Routes are optimized, but last-mile costs are still rising. Why?

Last-mile delivery now represents between 41% and 53% of total shipping costs, up from 41% in 2018. The global last-mile market is valued at approximately $201 billion in 2025 and projected to grow at a 12% compound annual rate through 2029. Every operator is being told the same thing: get this number down.

Most are trying. They are investing in route optimization software, renegotiating carrier contracts, pushing stop counts higher, and tightening delivery windows. A 2025 survey of global transportation professionals found that 96% already use AI in their operations, with route and load optimization among the top three use cases.

And yet still, 76% of retailers report that last-mile delivery costs have increased, and most describe home delivery as unprofitable under current cost structures.

But why?

The plan is not the problem

Route optimization has significantly improved in the last few years. AI-powered stop sequencing, zone-based carrier assignment, dynamic dispatch windows. These gains are real. The morning plan for a mature delivery operation is measurably more efficient than it was two years ago. Most large operators have already captured the bulk of what better planning can produce.

But the constraint has moved, and it is no longer in the plan, but instead in everything that happens after the plan is set.

96% of operators are already using AI for route optimization. The operators still seeing costs rise are not failing at planning. They are running into the limits of what planning, however good, can control.

Where the cost actually lives

A delivery plan is a prediction. It encodes assumptions about provider availability, pricing, zone conditions, and stop sequencing that are accurate at the moment the plan is built. The moment execution starts, reality begins diverging from those assumptions, and the cost lands across three places most operators never see together.

During the delivery

• A building entrance isn’t in the system and the driver spends twelve minutes finding it, cascading dwell time through the rest of the sequence
• A platform invoice arrives days later with a rate that doesn’t match what was quoted at assignment
• A window gets missed and a re-delivery gets scheduled before anyone has flagged a problem

After the delivery

• First-attempt failure rates run between 8% and 20% depending on geography and provider mix
• Each failure costs retailers an average of $17.20 in re-delivery, handling, and support
• Many customer support inquiries are about ETAs, cancellations, and delivery status volume that traces back to execution failures no single system flagged in time

The long-term revenue cost

• 85% of consumers say they will not shop with a retailer again after a poor delivery experience
• For time-sensitive orders (such as grocery, same-day retail, household essentials) a failed delivery means a customer who reorganized their day around a promise that wasn’t kept
• That revenue tail never appears on a cost-per-delivery report, because no single report sees the full picture

Across all three, the cost is measurable, but by the time it surfaces, the decisions that caused it are long gone. In response to this, the operators closing that gap are building systems that learn from every delivery they complete, to prevent the costs from occurring in the future.

The lever that compounds savings

Every delivery generates a data point. In a system built to use that data in real time, each completed delivery updates the network’s understanding of provider performance, zone conditions, and true route costs before the next order is dispatched. The invoice discrepancy that appeared three days late gets caught at assignment. The corridor where windows are consistently missed gets flagged before the next SLA breach. The provider that looked competitive last quarter gets reassigned before the cost of its underperformance compounds further.

For operators running gig fleets across multiple platforms, this is where the cost structure improves. The platform fee jump, the missed window, the re-delivery that never appeared on a cost report: each one becomes a signal the network learns from and adjusts to before the same issue repeats.

As volume grows, the network gets smarter. Every additional delivery adds another data point to a model already shaped by every previous one, creating compounding gains in provider utilization, cost per delivery, and first-attempt success rates.

The Gartner 2025 Future of Logistics Survey ranked digitalization of processes as the second-highest priority for logistics functions over the next five years, just behind customer experience. A network that continuously learns from its own operations is one of the few architectures designed to address both at once.

The cost of staying static

Every operator running last-mile delivery today is paying for the plan-reality gap. The costs are distributed across too many systems, too many providers, and too many line items to surface as a single number. That invisibility is what a learning system can begin to tackle.

The operators pulling ahead on cost-per-delivery are running systems that treat every delivery as data, feed that data back into the next decision, and get incrementally better at every dimension of the problem: provider selection, route reality, platform cost, first-attempt success.