Network-Level Routing in a World of Local Optimizers

Take a distribution network with three regional hubs feeding a shared pool of forty long-haul lanes, each with its own dispatcher optimizing that region's freight cost against that region's carriers. Each dispatcher is good at their job. Each one, most weeks, hits their regional cost-per-mile target. And the network they jointly operate still comes in over plan often enough that finance stopped being surprised by it — not because any one region is failing, but because lane-level optimization and network-level optimization are different objective functions wearing the same name, and a network built entirely from local optima routinely produces a worse global outcome than a network with a few deliberately "suboptimal" regional decisions.

A concrete version of how this happens

Say Hub A has a shipment that can move on a dedicated lane at $2.10 a mile, or get consolidated onto a partially-loaded truck already running from Hub B through A's territory at an effective $1.40 a mile once the shared load is accounted for. Hub A's dispatcher, optimizing locally, doesn't see the Hub B truck — it's not in their system, their carrier relationships, or their KPI. They book the dedicated lane at $2.10, hit their own cost target for the shipment, and the network eats a preventable $0.70-a-mile premium on a 900-mile lane, which is $630 on one shipment and compounds across every similar case that week. Multiply that pattern across forty lanes and three regions running independently, and the aggregate gap between what the network paid and what the network could have paid, given the freight it actually had moving that week, is where a meaningful share of that $180,000 typically lives — not in any single bad decision, but in hundreds of individually reasonable ones that never had visibility into each other.

Why this isn't a coordination failure you can fix with a meeting

The instinct is to say the dispatchers should talk to each other more, and in a three-hub network you can partially patch the problem that way. It stops scaling almost immediately. At ten or twenty distribution points, the number of pairwise combinations where a consolidation opportunity might exist grows far faster than any group of dispatchers can track by relationship and memory, and the ones with the best relationships across hubs end up doing informal, manual optimization that's neither systematic nor visible to anyone auditing network cost. What looks like a communication problem is actually a visibility problem: nobody has a single view of network-wide freight in motion, updated on a timescale that matches the decision, so consolidation opportunities get found by luck and tenure rather than by the system surfacing them as a matter of course.

What network-level routing actually requires

The fix isn't replacing regional dispatchers with a central planner who overrides every local call — that trades one failure mode for a slower one, and it strips out the local knowledge dispatchers genuinely have about carrier relationships and regional constraints that a central model won't capture well. What works is giving every dispatcher's routing decision visibility into total network cost before it's booked: a shared view of in-motion freight across all hubs, a live cost comparison that includes the consolidation option alongside the dedicated-lane option, and a routing recommendation that a local dispatcher can accept, override, or escalate — but can no longer make in genuine ignorance of what else is moving through the network that week. The dispatcher keeps the judgment call. The system stops letting that judgment call happen blind.

The result, six months into running it this way

In the network we're describing, moving from lane-level to network-level routing didn't eliminate the $180,000 gap in one quarter — some of it was structural, tied to genuine capacity constraints that no amount of visibility fixes. But it closed roughly two-thirds of it within two quarters, almost entirely by surfacing consolidation opportunities that had always existed and had simply never been visible to the person positioned to act on them. The more durable change wasn't the cost reduction itself. It was that the definition of a "good routing decision" shifted, organization-wide, from "the best price my region could get" to "the best price the network could get given everything currently in motion" — which sounds like a small semantic difference and turns out to be the entire difference between a network of good local decisions and a genuinely optimized network.