Our client, a leading Swiss retail company, receives hundreds of customer requests and questions every day, with an increasing number of requests each year. Since these requests need to be manually sorted into more than 20 different categories, their handling was getting more and more cumbersome, costly and time consuming. Once sorted, a request is treated by the agent specialized in the related domain. However, the quality of the request processing decreased under the load of increasing requests. Our client was looking for ways to speed up this process while at the same time improving the quality of the answers.
OpenWT was asked to provide an Artificial Intelligence system which would classify the different e-mails and directly send them to the mailboxes of the responsible agents. The AI system should also support the agent in answering the requests.
In close cooperation with the client, we identified the scope of the project and defined the desired functionalities and requirements for different use case scenarios. The client’s goal was to automatically redirect all e-mails received by the customer support to the most suitable agents and at the same time help these agents to process the request. Together with the client, the OpenWT team identified the elements which impact the classification process. Along with the classification functionality of the target system, we established a list of contextual information such as language identification and enriched customer information as well as a list of desired standard answers.
2. Train AI system:
An initial set of both structured and unstructured e-mails was provided to train the AI system using supervised learning. Unstructured e-mails consist of free text by the sender while structured e-mails originate from various types of contact forms. A succession of three main activities was performed in order to reach the goals defined during the Blueprint phase:
- Feature engineering: This part includes the cleaning and pre-processing of the e-mails. The cleaning concerns the detection of formatting elements and noise. The pre-processing includes a broad set of tools, with a major role being played by AI algorithms from Natural Language Processing. The goal of this step is to extract the most valuable information contained in the e-mails and to put them into a suitable form for further processing. Feature engineering is a crucial part and greatly benefits the accuracy of the AI system.
- Email classification: Various AI techniques were carefully trained on the data obtained in the previous step. This training procedure is iterative and one has to re-do some feature engineering in order to get the best possible classifications of the e-mails into the different categories. In the end we used a well-balanced mixture of AI techniques which made our classifier very stable and highly performing.
- Email enrichment: By combining Named Entity Recognition with a CRM data base we could enrich the incoming e-mails with valuable information about the customer. Furthermore, the AI system was trained in this step to use this additional information in order to propose to the agent a possible answer.
3: Feedback and Tuning
In this phase we switched to real-time processing of the e-mails. In close cooperation with the client, we integrated an automatic feedback-loop into our system in order to further learn from errors in the classification and proposed answers. This gave the AI system the ability to learn continuously and improve itself steadily.
Thanks to the Artificial Intelligence solution provided by OpenWT, the client drastically decreased the processing time of customer requests by e-mail and reduced operational costs. In addition, it significantly increases the quality of the answers delivered by support agents. With an accuracy of 97%, the solution demonstrates its reliability and robustness.