ML-powered Routing Tool – A foundational step towards developing continuous routing capability in Amazon India’s Last Mile (LM) operations
Amazon India’s delivery stations were stuck in a cycle of inefficiency – manually planning routes in batches throughout the day, limiting throughput, and increasing costs. Moving to continuous, ML-powered routing was the key to unlocking success.
Strategic Approach as the Charter Lead
Key Stakeholder Challenges
Picture this: a Bangalore LM delivery station receives around 2000 packages at 6 am and expects to receive another 2000 at 9 am. The station occupies a roughly 3000 sq. ft. area, which gets completely utilized for handling & dispatching packages. With the old routing tool, the operations team spent an average of 1 hour receiving packages, 45 mins preparing delivery routes, and 1 hour 15 mins dispatching the packages. This creates a bottleneck in timely processing of the next batch of 2000 packages arriving at 9 am in a space that’s blocked for handling the previous batch, ultimately delaying their deliveries. With the new ML-powered tool, that same planning now takes only ~10 minutes, which creates at least 5% more efficient routes, other than saving 35 minutes for every planning cycle.
To improve the station throughput, a foundational requirement was to “upgrade” from static route planning tool to an ML-powered route planning tool, which could be used to quickly create routes and incorporate multiple inputs like up-to-date traffic data, road entry points, vehicle capacities, etc.
Key Decisions Taken
Since the tool had the capability to influence routes by dialing up driver productivity, operations wanted to maximize productivity, while external partners wanted their delivery associates to continue getting familiar & easier routes with ~99% delivery success (meeting customer-promised delivery dates). If left unresolved, this would have resulted in optimized routes not being executed at all by delivery associates and potential driver & partner churn.
-
To manage partner expectations, I conducted regional calls with 3PL partners to present the cost-benefit analysis—how denser routes meant fewer stops, not more work. I also explained the pros of the new tool, considering updated geospatial and traffic data, preventing instances of routes spread on both sides of a busy road and reducing time spent on-road.
-
To set up operations teams for success, I presented simulation data to channel and operations leaders to establish why dialing up productivity to maximum would not result in delivery successes, as it would mean bypassing safety limits per vehicle (i.e., ignoring vehicle capacity of carrying shipments to allow drivers to safely ride) or non-uniform spread of denser vs. sparser routes throughout the day.
Results & Reflections
We saw a 6% improvement in overall productivity (calculated by measuring shipments carried per route or SPR, which increased from avg. 45 to 52) from optimized routes in 12 months of implementation, which brought down the cost per order totalling to annual cost savings of INR 20Cr. Further, to continue the momentum towards achieving continuous package processing capability as per the 3Y product vision, I enhanced the product roadmap for subsequent years with key deliverables to improve SLA adherence of promised delivery dates. These include bulk-deliveries for commercial destinations, package consolidating destined for the same location (to prevent different DA delivering diff. packages to same address throughout the day), and introducing DA-familiarity factor with certain pin codes to achieve ~99% delivery success.
Learning
-
User Research Depth: Initially I focused on route optimization, but user interviews revealed 40% of inefficiency was in package sorting. This taught me to dig deeper into the entire user workflow, not just the obvious problem statement.
-
Executive Alignment Strategy: Presenting theoretical gains to stakeholders created unrealistic expectations—15% productivity uplift was a theoretical ceiling from station-level modeling, but during the pilot stage, we identified three real-world constraints the algorithm hadn’t captured:
-
Route relearning curve for drivers when navigating unfamiliar, denser routes
-
Seasonal traffic variability, unique to traffic disruptions from India’s festivals calendar
-
Vehicle payload safety limits
-
-
When these constraints were identified during the pilot, I worked with the data science and engineering teams to incorporate them into the algorithm before full network expansion. The revised projections were presented to leadership with updated targets to reset their expectations. After full rollout across 400+ stations, we achieved a 6% improvement in shipments per route—from an average of 45 to 52 SPR—resulting in INR 20 Cr in annual cost savings. While below the original 15% projection, this outcome was fully explainable, defensible, and built on a more robust algorithmic foundation that would support the 3-year vision of continuous routing capability.
-
Change Management Scale: Underestimated the complexity of training 400+ stations simultaneously. Future rollouts should include dedicated change management plans, not just product guides and SOPs.
-
Partnership Dynamics: The pay-per-package pricing discussion could have been anticipated earlier. Including commercial implications in the original business case would have prevented stakeholder tension with 3PLs.