Walmart watches the weather to decide what food is going to sell better. Predictions powered an 18 percent increase in sales by having more steaks in stock when it’s warm, dry, cloudy and windy, and beefing up burgers when it’s hotter and less windy. High temperatures with a light breeze sell more salads; clear sunny days sell more berries. But you don’t have to be a retail giant to use AI to improve your supply chain.
Predictive analytics and remote sensors tell distributors when a fridge needs restocking with soda or a coffee vending machine needs topping up (avoiding the fine that Mars Drinks levies when a product is out of stock). Lakeba’s Shelfie robots will soon cruise the aisle at UK supermarket Co-op using image recognition to detect when products are sold out or shelved in the wrong place.
An online fashion retailer who deals with hundreds of suppliers uses image recognition apps on smartphones to check exactly which style of blue dress arrived in a shipment so they know what’s in stock.
On the other hand, Jet.com is very keen to know what’s not in stock, Director of Engineering Scott Havens told CIO.com. “Suppliers might say they have ten items in stock but they only have eight,” which leads to unhappy customers who can’t place an order when they expected to, he says. If an item isn’t in stock, Havens doesn’t want the site to even show up in a search. “In terms of customer experience, not talking to a customer is better than talking to them and disappointing them.”
AI’s spread to every corner of industry
Havens’ team uses machine learning to track how accurate different suppliers are at reporting which products they have in stock; the numbers can be wrong because they don’t update them often enough or because they just don’t give accurate reports. The machine learning system has halved reject rates for those merchants, he says, giving Jet.com a better view of its supply chain.
Windshield replacement service Autoglass is using image recognition in its phone app so customers know whether the chip in their windscreen can be repaired or needs replacing. Customers like knowing what the cost will be, but Autoglass knows whether a store can do the replacement from stock or will need to order the windshield and schedule a fitting once it arrives. The app is already handling 2,500 images a week with over 80 percent accuracy and the company is looking into using it to assess whether advanced driver assistance systems will need recalibrating after a repair or replacement.
Car dealerships are turning to AI to decide which cars to stock and who to advertise them, Sam Mylrea, CEO of PureCars, told CIO. “Unmoved inventory sitting on the lot costs hundreds of dollars each month in interest. By using AI, dealers can better understand which cars to stock based on consumer behavior and past purchase patterns. Armed with these insights, they can use personalized marketing outreach to target the right customer at the right time.”
AI can help in agriculture and food processing. Chinese winery Great Wall Wine uses weather stations and pesticide monitors to power the predictive analytics that says when to treat the grapes and when they’re ready to harvest. Microsoft is developing an IoT precision agriculture system for smaller farmers that uses machine learning to predict when they need to water crops or spread lime to improve the pH level of the soil. Bühler’s LumoVision grain-sorting system removes nearly twice as much contaminated grain using image recognition as its conventional machinery did. Instrumenting breweries and dairies can increase production and reduce production time by processing incoming materials faster and predicting when a batch of beer or cheese is ready so you can use the same equipment for the next batch more quickly.
Logistics firm R.R. Donnelley built a machine learning model using weather, traffic predictions and records of past shipping jobs to create such accurate cost estimates that the cost of building the system was covered in the first month, with the company winning 4 percent more bids in the first quarter. And Toyota might soon give Amazon’s warehouse robots some competition; it’s looking at using AI to train “palette drones” that can recognize patterns, learn the layout and flow of a factory floor and work in a swarm that decides for itself which type of robot to use to move each load.
Making the engine of business smarter
What all these projects have in common is a changing view of one of the most fundamental areas of business. Inventory and supply chain has traditionally been seen as a cost center, Ammon Matsuda, national leader for data and analytics at KPMG told CIO.com, “but really it’s an engine that takes in orders and makes things work.”
Using AI to make the supply chain more nimble takes it even further from being a cost center to an enabling function, especially in today’s challenging market. “It’s key to being more responsive to volatility, to handling far more variations of products than in the past and to handling more channels than before,” Matsuda believes.
AI tools are already getting a lot of traction in inventory management and that’s only going to expand, he predicts. “Think of robots being able to maneuver and do manual labor for tangible inventory management.” That’s not just in the warehouse. “You’ll start to see robots dealing with a less structured environment like going through a store, grabbing products, sensing where replenishment is needed or if the product is not placed exactly on the shelf above the price and label, so you have to have machine learning to know which label applies to which gap.”
AI can help with inventory aging for short shelf-life products like fresh produce. Matsuda says that companies whose names you’d recognize are “looking at how to deploy machine learning capabilities to sense when a head of cabbage is turning and predicting how much life is left so they know where in the supply chain to put it to turn it before it goes bad.”
He’s more cautious about using AI to predict changes in demand, because consumer patterns can change unpredictably, and machine learning won’t necessarily cope with those changes. “With machine learning you need to have a pretty good history of demand to sense patterns and come up with reasonable and valid predictions so you can make supply chain adjustment. Machine learning is fantastic at replicating the consistent behavior of humans but it’s pretty terrible at reacting to net-new situations and observations, so machine learning is appropriate when we can infer what a human could do but we use machine learning models because of the challenges of scale or speed.”
The real advantage of machine learning in supply chain prediction is rather more prosaic than uncovering the secrets of human behavior, Matsuda suggests; “the benefit is essentially getting companies to make more logical decisions.”
Putting AI to use
Businesses that want to take advantage of AI in their supply chain should start by optimizing existing systems and then bring in new technologies such as computer vision over time, Lance Olson, partner director of program management for Microsoft’s Cloud AI Platform told CIO.com. Take the historical analytics you already do and add predictions: “What is our supply demand going to be next month not just based on historical data but on predictions?” Or use canned solutions like the tools in the Azure Gallery that use machine learning to perform supply chain optimization for oil and gas, or a retail system that optimizes the quantities of products and the schedule for deliveries to stores and warehouses.
That can give you quick improvements that give you confidence in the process, he suggests. “If you were running optimally you could save 5.47 percent of your cost as opposed to the way you’re actually running things today.” Then you can move on to bolder, more revolutionary projects that need deeper data science knowledge and might require sensors and other hardware investment like cameras with custom computer vision models to do real-time recognition, microphones or looking at the non-visible spectrum with thermal imaging, lidar or radar.
“If we use a different mechanism entirely for inspecting our goods as they come through from suppliers, we might be able to eliminate an entire step in our process or two steps or ten steps. I have customers who have built data models where they look at trucks as they come into the warehouse that have giant piles of pipes or other equipment on them, and the models count the inventory as it comes in. That dynamic accounting and re-accounting every time the machinery is moving through the system enables them to short-circuit a whole bunch of other downstream issues that they might be trying to optimize for,” Olson says.
For industrial machinery, audio processing can help assess inventory during manufacturing and when parts arrive. “If you have anything that emits a sound, like a motor, and you’ve got microphones at the edge of the manufacturing line you can listen to the sound of the motor and understand the quality of that part,” Olson explains. “If you’re an automobile or airplane manufacturer and you’ve got engines coming in, as soon as you can run the engine you can listen to the engine as it comes in and do quality control on parts coming, and that might allow you to change the way you do manual inspections.” The same techniques could help food distributors and retailers monitor refrigeration units to predict failures in advance.
Olson agrees that AI will become a crucial part of inventory management as businesses build up their expertise and create custom monitoring and prediction systems. “It becomes a large scale system of intelligence where you’ve got a myriad of factors that you’re looking at and you’ve data pipelines flowing through the whole system. For any company that depends on inventory and supply chain, their sustainable differentiation over time is likely to include that as part of their core. How well can they optimize the system; how can they lower costs, how can they get the right product to the right customer at the right time, at the right quality and the right price.”