Machine Learning Assisted Troubleshooting Flows

Abstract

An interactive troubleshooting flow (TF) directs customer care agents in diagnosing customer service issues and resolving these issues remotely. This step-by-step process, which helps agents triage issues and further investigate customers’ concerns, is dependent on the retrieval of customer telemetry data. A single TF document is a graph that an agent and customer traverse together. Each node or step in this graph is a question for the customer, an automated diagnostic test, or an action completed by the agent. The process has typically been based on rules and domain heuristics with a pre-defined path determined by the initial description of a customer’s concern. In this paper, we illustrate how a machine learning approach using available telemetry upfront or sufficiently early in the troubleshooting process, can speed up the time it takes to resolve the customer’s concern by better predicting the likelihood that an onsite technician will need to be dispatched to fully resolve the issue. We further demonstrate how cost-savings of such an approach could be assessed using model prediction outcomes, to help with model deployment decisions.