Fighting Retail Fraud with Personalization and Classification AI Tools – with Experts from Instacart, Etsy, and Gap Inc.

Sharon Moran

Sharon is a former Senior Functional Analyst at a major global consulting firm. She now focuses on the data pre-processing stage of the machine learning pipeline for LLMs. She also has prior experience as a machine learning engineer customizing OCR models for a learning platform in the EdTech space.

Fighting Retail Fraud with Personalization and Classification AI Tools@2x-min (2)

This article is sponsored by Riskified, and was written, edited and published in alignment with our transparent Emerj sponsored content guidelines.

Identifying fraud is a high priority for the retail and eCommerce industry. That’s because fraud is a costly problem. In fact, when looking at apparel specifically, the average cost of return fraud rivals that of shoplifting. 

Online retailers encounter challenges when trying to implement policies to combat fraud. In their efforts to uncover more insights about their best customers, merchants simultaneously end up surfacing customer accounts responsible for fraudulent activity. If merchants implement policies that are too restrictive, they risk losing customers, which is ultimately bad for their bottom line. Therefore, they will always need to make exceptions for good customers.

The COVID pandemic helped to accelerate trends in criminal activity that were already in place before 2020. In fact, according to one fraud study, the true cost of fraud has risen 19.8% since 2019. There is a lot more to retail crime under the surface. The connection between organized retail crime and drug smuggling has been clearly established. Lawmakers are keenly aware of retail crime as a money-laundering scheme for drug cartels.

In the following analysis of conversations with leaders in the retail industry on Emerj’s ‘AI in Business’ podcast, we look at the challenges their companies face when they try to use AI to increase personalization for their customers and segment customers to identify fraudulent activities to reduce or eliminate them. 

The conversations highlighted below were part of a sponsored podcast series focusing on customer personalization and fraud detection. The guests interviewed included:

  • Eyal Raab, Vice President of Business Development and Sales for Riskified
  • Asha Sharma, Chief Operations Office at Instacart
  • Chris Nelson, Senior Vice President of Asset Protection at Gap Inc.
  • Gerald van den Berg, VP of Analytics and Strategic Finance at Etsy. 

Each guest brought industry experience to the conversations, particularly on how their companies use AI to combat retail fraud and improve customer personalization.

Throughout the four episodes featured below, these industry leaders emphasize:

  • The importance of providing customers personalized experiences when shopping online
  • The role of AI and machine learning in fraud detection
  • The overlap between improving customer personalization and identifying fraudulent accounts

This article further focuses on four key insights from their conversations:

  • Segmenting customers to apply policies appropriately and solve merchant problems: Using machine learning and data collection to assist retailers with modifying delivery approaches in areas where good customers are affected by fraudulent actors at a high rate.
  • Using just-in-time capabilities to determine fraud intervention rate: Machine learning allows online retailers to detect real-time consumer behavior patterns and enable merchants to intervene at a rate that is impossible in brick-and-mortar stores.
  • Combining link analysis and AI to balance restrictive policies: Using machine learning-based anomaly detection systems with link analysis to identify scams and isolate bad actors.
  • Sorting unstructured data inherent in a creative marketplace: Leveraging data analytics to integrate human judgment in workflows to understand and label patterns in the data to inform the machine learning feedback loop.

Segmenting Customers to Apply Policies Appropriately and Solve Merchant Problems

Episode 1 – Fighting Fraud in Retail and eCommerce

When considering online policies, Eyal mentioned that there are three key groups to consider:

  • Merchants whose goal is to sell as much as possible and produce revenue
  • Consumers whose goal is to purchase what they are looking for
  • Fraudulent actors who impersonate consumers and create a lot of noise in the system

Eyal highlighted a particular challenge in implementing policies where you might have good customers who live in bad areas. He mentioned the role that an accurate base layer developed by machine learning has in these solutions. Eyal warns merchants can worsen the problem if the base layer isn’t accurate. A good customer can be denied their refund when a package is stolen by someone else, and bad customers can slip under the radar and continue their fraudulent activities.

There is a high degree of nuance present when segmenting customers. Ultimately, though, segregating the abusers allows Riskified to solve more problems for the merchants. Merchants must consider what policies they want to apply to good and bad customers. For those good customers who live in bad areas, merchants can also use solutions that involve altering shipping patterns. Modern-day abusers are very sophisticated in their approach. When discussing how bad actors exploit the system, Eyal explained, “They reverse engineer policies, and they create multiple accounts.”

Using Just-in-Time Capabilities to Determine Fraud Intervention Rate

Episode 2 – What Artificial Intelligence Means for Retail

Emerj CEO Daniel Faggella recently talked with Asha Sharma, Chief Operations Office at Instacart. They discussed data’s role in Instacart’s solutions to the post-COVID challenges they face.

Without even considering the issue of fraud detection, the grocery space is simply a challenging business. Asha highlighted some of those challenges:

  • Grocery stores are large. (48,000 sq. ft. on average)
  • The sector experiences a disaggregated supply chain
  • Tens of thousands of SKUs
  • Razor-thin operating margins

Asha further comments on the challenges merchants in the grocery sector face: “The challenge of personalization is not only knowing the customer but enriching the experience so that it’s as good as their in-store experience or complementary.” 

She explains that it’s both a massive challenge and a massive opportunity for Instacart to meet those expectations. Still, it requires a new approach to thinking differently about fraud, personalization, and even inspiration. For Instacart, customer personalization and fraud detection are largely interwoven. Asha views the existence of fraudulent activity as an opportunity for Instacart. She recognizes that it’s not possible to change human behavior. As a result, Instacart’s approach is to use AI for fraud and loss prevention, intervening before it happens.

Combining Link Analysis and AI to Balance Restrictive Policies

Episode 3 – Retail Fraud & Loss Prevention in Data, Brick-and-Mortar and Beyond

Emerj Senior Editor Matthew DeMello sat down with Chris Nelson, Senior Vice President of Asset Protection at Gap Inc., to discuss the biggest challenges for retail and eCommerce in a post-COVID world. Chris recognizes that there have been both bad acts and bad actors throughout human history. However, he explains the difference in degree and its impact on businesses.

Christ mentioned that CEOs of major companies now identify loss as having a material impact on their profitability. One of the reasons Chris cited is that criminals previously focused on items with higher resale value, but now their efforts are across the board. Some factors that fuel fraudulent activity are social media’s popularity and online marketplaces’ existence. These online sites make it relatively easy for criminals to sell products at reasonable prices. Chris also mentions that for criminals who focus on the digital space, it’s relatively easy for them to get identification online to engage in fraudulent activity.

He also clearly recognized that it’s difficult for retailers to prosecute fraud cases. There are a few reasons for this:

  • Recent legislation has removed the consequences of focusing on crimes against people.
  • Criminals understand the chance of experiencing severe consequences is greatly reduced. They know the chance of getting prosecuted is a subset of the probability of getting caught and realize the odds are in their favor.

Trying to stop all fraud is not practical. Chris recognizes that companies still need to consider commerce when considering restrictions. He explains, “We pretty much could stop all fraud, but we would crush sales.”

Stopping fraud also requires more sophistication than in the past. Chris added, “We work through this philosophy that you must combine physical security, data security and virtual security.” The focus, though, needs to be on data. Chris continues, “We’re in a data-led world. Most scams have a data component to them.”

Part of Gap Inc.’s ability to identify and isolate bad customers involves using link analysis to hone in on bad actors. Link analysis helps his company connect different customers, whether connected by the same vehicle or seen together. The analysis is critical since many fraudulent activities are related to big criminal rings.

Gap is focused on making the most of the data. He is adamant that data is king and adds, “Intelligence is information designed for action.” Chris recognizes that machine learning can tell his company patterns they can’t see. It can tell his company where the connections are and what is the commonality between transactions.

Chris continues to tell Emerj that Gap has used this capability to investigate store shortages and identify the indicators present beforehand. They use machine learning to tell them what is happening with other KPIs in stores with a higher shortage rate. They can then identify whether it is a service equation issue, an issue with the shipment, or a problem with allocation. AI can show unusual indicators when the shortage is also high, though Chris. He recognizes the limitation, “It may not be the smoking gun, so to speak, but it can start to lean me again in the right direction.”

AI allows his team to look at a broader set of data and look at questions they would be unlikely to ask otherwise. He elaborates, “AI gives me the ability as a leader to have the team look at a broader set of data and start to answer questions I can’t even ask yet.” 

Sorting Unstructured Data Inherent in a Creative Marketplace

Episode 4 – Striking a Balance in Personalization and Fraud Detection

Recently Emerj Senior Editor sat down with Gerald van den Berg, VP of Analytics and Strategic Finance at Etsy, to discuss the role of data in personalization. They also discussed some of the challenges that Etsy is encountering in its specific sector of retail.

Customer personalization presents a unique challenge for Etsy because they are a creative marketplace. Etsy lists over a hundred million items, but unlike traditional retailers, Etsy doesn’t have UPCs or SKUs for its products. As a result, their data is quite unstructured, which poses a problem in understanding their inventory at a deeper level. When requesting information from their sellers, Etsy tries to reduce friction for the seller but also ensures they get the information they need to match products to buyers. Etsy requires certain information but makes other information optional. Ultimately, though, Etsy asks for the following:

  • Category Data
  • Taxonomy data
  • Descriptions of the items

Etsy does try to infer some information, particularly for things that might be harder to ask their sellers directly. Gerald mentioned that this is where machine learning and related tools can be constructive. While they want to make it easy for sellers to list items on Etsy, they benefit when they understand more about the sellers and their listings.

“Typically, our sellers are good and trying to do the right thing,” Gerald explains. “And so finding the bad actors among them is a little bit of a needle in a haystack type of problem.”

Gerald explains how Etsy views the concept of personalization. “When I think of personalization in the context of Etsy, I tend to think more from a buyer perspective, meaning ‘How are we matching the inventory that we have to a buyer’s needs?” 

He affirms that machine learning, data, and analytics tools are very applicable to the fraud problem and help Etsy identify which sellers they want to apply controls to and help them understand what those restrictions would look like.

Subscribe
subscribe-image
Stay Ahead of the Machine Learning Curve

Join over 20,000 AI-focused business leaders and receive our latest AI research and trends delivered weekly.

Thanks for subscribing to the Emerj "AI Advantage" newsletter, check your email inbox for confirmation.