contains a lot of valuable information
In the world of insurance, customer support is a key player. That is why our client, a Swiss insurance company, wanted to evaluate their recently established live customer support chat. More precisely, they wanted to find out what topics were addressed in the interactions between their customers and insurance agents, as well as receive insights on the questions that had been asked. Doing so, they could use the data to improve their system, and therefore support their client better and more efficiently.
How to get deep insights out of your customer support chats
As important and informative as customer chats are, they can often be arduous. The number of requests and questions is constantly increasing, and they can be as diverse as they can be complex. Manually extracting valuable insights from those correspondences is not an easy task and a time consuming one. Therefore, we applied a 4-step approach, leveraging natural language processing with artificial intelligence algorithms to be able to provide our client’s customers with the best possible advice. Also, once all the data was analyzed and the insights extracted, we were able to prototype a chatbot to automate the support chat.
We anonymized the chat protocols using the Swisscom Anonymization AI Enabler
Since the insurance chat protocols contain a wide range of sensitive information, we had to ensure data privacy. The anonymization process masks any personal identifying data such as names, phone numbers, addresses, dates and birthdays, as well as insurance policies and car registration plate numbers.
We used Pattern Mining and Information Extraction Techniques
With those techniques, we were able to identify the questions asked by the customers and the answers given by the insurance agents. We then structured the dialogues into the patterns of speech of the customer and agent, recognizing irrelevant turns such as greetings and requests for more information.
Using the k-means algorithm, we created clusters on the extracted questions
This allowed us to group the questions into topics, which resulted in a set of well-defined topics from the insurance domain. The number of questions that pertained to a topic indicated the prevalence of this topic in the chat interactions, allowing us to understand the overall distribution of topics.
Impact by leveraging AI
questions and answers paired
major topics identified
less time consuming