Text Mining to Identify Customers Likely to Respond to Cross-Selling Campaigns: Reading Notes from Your Customers

Background

Companies are spending large amounts on systems for managing and servicing customers, mostly in the form of CRM systems. But, many companies are failing to see the expected returns with an estimated 55% of CRM projects failing (Rigby, Reichheld, & Schefter, 2002) . Regardless of whether success or failure is achieved, companies are using CRM systems and related technologies to become more customer-centric. Customer-oriented companies have the ability to mass customize and personalize information for individual customers (Teo, Devadoss, & Pan, 2006). CRM enables businesses to acquire, get to know, and service customers according to their specific needs. CRM systems have three primary types of functionality, which are operational, analytical, and collaborative (Teo, et al., 2006). Of greatest interest in the present study is the collaborative functionality because it enables a business to interact with its customer to better service them.

Collaborative CRM systems are a means to engage customers in relationship marketing. Relationship marketing is attracting, maintaining, and developing customer relationships (Voss & Voss, 1997). Relationship marketing requires a deep and personalized understanding of the customer’s needs and characteristics, something service industries are well suited to do because of their personalized interactions with their customers (Voss & Voss, 1997).

This study develops and tests a method for using customer service notes captured from a call center and stored within a service company’s CRM repository with the goal of predicting which customers are likely to positively respond to a cross-selling campaign. It is well recognized that companies can increase their profits by identifying and altering marketing approaches to their different customer segments (Zeithaml, Rust, & Lemon, 2001). We show that text mining free-form customer service notes as a way to segment customers can result in predictive models that are as good as or better than models developed using only structured data.