A combination of cloud and machine learning upgrades saved Walmart millions of dollars last year — but not without implementation hiccups along the way.
Automation can enable product recommendation, personalization, chatbots and fulfillment centers at retailers like Walmart, according to Mansi Kamdar, principal product manager at Walmart Labs, speaking at the Women in Tech Summit on Wednesday.
Over the next two years, 53% of retail business leaders predict AI will have the greatest impact on customer intelligence, 50% expect it to impact inventory management and 49% say it will effect chatbots for customer services, according to a KPMG survey of 950 business decision-makers, including 150 from the retail space.
As more businesses turn to automation to realize business value, retail's wide variety of ML use cases can provide insights into how to overcome challenges associated with the technology. The goal should be trying to solve a problem by using ML as a tool to get there, Kamdar said.
For example, Walmart uses a ML model to optimize the timing and pricing of markdowns, and to examine real estate data to find places to cut costs, according to executives on an earnings call in February.
Using ML as a tool, Kamdar outlined the top challenges of implementing automated solutions and how Walmart overcame them in product development:
Challenge 1: Automation takes over humans
It's generally a misconception that ML and other forms of automation are coming to take over existing jobs. Because ML is a tool and not a solution, humans are still needed to guide the process, according to Kamdar.
The technology lightens the workload on humans while removing "a lot of the grunt work that could be done by a much more efficient, optimized system," Kamdar said. Automating part of the process to stock shelves, for example, frees up human resources to consider how stocking methods best serve the customer instead of focusing on menial processes
Repurposing existing employees — instead of eliminating their roles — can create an environment where automation and humans coexist.
Challenge 2: Automation comes with bias
Automation itself isn't necessarily biased, but bias can slip into the data fed to models. "It's extremely important for us that we have that human intervention to remove the bias from the data that we have," Kamdar said.
Eighty-seven percent of data professionals are concerned about bias producing discriminatory results in AI, according to the Alation State of Data Culture Report surveying 300 data and analytics leaders. Among professionals that have deployed AI, better modeling skills, cataloging and crowdsourcing are listed as the top ways to address biases.
Kamdar says eliminating bias is part of the retailer's cultural approach to ML where employees are trained to consider bias as they work with data. "It's an iterative process of learning where we need to go in consciously and remove those biases," Kamdar said.
Challenge 3: Predictions become less accurate over time
As ML processes are integrated over time, employees will need to monitor for shifts in accuracy, according to Kamdar. Data evolves and the environment changes, so teams track success by being receptive to feedback and improvement.
Kamdar recommends reinforced learning where teams consider data discrepancies and iterate along the way. Questions to consider include:
Are we learning from that feedback?
Are we reinforcing the learning with the changes in data?
How are we going to assess that this is truly learning in the right direction?
During the pandemic, many data sources underwent radical change to account for disruption. Companies had to retrain machine learning models to adapt, specifically to build more empathetic responses.