Implementation of virtual sales agents at e -stores

Customers who shop online at cyberstores (e-stores) benefit from the ease of comparison-shopping over the Internet. E-stores, however, experience a lower conversion rate (buyers-to-visitors ratio) compared to brick-and-mortar stores. E-stores also suffer from the lack of touch-and-feel experience, diverse customers, and the absence of salespersons. However, e-stores can improve their operations by allocating resources to visitors, improving website content/structure design, customizing visitors' experience at a personal level, and providing purchase recommendations. Intelligent agent (IA) technology provides a viable approach to processing available visitors' data for learning about their interest and intention and for allocating resources in real-time. A practical application is to use IAs to improve sales operations through allocating discounts. We implement this application through the following steps. A framework is provided for implementing IAs that support e-store operations including sales, demand forecasting, and order fulfillment. The framework shows how existing databases can feed IAs with required information and support IAs' communication needs. Regarding the implementation of sales agents, the framework is used for identifying and discussing important issues, including real-time identification of visitors' interest and intention. To address the need for a real-time model of visitors' interest and intention, a high-order Markov chain---a model commonly used for visitors' interest---is augmented with intention measures that can be calculated in real-time. Labeling visitors as buyers and non-buyers is considered as the basis for sales agents' decision engine and is formulated as a decision rule. As an example, a simple rule is provided based on a probability-of-purchase measure. Finding the rule's parameter is formulated as an optimization problem and illustrated for three loss functions. Other research contributions are: applying empirical Bayes estimation method as an alternative to Maximum Likelihood and Bayes methods; proposing a heuristic estimator for the prior distribution of the empirical Bayes method that outperforms method-of-moment estimator originally used for the prior; and developing a simulation program (SurfSim) to generate clickstream data and provide a platform for implementation purposes.