How Machine Learning Can Enhance Efficiency and Customer Experiences in the Retail Industry

by Rohit Kaila

While the massive surge in online shopping we experienced during the pandemic has ebbed in the last year, e-Commerce is still king, and delivering powerful customer experiences remains vital to cashing in. But not all retail categories are the same, and for some, delivering these powerful and personalized experiences is a significant challenge.

Take the home goods industry. Buying items for the home is very different from shopping for a new automobile, where many purchases are driven by the brand, or a six-pack of batteries, where the available options are limited. First and foremost, home purchases are highly personal and style based. Add to that the fact that customers are often not looking for a specific brand. They follow an “I’ll know it when I see” mentality.

This absence of search criteria presents a real “head scratcher” for retailers trying to figure out how to deliver a great customer experience, which is why more are turning to artificial intelligence (AI) and machine learning (ML). ML is a vital tool for online retailers because it can power through tasks with loads of relevant data and more. But it comes with a caveat—ML is not a “fix-all” solution. To maximize its potential, you should only leverage it under certain circumstances.

The top considerations for when to deploy ML should include the availability, accessibility, and quantity of data and the quality and relevancy of that data. Second, consider the problems that you want it to solve. Once you have a list in place, look for those items with the following qualities:

• They have a higher tolerance for forgiveness on “faulty” predictions.
• Offer a clear path to making improvements using feedback loops.

Problems that check these boxes are likely your best candidates for ML. I’ve gone through this exercise in three areas, keyword bidding, search, and personalization. In each instance, the volume and the relevancy of the data are “easily” accessible, and the inherent “noise” around predictions is “tolerable.”

Take search as an example. Let’s say a customer visits a retailer’s website looking for a sofa. They don’t have many search criteria to help narrow things down. In fact, all they have to include is the color blue. Using ML, the site ranks 50 sofas to show on the page, and after some review, the customer zeros in on number 5 and purchases the sofa.

Here is where predictability tolerance is vital—clearly, the ranking prediction was imperfect. Otherwise, the customer would have purchased the first sofa listed. But ML was able to sweep through a large selection of items in seconds and rank one high enough that the customer ended up purchasing. In this example, 5th was good enough of a prediction to secure the purchase.

When I speak with people about ML and this balance between risk tolerance and the extent of automation, I like to use the automotive world and autonomous driving as examples. For the industry, the goal is Full Driving Automation. This is an excellent example of full algorithmic automation in a high-risk application where the vehicle performs all driving tasks across all conditions, with zero human interaction/override required.

Next comes Control versus Suggestion. This is where the car makes the decision and does the driving. Once again, the risk is high and not tolerable because a single error can cause an accident and put the passenger’s safety at risk. Because of this, while there are controlled experiments going on in different cities in the world, it’s not available to the general public.

That leaves us with Partial Driving Automation. Here ML is applied in areas such as lane assistance, braking assistance, etc. In each case, it’s making suggestions but leaving it to the human to decide which ones to act on. Ultimately, the driver may determine that a warning isn’t valid, override it, and go about business. They may be slightly annoyed at the nuisance but no worse for the wear.

The example I provided earlier aligns with the partial automation example above. ML is optimized, and it makes suggestions that may help guide our actions. At the same time, we can tolerate and manage the risk if something goes wrong. Another area where this would apply is Ad bidding. In online retail, it’s not unusual for a business to enter millions of auctions every day. In each case, the goal is to submit a bid that results in the ad showing up high enough in Google Search to result in a customer purchase while optimizing the ad spend.

As with the search, the ad doesn’t need to rank number one. A win occurs when the purchases from those bids exceed the expected return on investment (ROI). If an ad misses the mark completely, teams can adjust and try again. In this example, the company’s marketing and ad teams can tolerate the risks inherent in these predictions.

For many online retail businesses, ML will be a critical ingredient to their success, but as in the automotive world, it’s not ready to take full control. Until then, remember that while ML can remove much of the heavy lifting, but your teams must find the right areas where it can help. I recommend stepping back and following the criteria outlined in this article. I’m sure you will find those areas where you can cash in with ML.

(The author is Rohit Kaila, Head of Technology and Site Leader Wayfair India TDC, and the views expressed in this article are his own)