“How AI is changing the game.”
“Why AI is the future.”
“Unleash the power of AI.”
We’ve all seen the headlines — in the news, at conferences, and from essentially every vendor in the tech space, ever. But with all of the hype around AI and machine learning, it can be difficult to ground the discussion in reality.
What specific business problems is AI helping to solve? Where is AI being used to drive measurable results? Not just in theory, but in practice. In other words…how can artificial intelligence move out of the lab and into the market?
Supply Chain Optimization — The New Imperative
Retail is one industry in which machine learning is being successfully applied to solve discrete business problems. From traditional, multi-channel retailers, to emerging, digitally native brands, the industry is undergoing massive transformation. Retailers are accelerating investment in innovative technologies that can drive sales, preserve margins, introduce efficiencies, or lend competitive advantage.
In enterprise retail, machine learning has traditionally been applied to areas like search, chat, advertising, pricing, personalization, and product recommendations in “front-end” (or customer-facing) systems. More recently, retailers have been applying machine learning to improve decision-making upstream of the customer, across the supply chain — in back-end systems and across business-critical functions like demand forecasting, planning and allocation, inventory management, order routing, delivery date estimation, fraud detection, and returns.
Why is supply chain optimization suddenly becoming a greater priority to retailers?
It’s clear that shoppers’ expectations have never been higher. Free same-day delivery is fast becoming “need to have” instead of “nice to have.” Customers expect to easily find the product they want, when they want it, wherever they are. They have little to no tolerance for out-of-stock items, long delivery times, or a poor user experience.
As retailers strive to keep up with increasing customer demands, the supply chain has grown more complex — and more critical. It is now a core component of corporate strategy, and retail executives are viewing the supply chain through a lens of value creation rather than cost-minimization.
Over the past decade, retailers have been investing in enabling the “omnichannel experience,” empowering the customer to find and purchase items on any channel — web, mobile, stores, marketplaces, social media — and with the same brand experience. But while the customer’s experience feels “seamlessly” connected, the legacy supply chain systems that enable that experience remain siloed and brittle. The network-wide transformation required to create a true omnichannel experience puts substantial burden on the entire order lifecycle — from demand forecasting and buying decisions, to inventory assortment, allocation and availability, to expected delivery dates and last mile logistics.
A key component of an omnichannel strategy for retailers with a brick-and-mortar presence is store-based fulfillment. According to the 2019 Internet Retailer Top 500 report, 51.8% of retail chains in the top 500 now offer in-store pickup, a capability that drives in-store traffic and increases average basket size. Using stores as fulfillment centers improves shipping times by leveraging store proximity to customers, expands assortment online by tapping into store inventory, and offers convenience and flexibility to shoppers who want to buy items online and pick them up in store.
A successful store-based fulfillment model requires strategic investment in people, processes, and technology to bring a true decentralized multi-node distribution network to life. For most retailers, this means transitioning from a traditional hub-and-spoke model to leveraging hundreds or thousands of interconnected mini-FCs and micro-DCs from which ordered items can be sourced. While an expanded network creates new opportunities to evolve the customer experience and capture revenue, it also introduces vast complexity into the supply chain.
Machine Learning Use Cases — Inventory, Fulfillment, & Markdowns
This combination of market factors — investment in omnichannel capabilities, evolving distribution networks, and the maturing of store-based fulfillment models — has generated copious amounts of data across the supply chain, creating fertile ground for machine learning. At the same time, the availability of cloud computing and parallel processing has made it possible to optimize and execute learned decisions at web scale and speed.
I’ve been fortunate to be part of a wide variety of discussions about AI, some in the clouds and some rooted in reality. Below are three examples where I see machine learning being used to drive better decision-making across the commerce supply chain.
Example 1: Inventory Availability
While transforming stores into fulfillment centers is a critical part of omnichannel strategy, the actual implementation reveals just how differently stores operate from fulfillment centers.
- Operations: Distribution centers are highly controlled, organized environments, operating with high inventory accuracy and predictable, repeatable processes. Conversely, stores are inherently variable environments, impacted by real-time sales, human error, shrink, cross-store operational differences, and a host of other externalities.
- Inventory: Offering store inventory to online shoppers may “save the sale” and unlock additional demand. However, systemic in-store inventory inaccuracy makes pinpointing exactly which units are available to promise — by item, by location, and by purchasing channel, at any given point in time — very difficult.
- Findability: The ability of a store associate to actually find an item in store varies by a multitude of factors — from product attributes like size, color, style, vendor, or price, to operational factors like store backlog and historical pick-rate, to demand signals like traffic patterns and item sell-through rate. The “findability” of items also dynamically changes depending on the time of day, day of week, promotional calendar, and season.
These fundamental differences mean retailers have to manage a trade-off when choosing how to expose store inventory online:
Expose as much inventory from stores as possible to maximize demand and increase assortment…but risk canceling items that aren’t available after they’re promised to the customer?
Or, expose only what has a high likelihood of being available in stores…but potentially miss out on demand and inhibit the success of an omnichannel program you’ve invested a lot of time and resources into?
Taking a page from modern portfolio theory (often used by hedge funds to determine the optimal mix of investments), retailers can leverage machine learning to determine the optimal mix of inventory to make available in order to balance risk (cancels) and reward (demand). This is the Efficient Frontier of inventory management, and a key source of competitive advantage for retailers who get it right.
To summarize, this amount of variability in the store network and uncertainty regarding outcomes — coupled with massive volumes of data — is a perfect use case for applying AI. Using historical and real-time data, retailers can leverage machine learning algorithms to:
- Find patterns and identify which inputs correlate to in-store fulfillment rate
- Estimate the probability of an item being in demand and successfully picked at a given location, at a particular point in time
- Predict the optimal inventory levels to make available for purchase by channel and then serve those recommendations to a variety of operational systems, like order management or enterprise resource planning
Example 2: Split Shipment Reduction
So, our Frontier-balancing, AI-savvy retailer has determined where to allocate inventory and what to make available for purchase online. Now — where does the retailer source orders from across an expansive, decentralized fulfillment network?
Fulfillment is a meaningful source of costs in the retail supply chain. When an order is placed on the website or mobile app, a decision needs to be made regarding which “node” (i.e., location — think store, fulfillment hub, distribution center) to fulfill it from, and how. Ideally, a retailer wants to source the items in an order from a single node, since splitting the order and sourcing items from multiple locations increases shipping and packaging costs.
But how do we know whether a single location can successfully pick, pack, and ship all of the items in the order, in the time required to meet the customer promise date? And, assuming there are multiple nodes that could meet the delivery target, which one is the most profitable location to source from?
This is where web-scale cost computation and machine learning come in. To make a node assignment decision today, most retailers rely on the simple heuristics built into an order management system’s logic. Sequential rules are used to arrive at a decision, with each rule narrowing the options in a “first-then” format (i.e., first minimize the number of shipments, then ship closest to the customer).
The limitation of this approach is that filtered nodes can’t be considered against future, competing factors. For example, a rule that requires shipping from the store closest to the customer might end up sourcing from a store that has a poor pick-rate and thus higher labor costs. Or, a rule that requires shipping in a single package might source from a store that can pick all of the items in the order but has to pull from full-price inventory to do so. Instead, it may be more cost effective to split the order across two stores that can source from stale or marked-down inventory.
An algorithm is better equipped to:
- Optimize for multiple objectives simultaneously
- Calculate the total estimated cost to serve from any given node
- Consider all possible permutations with which to fulfill the order
- Apply constraints such as delivery date or store capacity
Machine learning can predict the propensity of a location to split an order, the partial fill-rate by location, and the cost to ship across carriers, rates, dimensional weight, and transit times — all before an order is allocated.
Intelligent, data-driven order routing can achieve a reduction in split rate and overall shipping costs, without abdicating the opportunity to optimize for other objectives, like speed to customer or labor efficiency.
Example 3: Markdown Optimization
According to Deloitte’s Global Powers of Retailing 2019, the retail industry has a composite net profit margin of 2.3%. Coupled with increasing competition from digitally native brands to offer expanded assortment and competitive pricing, retailers are feeling the squeeze from shareholders to maximize profits, while still investing in e-commerce and omnichannel capabilities. Retail executives are under constant pressure to reduce COGS and operating costs, while delivering goods faster and cheaper.
Avoiding unnecessary markdowns, as well as selling through stale inventory to make room for newer, full-price items, is paramount to preserving margin. Holding costs are high, especially for fast fashion and retailers with seasonal merchandise, and inventory must either turn or be discounted dramatically — or even liquidated.
In an ideal scenario, retailers would perfectly match supply with demand and purchase the right products in the right quantities, by channel, and never under- or over-allocate inventory. But in real life, this is still more of an art than a science, and despite best efforts, gaps exist that create out-of-stocks and lost demand, or deep discounting due to the cost of held capital.
Optimizing markdowns is one way to mitigate the pain of over-purchasing. There are many factors that impact markdown amount and cadence — from pre-season planning to in-season replenishment. But when markdown strategy isn’t quite perfect, routing orders to the right location at the right time is key to averting additional profit erosion. All else equal, why not fulfill an online order using items already in markdown or expected to be in markdown, rather than from items at full price (or expected to be sold at full price) in store?
Preserving margin by circumventing markdowns or sourcing from stale inventory requires an as-accurate-as-possible view of demand for both in-store and online merchandise. In addition to forecasting demand by channel, retailers must also be able to predict the expected revenue for the last unit sold in a given location, and take into consideration the future selling price at the time of anticipated in-store stock-out.
Forecasting demand is notoriously difficult, but improvements in open source algorithms and massive amounts of historical time-series data are enabling better, faster predictions. In the past, a major challenge to generating these types of forecasts was collecting the right data, getting it into the right format, and processing it in a timely and cost-effective manner. Today, a host of tools and technologies exist to facilitate data ingestion, engineering, validation, labeling, tuning, and processing — from data lakes and ETL operations, to automated services that enable data scientists to build, train and deploy models in minutes.
Getting clean training data into the algorithm is the first step. From there, the model can estimate the future demand and margin potential of an item at a given location, trade it off against other possible locations in the network, and choose the location with the lowest opportunity cost. Applying AI to markdown optimization can unlock substantial margin improvement, at a time when retailers need to maximize profit on every dollar.
Change or Fail
In retail, we can see how AI is able to move from the theoretical to the practical, resulting in tangible business impact — driving sales, preserving margins, introducing efficiencies, and lending competitive advantage. Machine learning drives better decision making across the end-to-end supply chain, and has the mutual benefits of both improving financial outcomes for the business, as well as enhancing the customer experience.
In today’s constantly evolving, highly disruptive market, retailers must be willing to embrace innovation and endorse cutting-edge technologies like AI. To gain both adoption and acceptance, AI needs to be applied in real life to real problems that deliver real results. But ultimately, business is conducted by people who have to make critical decisions in an environment ripe with volatility, ambiguity, and uncertainty. As Art Peck, former CEO of Gap Inc., put it at last year’s industry conference Shoptalk, the choice facing retailers is: “Change or fail.” Leveraging AI to make smarter, data-driven decisions isn’t just good business — it’s paramount to survival.