Skip to main content

How Amazon’s Pandemic Planning Drove $500M in Merchandising Decisions

When people think of efficiency, Amazon is often the first company that comes to mind. Behind its vast delivery network of fulfillment centers and delivery trucks is a demand planning operation that enabled the e-commerce giant to adapt to rapid consumer shifts early on in the Covid-19 pandemic despite high volatility and extreme uncertainty.

At the AI4 Retail, Supply Chain & Marketing Summit, Kelsey Conophy, senior technical product manager, Core AI at Amazon, revealed that over the course of the first six months of the pandemic, her team drove more than $500 million in buying decisions based on the forecasting methodology they developed, increasing Amazon’s allocation of key products by 40 percent after the first Covid wave.

Although the coronavirus pandemic is an unprecedented event, Amazon still sought out historical data that could provide insights into supply chain demand shifts, including 2008 recession data, country-level data since the global start of Covid, past unexpected weather events and even tax refunds. Throughout the pandemic, the team also integrated external data sets including Covid infection rates and replication patterns, mobility data, economic impact projections and government lockdowns, among other factors.

From there, the team identified opportunities to expand and adapt research into product availability correction, demand spike detection and demand scenario production, so that they could differentiate between the present experience versus a traditional seasonal pattern.

Related Stories

“In March and April, we have a lot of customers buying typically spring seasonal lawn and garden products, and those are lots of seasonal products that we didn’t necessarily want to have to act on,” Conophy said. “How do you develop the distinction between those seasonal versus cases that are unexpectedly spiking?”

During the research process, the team focused on rapid iterations to accelerate their development, creating and launching their Covid forecasting methodology in less than two weeks.

“Based on our conversations with our customers and our partners, we identified that the key was to be able to provide visibility at an aggregate product-group level, meaning: How much toilet paper demand do we have? How much hand sanitizer demand do we have? As opposed to: how much two-ply Cottonelle four-pack demand do we have?” Conophy said.

Conophy said the adjustments within the different methodology also covered different demand scenarios, such as whether there would more Covid waves or new lockdowns, or what would happen if some states returned to in-person schooling.

Since the team developed a new forecast not just for the pandemic, but also encompassing the next four years, this methodology also informed more than $50 million in strategic long-term investments, according to Conophy.

“The most exciting thing was that we really increased our ability to allow Amazon and our customers and our partners to really understand and adapt to unexpected events, which was something we really didn’t have the capability to do before,” Conophy said.

What apparel retailers can learn from satellite data

In another session at the AI4 event, Niels Wielaard, CEO of Satelligence, stressed the importance of AI in giving consumers traceable insights into whether the products they buy negatively impact the environment and social surroundings.

While Satelligence specifically leverages risk and performance satellite data analytics to understand agricultural risks to the supply chain such as deforestation, forest fires or drought and their impacts on CPG products, the lessons here can better inform sourcing strategies of apparel and fashion brands that are concerned about the locations of farms, mills or factories, or need proof of compliance.

“Despite the Covid pandemic, we saw last year that the need for sustainable products has only increased, and the efforts that a lot of consumer goods brands and retailers and growers are putting into getting better visibility into their supply chains has only increased,” said Wielaard.

Leveraging the satellite data, Satelligence can determine which forests have been cut down, for example, and cross-reference that data with where its own partner brands, suppliers, manufacturers and traders are sourcing their products, to build digital supply chain models that show which companies are at the greatest risk.

“We now have the data from the satellites to know where low-risk areas are, so when you’re sourcing from those areas now, we can now see which areas are okay and which areas are risky,” Wielaard said. “Our information to pinpoint the high-risk areas enables users to engage with suppliers and discuss what’s happening in the region. We’ve seen that this engagement leads to transformation and the stopping of deforestation within the supply chain.”

Two reasons AI can bring accurate delivery predictions

When used correctly, AI can be a practical solution across other supply chain areas, namely logistics. Dana von der Heide, founder and chief customer officer of Parcel Perform, identified that many of the shopper frustrations with parcel delivery companies like UPS or FedEx is that they only know where the package is, but not when it will be delivered.

Only five percent of 642 last-mile carriers analyzed by Parcel Perform actually provide shoppers with an estimated delivery date, von der Heide said in the session.

“[T]elling customers when the parcel will arrive on a tracking page or email notifications of course adds to the convenience of customers, but also reduces the amount of redeliveries of parcels and enhances the brand environment with your customers and the retailers,” von der Heide said.

Von der Heide said that machine learning can solve this issue by accurately predicting when the parcel will arrive largely because logistics processes follow repeating patterns, and the defining data points are already available to use. In the case of Parcel Perform, the platform tracks 50 million packages per month across more than 600 carriers.

Parcel Perform needs two key sources to train its prediction engine, according to von der Heide: shipper/merchant data and tracking/carrier details. In one case study applying the Parcel Perform platform to a European marketplace, von der Heide noted that the platform was able to narrow down the right delivery date at checkout 88 percent of the time despite the marketplace’s work with 20 different carriers across markets.

Although the marketplace initially had an expected delivery window of three to seven business days prior to working with Parcel Perform, prediction accuracy jumped to 96 percent upon dispatch and 99 percent once the product was dispatched for delivery.