The Art of Deployment: How UPS Overcame Resistance to Machine Learning and Saved Millions
Abstract:
Machine learning’s potential for revolutionizing industries is undeniable, yet its deployment often falls short of delivering promised value. This article delves into the challenges of machine learning implementation, particularly the resistance to change and the need for effective change management. Drawing upon UPS’s successful deployment of Package Flow Technology (PFT), a predictive analytics system for optimizing package delivery routes, the article highlights the importance of managing change, fostering understanding, and providing incentives to drive adoption.
Introduction:
Machine learning has emerged as a transformative technology with the power to revolutionize various industries. However, despite its immense potential, many organizations struggle to successfully deploy machine learning models and capture their intended value. This article explores the challenges associated with machine learning deployment, particularly the resistance to change and the need for effective change management. To illustrate these challenges and demonstrate successful implementation, the article presents a case study of UPS’s deployment of PFT, a predictive analytics system that revolutionized its package delivery operations.
Section 1: Resistance to Deploying Machine Learning
The article begins by highlighting the prevalent issue of machine learning projects failing to achieve business value due to deployment challenges. It emphasizes that the technical aspects of machine learning, while crucial, are often not the primary obstacles. Instead, it is the resistance to change and the lack of focus on the business value that hinder successful deployment.
Section 2: UPS’s Package Delivery Optimization Challenge
The article introduces UPS’s challenge of optimizing its package delivery process to reduce mileage and driver hours. It describes the complexity of the problem, involving the assignment of packages to delivery trucks based on incomplete information and the need to plan truck loading well before all deliveries are known. This section sets the stage for the introduction of PFT as a solution to address this challenge.
Section 3: The Delivery Paradox and the Need for Prediction
The article introduces the concept of the Delivery Paradox, a central dilemma faced by UPS in planning truck loading. This paradox arises from the fact that optimal planning requires knowing all deliveries in advance, which is impossible due to the unpredictable nature of package arrivals. The article emphasizes the importance of prediction as a means to resolve this paradox and enable effective planning.
Section 4: Predicting Tomorrow’s Deliveries with PFT
The article describes PFT as a predictive analytics system designed to address the Delivery Paradox. It explains how PFT generates predictions for tomorrow’s deliveries, augmenting the list of known packages and enabling the planning process to begin earlier. The article also highlights the iterative nature of PFT’s predictions, which are continuously updated until the trucks head out for delivery.
Section 5: Why Stodgy Machine Learning Projects are the Sexiest
The article challenges the notion that only flashy machine learning projects, such as self-driving cars or AI systems that play games, are truly valuable. It argues that seemingly mundane projects, like UPS’s PFT, can deliver significant business value by directly improving established, large-scale operations. The article emphasizes the importance of focusing on the ends—process improvements—rather than the means—technology.
Section 6: Shift Happens: When a Legacy Process Goes Digital
The article discusses the process changes introduced by PFT, highlighting the shift from a legacy process of manual truck assignments to a centralized, semi-automated system based on predicted deliveries. It explains how this change aimed to improve efficiency by reducing mileage and driver time, and by centralizing decision-making. However, the article also acknowledges the challenges faced in overcoming ingrained habits and ensuring compliance with the new process.
Section 7: To Manage Change, Change Management
The article emphasizes the importance of change management in successfully deploying machine learning projects. It highlights the need for dialogue between decision-makers and those impacted by the change, as well as the importance of emotional commitment from employees. The article discusses the efforts undertaken by UPS to drive adoption of PFT, including training, incentives, and performance monitoring. It also highlights the role of fostering understanding and conviction among staff, as well as providing rewards for achieving leading indicators of success.
Section 8: The Outcome for UPS
The article presents the positive outcomes achieved by UPS as a result of PFT’s deployment. It reports significant savings in mileage, fuel consumption, and emissions, as well as improved efficiency in package delivery. The article credits PFT with saving UPS millions of dollars annually and earning it industry recognition and media attention.
Section 9: The Future of Prediction
The article concludes by emphasizing the need for a comprehensive approach to machine learning deployment, encompassing technical requirements as well as change management strategies. It highlights the importance of aligning stakeholders, ensuring model transparency and understandability, and providing clear metrics for evaluating model performance. The article also calls attention to the often-overlooked requirement of managing the change that model deployment entails, emphasizing that capturing the potential value of machine learning requires proactive management of operational changes.
Acknowledgments:
The article acknowledges the contributions of various individuals and organizations, including Samuel Bodily, Eric Tassone, Michael Albert, Alex Cowan, Sasa Zorc, Rupert Freeman, and Jack Levis. It also expresses gratitude to the Editor-in-Chief and anonymous reviewers of Harvard Data Science Review for their feedback in improving the article.
References:
The article provides a comprehensive list of references, citing academic research, industry reports, and news articles that support the key points discussed throughout the article.