When retailers implement artificial intelligence (AI) to power merchandising decisions, many fall into the pit of killing a product too early because they focus too much on margin or revenue instead of looking at the right measurements that determine early product success.
At the Ai4 Retail, Supply Chain and Marketing Summit, Elena Parfenti, director of scaled analytics at Nike, said retailers must lay out what objective they are trying to achieve ahead of measuring, whether optimizing for revenue, margin or some other metric entirely.
From there, Parfenti recommended that retailers leveraging AI should design product KPIs (key performance metrics) that speak to the quality of forecast, and bridge them with business KPIs such as sell-through or weeks of supply that are related to the product’s objective.
“Only after you bridge over the product KPIs to business KPIs, can you actually make the connection with your financial KPIs (margin/profit, EBIT or revenue), and the way you are translating those KPIs into each other has to be designed well before you go live,” Parfenti said.
The ROI for the “right” AI-based decision is generally higher the earlier it is implemented in the retail cycle, starting almost immediately at the planning stage, which lasts four to 18 months, Parfenti said. This goes into the buying and allocation stages of the cycle, but the impact starts to lessen once it reaches the in-season pricing stage.
“Later on, the decisions are less consequential because you are reshuffling the inventory you have already committed to buying,” Parfenti said. “Although the benefits are not so high, the later stages of the retail cycle are much friendlier to analytics implementation and machine learning, because they don’t need so much time for forecasting. When you are in season and you have days, weeks or seconds to say what the right decision should be, your forecasting accuracy will be very high.”
Optimizing a product launch with AI technology comes down to three steps, according to Parfenti. The first is finding a “friendly” AI problem that can be solved and useful for the company, while the second is designing the product and implementation that is actually compatible with the business model. The last step, measuring all the way through the end of product launch, is too frequently ignored and seen as too much of an afterthought.
Parfenti highlighted an “unfriendly” example from her time at predictive analytics company Celect (since acquired by Nike), in which one client wanted to know in June what sizes of Manolo Blahnik stilettos should be kept in each of its stores—nine months in advance.
While Celect had the historical data from March through the start of the program, it had no algorithm to accurately predict sales that far out because it only had eight major observations to go on.
“We recommended to our clients at the time, ‘Can we rethink the problem?’ Maybe the right question to ask is, ‘How many pairs of shoes do you need for North America for the month of March, April and May next year when your season starts?’” Parfenti said. “But let’s start with a rough understanding what that volume is, because with such a question, we can aggregate all the historical data of sales, and then answer a higher-level question of high granularity with more accuracy.”
From there, they built a framework on two dimensions: lead time before sales and level of granularity into an area of sales. This means instead of predicting stiletto sales across North America, they would prioritize predicting specific shoe sales in the Florida store on a specific day on a specific month.
“The higher the granularity, the friendlier the model it is,” Parfenti said. On the other hand, the lower the granularity, the more necessary it is to be closer to the decision. “A classic question here would be, ‘How much should I discount product X store-wide tomorrow?” That’s a great question to ask.”
In making the AI implementation compatible with a retail business model, Parfenti cited how a retailer wanted to determine the makeup of three men’s shoe categories in one store. One problem persisted: three different managers operated the men’s running footwear, men’s tennis footwear and men’s fashion sneaker categories under their own silos, thus defeating the purpose of any algorithm that was being implemented.
“You don’t want to keep those types of shoes siloed,” Parfenti said. “You want to open up the store and allow the algorithm to decide what share should be increased based on demand signals.”
Upon taking away the silos, Celect simply determined that tennis footwear required more space because demand asked for it.
For the third step, Parfenti noted that the key to a successful measurement is designing it before going live with the product, instead of waiting months to ensure that it works.
“Measurements should receive just as much attention during the product build as any optimization or process redesign,” Parfenti said.
To measure the impact of the decisions made through the AI, the retailer must seek out the percentage of AI recommendations that were executed so it can differentiate which decisions were controlled and those that weren’t.
Through the process, retailers must keep an eye on what constraints beyond the product itself are affecting end results, making it crucial to build a data infrastructure that monitors the product.
Walmart’s AI strategy improves long-term sourcing, saves productivity and money
During the event, Walmart’s Travis Johnson shared insights into how the retail giant is using AI to provide better sourcing insights to build long-term category strategies. Johnson, Walmart’s director of procurement technology, noted that category management is a major AI-driven initiative that is already automating administrative tasks and leveraging data to generate optimal category strategies.
In partnership with a third party, Walmart developed a patented AI Guided Strategy Creation technology, which enables sourcing professionals to efficiently develop, validate and implement category strategies, already seeing 60 percent productivity gains versus a normal manual process. With approximately 560 category strategies being developed across all markets at 45 manhours eliminated per strategy, 23,000 manhours are “saved” overall.
“We believe that we can see upwards of about $6 million in savings as we redirect or repurpose the time that normally would have been spent within a normal category strategy,” Johnson said in the session. “Leveraging this AI solution, we’re able to automate and drive efficiencies there, and then repurpose that time for other more strategic initiatives that will generate an increased savings.”
In Canada, Walmart is leveraging chatbots to conduct automated negotiations with hundreds of long-tail suppliers, with the AI platform identifying and accepting accurate “successful” negotiated payment terms.
Thus far, 20 percent of confirmed suppliers have reached a deal, with Walmart Canada calculating a 4x ROI in the process. Ahead of the negotiation process, Walmart identifies acceptable tradable terms upfront such as what the supplier could agree on and what Walmart is willing to pay.
“We define those contract parameters, what the value is for each of those contracts, and then allow the tool, the AI capability to look at the most optimal scenario for Walmart in order to achieve what we see as generating the appropriate value and return on investment,” Johnson said. “The value is generated based on the autonomous negotiations that take place with the supplier. We send a link out to them, we communicate with suppliers beforehand as to what’s occurring and why are we doing it, and then the suppliers will engage with the bot.”
Looking ahead, Walmart sees future considerations for AI on important more procurement initiatives, including spot sourcing, supplier optimization, geographical leverage, dynamic discounting and e-auctions.
StyleSage CEO: AI a necessity for post-Covid A/B testing
As apparel retailers look to leverage AI solutions out of the pandemic, they must better grasp shopper habits and trends so that they quickly get the right products on the first and second pages of their on-site search results, according to StyleSage CEO Jade Huang.
StyleSage, an analytics platform that analyzes data across 3,000 retailers and 90,000 brands to visualize e-commerce activity that fuels “in-and-next season decisions,” is currently helping apparel clients adapt to the post-vaccine world by using Google search trends to tailor recommendations to their upcoming assortments.
“You need some of those clues in order to know what to A/B test on their digital shelf,” Huang said. “When you have these massive amounts of data, having the machine learning capabilities is necessary to surface up the right insights to figure out what we’re going to test and invest in in our inventory. For global retailers that are in varying stages of lockdown, it’s important for them to understand: for each market, are these data points sharing something slightly different? This enables us to develop a more localized strategy for these countries.”