US Regional Last-Mile Delivery Network
34% lower delivery costs through ML-driven demand forecasting and route optimization.
The client is a regional last-mile delivery company operating across the mid-Atlantic United States, serving approximately 140 e-commerce and retail clients. They run a fleet of 85 vehicles out of three distribution hubs, handling an average of 9,400 deliveries per day across a territory covering parts of Pennsylvania, New Jersey, Delaware, and Maryland.
The company had grown rapidly — 40% year-over-year volume increase over the prior two years — on the back of e-commerce acceleration. But the operational model that worked at 5,000 deliveries per day was failing at 9,400. The core problem was planning. Each morning, route planners at the three hubs manually built delivery routes using a combination of a legacy routing tool and personal knowledge of the territory. The process started at 4:30 AM and had to be completed by 6:15 AM when drivers began loading. The planners were experienced and good at what they did, but 9,400 stops across 85 vehicles in 105 minutes left no room for optimization.
The consequences were visible in the numbers. Average vehicle utilization was 64%, meaning trucks were running with a third of their capacity unused. Failed first-attempt deliveries ran at 14%, each one generating a redelivery that cost the company approximately $8.70. Fuel costs had increased 22% over the prior year, outpacing the 8% increase in fuel prices, indicating a structural routing inefficiency. The company was spending $3.2M annually on contract drivers to handle overflow that the existing fleet should have been able to absorb.
Demand prediction was essentially guesswork. Monday volumes were typically 30-40% higher than Fridays, but specific daily volume by hub and by zone was not forecasted. When actual volumes exceeded the morning plan — which happened roughly twice per week — the response was reactive: contract drivers pulled in at premium rates, packages held for next-day delivery, or existing routes extended past shift limits, triggering overtime.
The CEO had evaluated three commercial route optimization platforms. Each quoted $180K-$350K annually and required six to nine months of implementation. None addressed the demand forecasting problem, which the CEO correctly identified as the upstream root cause.
Millennial AI structured the engagement around two connected problems: knowing what tomorrow looks like, and then planning the best response to it. The forecasting model was built first because accurate demand prediction directly affects route planning, fleet allocation, hub staffing, and contract driver decisions.
Diagnose: Data Assessment & Baseline (Weeks 1-3)
Three weeks were spent extracting, cleaning, and structuring 26 months of delivery data from the client's TMS, fleet GPS records, and client order feeds. The data was messier than expected. GPS records had gaps for 12 vehicles, the TMS categorized deliveries inconsistently across hubs, and client order data arrived in four different formats. Building a clean, unified dataset took time, but nothing else worked without it. From this data, a precise operational baseline was established: actual cost per delivery by hub, by day of week, by zone, and by vehicle type. That baseline revealed that the cost differential between the best-performing and worst-performing hub was 41% — a gap the company had not previously quantified.
Design & Deploy: Demand Forecasting Model (Weeks 3-8)
The forecasting model was built to predict next-day delivery volume by hub and by delivery zone, with a 72-hour forward view updated daily. The model used historical delivery patterns, day-of-week and seasonal trends, client-provided order forecasts (where available), and external signals including weather forecasts and regional event calendars. A gradient-boosted ensemble approach outperformed the LSTM and Prophet models tested during development. The model was validated against eight weeks of held-out data and achieved 96.8% accuracy at the hub level and 91.4% at the zone level — compared to the planners' implicit estimates, which post-hoc analysis showed were accurate to within 15-20%.
Deploy: Route Optimization Engine (Weeks 6-11)
The route optimizer was built to consume the demand forecast and produce optimized route plans for each hub. The engine accounted for vehicle capacity, delivery time windows, driver shift limits, real-time traffic patterns, and the physical constraints of each vehicle type. The system generated a proposed route plan by 3:00 AM — three and a half hours before the old manual process began — giving planners time to review, adjust for local knowledge, and approve. Planners could override any route. The system was a starting point, not a mandate. Within four weeks, planners were accepting 88% of proposed routes without modification.
Scale: Real-Time Dashboard & Operations Integration (Weeks 11-14)
A real-time operations dashboard was deployed at each hub, showing live vehicle positions, delivery completion rates, predicted versus actual volume, and flagging routes that were falling behind schedule. The dashboard also surfaced next-day and 72-hour forecasts alongside current fleet and staffing levels, enabling hub managers to make contract driver and overtime decisions a day in advance rather than reactively on the morning of. Integration with the client's existing TMS ensured that route plans, delivery confirmations, and exception data flowed bidirectionally without manual intervention.
34%
Reduction in per-package delivery cost
Blended across all three hubs over the first full quarter
96.8%
Forecast accuracy (hub level)
Next-day volume prediction, validated over 12 weeks
82%
Vehicle utilization
Up from 64%, reducing wasted fleet capacity
$1.9M
Annual contract driver savings
Contract driver spend reduced from $3.2M to $1.3M
6.1%
Failed first-attempt delivery rate
Down from 14%, saving an estimated $8.70 per avoided redelivery
3:00 AM
Route plans ready
3.25 hours earlier than the prior manual process
The three route planners, who had initially been skeptical about the system, became its strongest advocates after the second week. The overnight route generation gave them time to focus on exception handling and driver communication rather than building routes under time pressure. The CEO reported that the contract driver cost reduction alone covered the full engagement cost within the first quarter. The company has since used the demand forecasting model to evaluate two potential new hub locations, using predicted volume distribution to model the economics before committing to a lease.
“Our planners are the same people — they just have better information now. The routes are ready before they walk in, the forecast tells them what tomorrow looks like, and they spend their time on the problems that actually need a human. We should have done this two years ago.”
CEO, Regional Last-Mile Delivery Network
Fleet costs growing faster than volume?
If you are planning routes manually and reacting to demand instead of predicting it, we can show you what a better model looks like.