To catch up in 21 seconds. Machine learning case for the YUM! Brands project

Project Objectives

Considering the active growth of KFC in Russia and the opening of more than 100 restaurants annually, there was a need to reduce the costs of ICL Services for the pre-processing of requests and their classification, as well as to increase the correctness of the appointment of responsible employees and the productivity of first-line support specialists.

Service Desk Robotisation

At the bot development stage, several algorithms were selected and artificial intelligence training began using a sample of requests from the Customer’s ITSM system. After several months, the first bot release was ready and launched for testing. The bot classified only those requests that it considered correct with a confidence of 98% and higher.

In June 2018, the bot began to work independently. Its functions included classification of requests, their initial processing and escalation to the second and third lines of support. After 2 months, the effectiveness of the bot amounted to about 35% of all user requests. By the beginning of 2019, the automatic request routing was already fully operational: the bot no longer just classified requests, but distributed them taking into account responsible employees (agents). Requests were assigned to agents available at a certain period of time, taking into account the shift schedule, the workload of the agent team and their level of competence.

In addition, the Autonomous Bot functionality was developed which was launched by a trigger and created or changed user accounts. In total, the project development cycle took only 20 business days.

Results

ICL Services solution allowed distributing requests automatically, providing accelerated classification of requests (21 seconds instead of 5 minutes). Using machine learning, the bot processes more than 35% of requests independently, and another 15-20% of requests are assigned using bot tips about the most effective solution. In general, the accuracy of classification within the framework of the project has increased, as machine learning algorithms do not make human mistakes.