Commentary

Leveraging machine learning to improve order fulfilment throughput

Prithwijit Mukherjee, global business manager at Wipro, discusses the impacts of machine learning related to BOPIS and identifies challenges retailers face and solutions to address the problems.

Photo by iStock.com

April 29, 2019

By Prithwijit Mukherjee, global business manager, Wipro

Buy online pickup in store or click-and-collect is now a key omnichannel initiative with retailers for driving in-store traffic by identifying potential customers online and luring them to buy other products once they arrive. Yet BOPIS order fulfillment operations remain a key challenge for retailers.

Once BOPIS orders are placed online by customers, some common issues faced by retailers for fulfilling these orders include:

Lack of Item availability: Items ordered are often not in the store and need to be procured from regional or central distribution centers, causing fulfillment delays.
Task prioritization and assignment: Associates need to perform a variety of tasks like selling using Point of Sale terminals, stock receiving and put-away, customer service tasks, etc. Hence, BOPIS tasks are often lost in the pipeline, leading to aging and ignorance.
Missing items at pick locations: Pick locations are either backroom or store shelves, from which items are often displaced by a customer or a fellow associate assigned to other tasks before the BOPIS items get picked. This results in items missing from their pick locations.

All these challenges lead to fulfillment delays, while customers expect quick pick-up cycles. Consider that 60% of U.S. online adults who use BOPIS expect orders to be ready for pick-up within two hours of placing the order. Delays lead to order cancellations, causing significant business impacts such as customer churn and loss of cross-selling opportunities at the store when the customer arrives for order pick-up.

To overcome these challenges and meet customer expectations of timely and perfect order fulfillment, retailers need to succeed across three banners:

Consumer experience: This is a highly complex parameter and is characterized by pick-up location of choice, in-store parking facilities, pickup window duration flexibility and is often influenced by behavioral traits like willingness to wait for perfect fulfillment, and willingness to accept substitutes for a timely fulfillment. Some customers don't prefer delays and may be willing to accept a substitute product when an ordered item is not available. However, others may be highly brand specific and will be willing to wait until the preferred brand is available.
Workforce availability: This is a key lever for meeting successful order fulfillments from the store. Store associates will always be busy with regular store operations, finding it extremely hard to have associates dedicated solely for BOPIS operations like picking, sorting and packing.
Product availability and sourcing: Inventory management and control are tricky in omnichannel. It requires sufficient data and training algorithms to accurately predict demand across different channels. It's very likely that the store selected for pick-up by the customer will not have all the products ordered. Hence it's imperative to have an appropriate sourcing strategy to tackle some common questions like which SKUs should be stocked in the store and how much? Which SKUs can be stocked in a central DC and which ones in regional DCs? Efficient planning across these dimensions will enable a retailer to have products ready for a timely pickup at the store.

Machine learning can bring about incremental benefits for retailers under each of the above banners by use of various prediction algorithms based on historical data. The objective of this article is to illustrate the use of machine learning in the banner "workforce availability" and improve operational throughput through workload reassignment and order prioritization in real time.

Solution approach

Before understanding how we can use machine learning, let us try to understand the various factors that impact BOPIS fulfillment throughput. This is measured by the time taken for an order to be ready for pick-up from the time the order was placed. Some common factors are –

Time of the year/season (month: This has a direct impact on the volume of BOPIS orders assigned to a store, directly impacting workload
Number of associates in the store: The higher the workforce, the higher the expected throughput
Time of day: Higher the store traffic, the higher the chances that associates are busy and orders not getting assigned
Retail store space: The larger the store, more time is taken by associates to complete the picking activity
Types of SKUs ordered by a customer: Higher the variety, more complex is the picking, packing and sorting which results in a longer cycle time
The average age of workforce: Elderly associates will have less productivity than younger associates
Number of open orders in the pipeline: The higher the number of orders not processed by store associates, higher is the pending workload

A retailer may want all orders to be ready for pick-up within two hours, against which they will measure operational throughput, i.e. the percentage or orders that were ready for pick-up within two hours. Supervised machine learning models like Logistic Regression or Naïve Bayes Classifier can be used to understand the impact of the above attributes on throughput and to predict if a new order, once received, will be ready for pick-up in two hours or not.

A BOPIS order can be assigned in real time if it's predicted that it will not be ready for pickup within two hours. Thus, once the model has been incorporated, these steps need to be taken:

Given a BOPIS order is assigned to a store, predict the likelihood that it will be ready for pick-up within two hours.
For low likelihood (less than 50%), assign to an associate with available capacity and notify associate.
For high likelihood, it is expected that the service level will be met and hence the order may be pushed to the usual task pipeline of associates.

Example:
Given store X receives an order with 3 SKUs at 3:30 p.m. on a Tuesday in April, with eight associates in the store, with the average age being 25 and given the store is medium. If the predictive model estimates a 35% chance that the order will be ready for pick-up within two hours, then this is a case of a low likelihood scenario which should be assigned to an associate with the available bandwidth to fulfill this order. This ensures that order processing starts and the assigned associate initiates picking and meets the service level.

Outcomes and benefits for the retailer

Retailers can benefit in the following ways from this solution

Orders are proactively assigned to associates if there are high chances of missing the service level, thereby ensuring customer satisfaction.
The insights from the model can aid in staffing strategy by determining the profile of associates required for successful service level in terms of associate age and headcount needed for specific seasons, months and days.

Thus, it must be understood that this kind of initiative will involve the following team mix:

Machine learning engineers and data scientists to develop the model.
Operations team to define the business problem.
Retail consultants to bring in a fresh perspective on how to approach the problem and identify all possible factors by working with the operations team.
A product development team to build the necessary software that notifies associates in real time based on the insights from the machine learning model.