We recently launched Machine-learned Service Times 2.0, a product enhancement we’re particularly proud of, and I wanted to share my perspective on why this is so significant to the last-mile fleets that we work with, and for the industry overall.
How Machine-Learning Impacts Last Mile Delivery
At Wise Systems, we’ve long believed in the innovative application of machine learning and AI to last-mile operations. In fact, that was one of the foundational beliefs we held when starting the company out of the MIT Media Lab. Our original hypothesis was that if we could gather and use fleet data, we could transform last-mile operations, and that’s what we’ve been doing ever since.
We gather and use huge volumes of data to improve performance for our customers in innumerable ways, ultimately increasing fleet efficiency and utilization, reducing customers’ carbon footprint, and helping our customers better serve their customers. Those are the ways we measure success, and that success is built around data and the way we use data inside of a sophisticated, and easy-to-use platform. Back to the data. Once you have data it’s time to put that data to work in the most beneficial ways, using the latest tools and techniques available to solve customers’ real-world problems.
Data-Driven Routing Solutions For Fleet Managers
As we looked at our customers’ operations, it became clear that having accurate service times was a big issue. It turns out that despite their best efforts, dispatchers and fleet managers have difficulty accurately predicting the amount of time it takes for a delivery driver to arrive at a location, deliver merchandise or service, complete the transaction with the customer, and return to the vehicle. Not only are there large variations between the amount of time it takes to service one customer vs. the other (think small retail location vs. big-box retailer), but there are also variables including the day of the week, the driver, the vehicle, the amount of product or type of service being delivered. So, many planners and/or dispatchers simply assign an average amount of time to all stops that the fleet delivers to. Others assign different numbers to different stop types in a best-guess effort.
In theory, those approaches seem reasonable. But, in practice, there are inevitable flaws that cause delays or missed time windows, which is why we originally developed Machine-Learned Service Times. In working with clients, we found that our original model — which looked at service time time alone — improved customers’ ability to create accurate, more efficient routes.
With this enhanced model, we’ve expanded the number of variables to include delivery size, weight and type, stop priority, as well as driver, day of week and month. For our customers, this translates into even more accurate service-time predictions which then serve as the basis for continuous and automated improvements.
While many in the industry are focused on the future potential of machine learning in supply chain applications, our perspective is different. We’re focused on delivering pragmatic, real-world applications that can be used today.