Retail in High Definition: Using Video Analytics in Salesforce Management

Abstract

Abstract Staffing decisions typically account for a large portion of a retailer's operational costs. The effectiveness of these decisions has often been analyzed by relating staffing levels to revenues. However, such approach does not explicitly consider the mechanisms by which the staff can contribute to generate revenues, such as customer assistance. This motivates the development of a fast, efficient, high-frequency method to measure customer assistance in real time. The method relies on the use of short videos that track only a portion of a customer's shopping path. The recorded videos may not track all the relevant information to identify a customer-employee interaction, i.e. they might be censored. Accordingly, we develop a survival model to analyze these data, defining unbiased estimates of customer assistance. This methodology also gives insights into how staffing decisions translate into different levels of customer assistance under different congestion scenarios. For example, when the store is congested, increasing the staff from one to four employees can increase the fraction of customers receiving assistance from 38% to 45%. Furthermore, these assistance rate measures can in turn be used to assess the economic impact of assisting customers in terms of conversion or basket size. This introduces important estimation challenges related to the endogeneity of customer assistance (e.g., if the customers that are more likely to purchase are also more likely to seek assistance) and the measurement error in customer assistance rates. We address both issues using an instrumental variables approach that relies on variations on service capacity as a driver of exogenous variance in customer assistance. In particular, we find that raising the assistance rate from 50% to 60% (a one standard deviation increase from the average) increases conversion by about 5 percentage points, corresponding to a 18.5% increase in transaction volume. Finally, we show that the approach developed in this work is useful to support store staffing decisions.