Use big data and Artificial Intelligence (AI) to wipe out waste from last-mile deliveries (LMD). COVID-19 has put disruptive pressure on last-mile B2B and B2C (business-to-business and business-to-consumer) deliveries. Customers demand higher volumes at lower costs in shorter times. As providers scramble to cope with the ruckus, they risk their margins because of the manual, wasteful ways of meeting the demand.
Today, all vehicles leaving the warehouse at full capacity gives a semblance of high vehicle utilization, but a full truck may also mean undue kilometers traveled or time worked. Most supply chain practitioners rely on experience, familiarity with delivery areas, and intelligent guesswork to draw up their delivery clusters and route plans. Unfortunately, this approach for LMDs rarely yields the optimum. Additional delivery jobs that route planners “must” accommodate even if the job requests come in after the deadline throw the plans into a tailspin. They have to move jobs from one vehicle to another, change routes, hire additional vehicles, or a combination of these. The result — vehicle volume or weight capacities aren’t fully utilized, and routes aren’t the most cost-efficient paths.
Below, I advocate for a LMD optimization framework that uses AI, big data, and state-of-the-art quantitative models to improve deliveries. Let’s examine the six elements of this framework.
1. Objectives: For a varying set of load and delivery locations per day, the overall objectives are: to a.) reduce the number of vehicles to deliver the same load, b.) reduce the total distance traveled by all vehicles, and c.) speed up the clustering, load planning, and route optimization process from hours to minutes.
2. Dynamic Configuration: Define clusters, load distribution, and routes dynamically. These should vary as orders and delivery locations vary. Many practitioners delimit clusters based on geographical boundaries, e.g., zip codes, barangays, towns, or cities. They assign clusters to delivery vehicles, and vehicles deliver to the assigned clusters whether or not the vehicles are fully loaded. In contrast, route planners should define clusters dynamically, i.e., based on the jobs for that day. If the jobs straddle the boundary of cities A and B, the cluster should cross that boundary instead of one cluster limited to A and another limited to B. The configuration of clusters and routes tomorrow will differ from today’s because the delivery jobs for tomorrow will be different.
3. Delivery Constraints: The definition of clusters and routes are subject to constraints. The more constraints, the higher the cost to deliver. It is costlier to deliver within a narrow time window than it is within a wide time window. Typical constraints are delivery time window, vehicle capacities, service times at delivery locations, truck bans, narrow roads, and traffic.
4. Hacks: First, tag addresses with accurate GPS coordinates (a.k.a. geocode the addresses), and update volume and weight of parcels or SKUs. All this data is crucial for the algorithms to work well.
Second, cluster and route using road distance, not straight line distance. Locations that are visually near each other may not be so by road distance.
Third, ensure that the reduction in the number of vehicles required is not offset by an increase in distance traveled or time worked. Off-the-shelf and uncalibrated routing algorithms pretend to create optimum clusters and routes with deceivingly excessive offsets.
Fourth, confirm that the optimization models you work with are fast enough to produce improved plans within the time limits required by your operations. An initial step of these algorithms is to evaluate the different permutations of distances between all delivery points. For 50 delivery locations, there are at least 2,500 permutations. For 5,000 delivery locations, there are at least 25 million permutations. The algorithms should be fast enough to iterate through these permutations in minutes and not hours. Powerful algorithms can reconfigure clusters and routes midway through the deliveries to incorporate new pick-up points for on-demand delivery or voluminous courier operations.
Finally, be critical about manually changing clusters and routes without substantial evidence to do so, especially in the early stages of shifting from manual route planning to AI-driven route planning. It takes getting used to trusting the output of the models.
5. Economic Gains: I worked with a supply chain company with distribution centers across the country that delivers to traditional and modern trade, convenience stores, hotels, and restaurants. We attempted to reduce the delivery trucks required to deliver to 40-60 locations daily. This exercise yielded an average 17% decrease in trucks required or a 17% transport cost savings.
6. Pricing Model: A prevalent charging method in the LMD business today is a fee per truck per day. Under this model, it makes no sense for a delivery provider to reduce the number of trucks through optimization because such reductions only mean less revenue. To address this disincentive, delivery providers should charge a fee per cubic meter (CBM) per drop zone. Identify concentric drop zones from the distribution center. The farther the drop zone, the higher the fee per CBM delivered. For dense items, use a fee per kilogram per drop zone instead. Pricing per CBM or kilogram is proportional to the service you provide and incents the reduction of delivery vehicles.
This framework I outlined boosts margins and utilization. The advent of big data, machine learning, and computing power has made the framework less challenging to implement, and more effective for dynamic clustering and constraint management. I hope these ideas encourage you to rethink the efficiency of your LMDs in the context of current technology innovations.
This article reflects the personal opinion of the author and does not reflect the official stand of the Management Association of the Philippines or the MAP.
Cliff Eala is a member of the MAP NextGen Committee and is the President and CEO of technology firm Synerbyte Limited