How Northwind Logistics Cut Fuel Costs 14% Fleet-Wide

Problem

Northwind Logistics dispatched routes the same way for years: each morning, dispatchers assigned drivers to zones based on familiarity and rough geographic sense. That worked when the fleet was 60 vehicles. Past 200, it started costing real money — dispatchers couldn't hold that much complexity in their heads, and fuel prices were climbing at the same time the fleet was growing.

Leadership's brief wasn't "add an algorithm." It was narrower and harder: cut fuel spend through better routing, without asking drivers to adopt new hardware, and without dispatchers losing the ability to override a route when they knew something the system didn't.

Research

We spent a week riding with dispatchers across two shifts before writing any code. The goal wasn't to learn routing — it was to learn why dispatchers already distrusted automated tools. Northwind had tried a routing add-on two years earlier that got abandoned within a month because it produced routes with no visible reasoning, and dispatchers reverted to manual planning rather than trust a black box.

That history reshaped the brief. The technical problem — geospatial route optimization — was the easy part. The real constraint was building something dispatchers would actually keep using after the novelty wore off.

Strategy

We proposed a "propose, don't dictate" model: the system generates an optimized route, but any dispatcher can override any leg with one click, and the override reason gets logged for future tuning. We also scoped a 40-vehicle single-region pilot before touching the full fleet, specifically to validate real-world fuel savings against actual driving conditions rather than simulation numbers.

Design

The interface surfaces the reasoning behind every route — distance, live traffic, delivery window constraints — instead of just handing dispatchers a single optimized line with no explanation. That single decision did more for adoption than any part of the optimization algorithm itself: dispatchers who can see why a route was chosen are far more likely to trust it on the days it's right and forgive it on the days it needs an override.

Implementation

The platform runs on Next.js with PostGIS handling geospatial routing logic and the Google Maps Platform supplying live traffic data. It's browser-based by design — no new in-cab devices, running on the tablets already mounted in each truck, which removed procurement and driver-training friction that would have otherwise added months to rollout.

Challenges

The hardest problem wasn't the routing math — it was earning back dispatcher trust after the previous failed tool. Some dispatchers manually overrode nearly every suggested route for the first two weeks of the pilot, effectively ignoring the system while still technically using it. Rather than push back, we let it happen and used the override log as data: which legs got overridden, and why, told us more about real-world constraints than any amount of upfront interviewing had.

Solutions

We ran three weeks of parallel routing — dispatchers planned manually as usual while the system quietly generated its own routes in the background for comparison — before the pilot went live for real. That parallel period is also where the override-logging paid off: two recurring override patterns turned out to be legitimate local knowledge (a permanently congested intersection, a client with a strict early-morning-only delivery window) that we fed back into the routing constraints, which cut override rates by more than half in week three.

Results

The 40-vehicle pilot validated fuel savings against real conditions, and the platform expanded to the full 230+ vehicle fleet over eight weeks, region by region, with each region's savings checked against its own baseline before moving to the next. Fleet-wide, fuel costs are down 14%, on-time delivery sits at 96%, and — the number that mattered most internally — dispatcher override rates settled at a low, stable level rather than climbing back toward "ignore the system," which is what killed the previous tool.

Lessons Learned

Letting dispatchers override routes with one click was worth more than any optimization refinement we made. A system that's 95% as accurate but visibly trusts the human operator will get adopted; a system that's more accurate but feels like it's taking control away won't, no matter how good the numbers are.

Future Improvements

Northwind and SmartFlow AI are on an ongoing monthly retainer refining the routing model as new depots and delivery zones come online — one of the few engagements still actively evolving well past initial launch, since fleet routing constraints change as the business grows.