A brief insight into AI supported customer ticketing
Our company made a special AI interface for sorting and classifying customer support requests. Every client creates a ticket in case of a question or a request for help. The ticket can be created on our customer’s website or emailed. Even though, there are two distinct options available (logistics question or technical support question) and of course, the third, “general” option. In the past, the customer support department spent a lot of time sorting the wrongly submitted requests by hand and it took a lot of working hours to just properly classify them. Since the client’s satisfaction is the primary drive of our customer’s business, the response time and quick resolution of customer issues are vital. The AI solution analyzes every customer request and classifies it. The classification is conducted directly in the customer support tool (via Zendesk platform). During the development of our AI solution, we’ve encountered several challenges: We had to prepare the proper training and test sets from the existing tickets We had to properly verify the results The model had to be the integrated into the existing working process Since this is the NLP problem (Natural Language Processing), we used the BERT embeddings for text representations and classified the tickets with SVM (Support Vector Machine) algorithm. In essence, the algorithm has to read the customer’s ticket and based on the recognized text, it forwards it to the logistics support or the technical support deportment. The is achieving a balanced accuracy of more than 90% right now and we keep improving it. Priority of the tickets The second task for us was to determine the priority of the received ticket. Based on the text in the ticket, the model has to prioritize it and assign it to the employee, who is either the best qualified to solve it or who has the most suitable workload in order to solve the ticket in time. This is a typical regression problem, with ticked evaluation on scale from 1 to 100. The final result is a classification of the ticket into three priority groups. Distribution After this initial step, another process takes over and assigns the ticket to the agent, who is the most suitable to solve it in time. The solution has been in use for several months and it has drastically reduced the unnecessary workload of the customer support personnel. After initial mistrust, everybody are now taking the AI process for granted which is in the essence the sole purpose of such integrations. The model runs on the Microsoft Azure, currently still on Azure Virtual Machine and connects to the Zendesk with the use of the API interface.