Many companies seem eager to leverage artificial intelligence and machine learning capabilities, if for no other reason than to be able to let their employees, customers, and business partners know that they’re on the leading edge of technology progress.
At the same time, a lot of businesses are looking to enhance the experiences of customers and channel partners, in order to increase brand loyalty, boost sales, and gain market share—among other reasons.
Some have found a way to combine these goals, using AI-powered tools to improve the way they deliver products, services, and support to their clients and business partners. Here are two examples.
G&J Pepsi: Predicting stores’ product needs
G&J Pepsi-Cola Bottlers began its foray into AI and machine learning in January 2020, when it partnered with Microsoft to better understand the AI and machine learning components within Microsoft’s Azure cloud platform.
With guidance from Microsoft’s data science team, “we spent time understanding the environment, required skill sets, and began ingesting various data components within Azure ML to provide predicted outcomes,” says Brian Balzer, vice president of digital technology and business transformation at G&J Pepsi.
A year earlier, G&J Pepsi’s executive team had approached its digital technology organization about providing predicted orders and store shelf optimization for its Pepsi products. “This was driven by the large amount of manual labor required to service our customers with the vast array of products, brands, and SKUs we offer,” Balzer says.
The company carries more than 250 different SKUs, and typically most of those products are in stock at any number of stores across its markets. The senior executives wanted the company to have an automated order mechanism to speed up processes and improve results.
Order writers at the company are required to know each store, consumer buying behaviors, sales activities, promotions, competitor tactics, weather changes, and more, Balzer says. “All of this is done manually and based on their own experience,” he says. “Some may be great at juggling all of this, but it’s time-consuming and is very dependent upon an individual.”
Furthermore, it can take individuals a long time to acquire this knowledge, Balzer says. “What if they leave the company? All of that knowledge goes with them and the next person has to be trained and learn it on their own,” he adds.
The reordering process is typically handled manually, with staffers counting empty spaces on shelves and in backrooms. “Much of this work is acquired knowledge from years of experience in each store,” Balzer says. “We began collecting this data and pumping it into the Azure ML models that are already built within the platform. We spent time tweaking those models with the more data we piped into it.”
As various types of data are fed into the machine learning models, they generate a predicted order. G&J Pepsi is in the midst of rolling out the automated order platform to all frontline employees currently servicing Kroger stores, and it plans to roll it out to those servicing Walmart stores in the coming months. The company is looking to use the same technology to begin determining shelf optimization for its convenience and grocery store segment.
“One of the biggest challenges any beverage company faces is determining what products to have in the cold spaces” within retailer stores, Balzer says. This requires having a clear understanding of how much quantity of a particular product should be available in each store, the proper location within the store coolers, and the profit potential for those products, he says.
“This can be a complicated formula, and one that changes market to market,” Balzer says. For instance, infused water or teas might sell more quickly in an urban location than in a rural market, whereas the opposite might be true for an energy drink. Developing the proper sets of products and optimizing storage space is critical to G&J Pepsi’s success.
The machine learning tool the company has developed, Cold Space Allocator, takes into account all of the variables and lays out an optimized product selection for each customer within each market. “It will also provide recommendations of products that might be outperforming in similar locations to replace slower selling products,” Balzer says. “Product optimization is an immense market advantage when done properly to meet consumer demands.”
The company can also use the data to show its customers which products are increasing their profits the most and which are in the most demand.
Since implementing the automated order platform, G&J Pepsi has seen a dramatic improvement in ordering efficiency. The time required to write orders has fallen from more than 60 minutes per store to about 10 minutes.
The company did face a few challenges as it began deploying the new technology. “The first and most important was to focus on the process,” Balzer says. “A great technology on a bad process will fail every time. It’s critical to fix process issues before implementing technology. We took time to partner with our frontline employees to understand how they manage their current processes, gain buy-in, and fix any process issues.”
For example, for the predictive order process to work, the company needed to ensure that all frontline employees were servicing customers the same way. “That means they need to walk the store the same way, identify backroom stock first, understand promotions, sales activities, etc.,” Balzer says. “They also needed to understand how buying behavior impacts our ability to provide a predicted order and when they should or shouldn’t adjust.”
G&J Pepsi also needed users to buy into why the automated order platform is valuable to them, how it makes them more efficient, and how it improves their ability to service customers. The employees’ had some concerns of their own.
“They needed to be reassured that we were not removing their job,” Balzer says. “We’re actually making their jobs easier and giving them time back to service more customers or spend more time with store managers to focus on selling. As they have more time to build relationships with each store, they will see improved results from growing those relationships and our brands.”
Zipline: Delivering medical supplies where they’re most needed
Zipline is a drone delivery service whose stated mission is no less than to provide every human on Earth with instant access to vital medical supplies including blood, vaccines, and personal protective equipment. The company’s drones have flown more than five million miles in multiple countries and completed more than 115,000 commercial deliveries, including bringing supplies to hospitals and clinics in some of the world’s most remote communities.
The company designs, assembles, and operates its unmanned aircraft system in the US and is progressing toward FAA certification of its drones and air carrier certification for its US operations.
“AI and machine learning were more or less ‘baked in’ to Zipline from the start,” says Matt Fay, data team lead at the company. “I don’t think you could design a cooperative fleet of autonomous aircraft without those tools.”
In the early stages before Zipline was flying hundreds of flight hours each day, developing intelligent behaviors needed less data-driven methods, because the company lacked the kinds of data sets that make those algorithms work, Fay says. “It wasn’t until we had begun flying, delivering medical products every day in Rwanda, that we had collected enough data to require new tools,” he says.
The company’s motivation at the time was two-fold, Fay says. “First off, we wanted to migrate from a local workflow—individual engineers downloading and analyzing a batch of flights on their own machines—to a cloud-based approach, where our entire flight history was already available,” he says.
Second, Zipline wanted to build an analysis environment, with powerful batch processing capabilities and a common, collaborative workspace. The software team was already fluent in Python, so the company deployed Jupyter Notebook, an open source web application that allows users to create and share documents that contain live code, equations, visualizations, and narrative text, running on a cluster of Apache Spark analytics engines.
A key component is a data science and machine learning platform from Databricks, which combines a scalable cloud-based computing environment with data streams from all aspects of Zipline’s operations—everything from flight logs to maintenance to tracking the provenance and status of parts and inventory at each distribution center.
“Because Databricks is a shared, collaborative environment, we’re able to invest in the platform: building our own set of utilities for batch processing, maintaining a plotting library of our most helpful data visualizations for flights, building a simple set of tutorials and training curriculum to onboard new team members,” Fay says.
“When most folks think of ‘data democratization’ initiatives, they’re usually thinking of dashboarding platforms that give access to analytics,” Fay says. “While that’s an important part of any strong data team’s arsenal, with [the Databricks platform], we’ve been able to democratize data science, giving everyone at the company the ability to combine, explore, visualize, and act on all of Zipline’s data.”
This broadly available capability has helped Zipline provide better service. The company’s customers, the health systems it serves, “rely on us to reliably deliver essential medicines on time,” Fay says. “Achieving this requires more than just a reliable aircraft; it takes sufficient operational capacity at each step of the process involved with fulfilling an order.”
An emergency delivery can be delayed for any number of reasons, everything from not enough staff on hand to pick and pack each product, to running out of fully charged aircraft batteries. “In order to understand the tradeoffs and bottlenecks in the larger system that is a Zipline distribution center, our team built an event-based simulation tool, modeling every step involved with delivering medical products,” Fay says.
Without tuning this simulation to “real-life data” taken from Zipline’s operations, “this tool would be uselessly inaccurate,” Fay says. “Only with that calibration complete are we able to ask and answer all kinds of invaluable hypothetical questions: ‘How will opening three new delivery sites impact our on-time rate at this distribution center? If we increased our charge rate by 10%, how many fewer batteries and chargers might we need? What is the best algorithm for dispatching aircraft?’”
Zipline has found that the insights from this tool impact practically every team at the company. “For that reason, along with the ease of continuously calibrating and updating the model, we’ve chosen to host it in Databricks,” Fay says. “This enables analysts with different needs across the company to see the same simulation results, and investigate the relevant parts.”
For Zipline customers and their patients, the technology has meant more reliable delivery of vital supplies.