Last-Mile Delivery Cost in 2026: Why the Number You Price New Business With is Eight Times Too Low

Last-mile delivery cost is usually expressed as an average: total last-mile operating cost divided by successful deliveries. That number is correct for reporting and wrong for almost every decision it gets used for, because decisions about new business are incremental rather than average. The trouble is that the incremental number is easy to calculate badly. Adding an account to today’s routes and measuring what it adds produced $0.70 a stop in this model, while the same account measured against the fleet the operation ends up running cost $5.51, a difference of 7.8 times from the same shipments. Locus, the world’s first Decision-Intelligent, Agentic TMS, plans the incremental work against the full constraint set and the fleet it actually requires, which is the only way the second number gets produced.

Why the Incremental Number is Usually Wrong: The Business Case

The reason a marginal figure looks so small is that most last-mile cost is not distance. It is the driver, the vehicle and the day. ATRI’s 2025 analysis of operational costs found that excluding fuel, marginal costs rose 3.6 percent to $1.779 per mile, the highest costs ever recorded for non-fuel operating costs, with driver benefits alone at $0.197 per mile. That is long-haul rather than last-mile work and the absolute figures do not transfer, but the structure does: the expensive things are attached to the vehicle and the shift, not to the kilometer. Last mile concentrates that further, because service time at the door means a vehicle covers far fewer miles in a shift, so the fixed daily cost is spread over less distance and dominates the per-kilometer component even more heavily.

Which means adding stops to a route that is already running looks almost free. The vehicle is out, the driver is paid, and the only new cost is a short detour. McKinsey’s work on out-of-home delivery puts the same effect from the other direction, finding that raising drops per stop from one to five cuts labor and vehicle cost by more than 50%. Density is close to free. That is a real economic property and it is also a trap, because the route only stays already-running until the work outgrows it.

At that point the next stop does not cost a detour. It costs a vehicle, a driver and a shift, and the cost curve steps rather than slopes. An operation that quotes from the flat part of the curve and delivers from the step has priced work it cannot serve at the price it quoted.

The stakes follow from where the money sits. With last mile running 60% to 70% of total parcel delivery cost, a systematic understatement in incremental costing is a systematic margin leak, and it grows with the business. The World Economic Forum projects 36% more delivery vehicles in the top 100 cities globally by 2030, which is the same pressure arriving as volume.

How a Bad Incremental Number Gets Built

1. The account is dropped onto the current plan

Someone adds the prospective stops to a recent day’s routes and reruns the optimizer. This is the fastest way to get a number and it is the one most often used.

2. The optimizer absorbs them into existing routes

With slack in the shift, the new stops slot into routes that already pass nearby. Total distance rises a little and total cost rises very little.

3. The result is divided by the new stops

The difference in cost divided by the number of added deliveries produces an incremental cost per stop, and it is a small number because nothing was added to the fleet.

4. Slack is finite and unevenly distributed

The absorption only worked because those particular routes had room on that particular day. Another day’s geography, or a peak week, and the same stops force overtime or another vehicle.

5. The quote is made from the flat part of the curve

Price is set against the small number. The work is won, and it is served from a fleet that has to grow to accommodate it plus everything else that arrived in the same period.

6. The shortfall appears as a general margin decline

Because the cost lands on the fleet rather than on the account, it never reconciles back to the deal that caused it. The operation sees costs rising faster than volume and cannot point to why.

What the Model Shows

The model builds a depot operation from stated inputs rather than observed customer data: roughly 118 stops a day across an urban core and a periphery out to 36 km, vans at 380 units capacity, an illustrative $175 per vehicle-day for driver and vehicle, $0.26 per km running cost, $34 an hour for overtime beyond an eight-hour shift, and 4.5 minutes of service per stop. Routes are built by balanced angular clustering with nearest-neighbor sequencing and local search. The fleet is sized to the cheapest feasible option, which lands at four to five vehicles running roughly 24 stops each over about 5.8 hours. A 20-stop account is then added and costed two ways.

Average cost per stop was $7.93. This is the number the operation reports and the one most rate cards are sanity-checked against.

Added to a fixed fleet, the same account cost $0.70 a stop. That is 0.09 times the average. Across distance bands it ran from $0.39 in the core to $1.15 for an account outside the existing footprint. Every one of those figures is a correct answer to the question “what does this add to today’s routes.”

Costed against the fleet actually required, it cost $5.51 a stop. That is 0.69 times the average, and the understatement between the two treatments averaged 7.8 times, ranging from 4.1 times for a distant account to 14.1 times for one in the core. The core case is the worst because core stops absorb most easily into existing routes, which is exactly what makes the fixed-fleet figure most misleading there.

Account size does not rescue the estimate. Costed properly across sizes from 5 to 80 stops, incremental cost per stop stayed between 0.63 and 0.89 times average, with no size at which it approached the near-zero figure the fixed-fleet method produces. Larger accounts do not become dramatically cheaper per stop; they simply force the fleet decision sooner and more visibly.

Geography matters far less than the accounting. Varying both the distance band and how tightly the account clustered, incremental cost per stop moved only between 0.60 and 0.89 times average. A 20-stop account dispersed across a 9 km radius 40 km from the depot was not dramatically more expensive per stop than a tight cluster in the core. The choice of whether to count the vehicles moved the answer by 7.8 times; every physical variable tested moved it by less than 1.5 times.

This sharpens a known pricing trap rather than contradicting it. The standing advice for shared networks is that marginal-cost pricing loses money, because each client’s increment is measured against a network the others also pay for, so the increments sum to less than total cost. That holds, and this model adds the reason the trap is so easy to fall into: the increment most operations calculate is not even a true increment. It is an absorption figure taken against a frozen fleet, and it sits roughly eight times below the real one. An operation can therefore under-recover twice over, once because increments do not sum to total and once because the increment itself was understated.

What the model does not settle. It assumes the operation will resize its fleet rather than degrade service, it uses a single vehicle type, and it prices a single account against a single baseline rather than a portfolio arriving over a year. It also says nothing about what the account should be charged, which depends on competitive position and the rest of the book.

Three Costing Bases and What Each Is For: Key Differences

What to Look for in Cost-to-Serve Modeling

Fleet size treated as an output, not a fixed input

A costing run that holds the vehicle count constant can only ever produce the absorption figure. The model has to be free to add capacity, because that is what the operation will do.

The capacity step visible in the output

The useful chart is not a smooth cost curve but a stepped one, showing where the next vehicle becomes necessary. An operation should know how much volume it can take before the step and what the step costs.

Costing across multiple days, not one representative day

Absorption depends on the geography of a particular day, because whether a route has room for three more stops is a question about that day’s orders. A single-day estimate lands somewhere on a wide distribution, and it tends to land on the favorable side because the person running it picks a normal day rather than a peak one. Costing across a month of real order files removes most of that bias for very little extra effort.

Portfolio-level reconciliation, not account-level addition

Because incremental costs do not sum to total cost, a rate card built by pricing each account at its own increment under-recovers by construction. The reconciliation has to happen at the book level.

Realized cost compared back to the quote

Once the account is live, the cost actually incurred should be comparable with the cost quoted. Without that loop the estimate never improves, and the error repeats on the next deal with more confidence than before. The practical test for a platform is whether it can show, for a named account twelve months in, the vehicle days and distance attributable to it against what the model predicted.

Cost Structure in Practice

A beverage distributor with depot-based mixed fleets. Vans, trucks and motorbikes serving thousands of small retail points a day, previously planned in spreadsheets with over an hour of manual work before any vehicle moved. Fuel consumption fell 37% and orders per delivery trip rose 22%. The second number is the cost-structure number: more drops per trip is the density effect being captured rather than given away.

A Fortune 50 parcel and logistics network. More than a million freight shipments a year across 51 sites and a 4,500-strong driver pool, each site measuring utilization against its own plan. Centralizing raised weekly execution from 75% to 92% and surfaced more than $14M in unused capacity, including $565K at a single site. Unused capacity is absorption headroom nobody could price, because no one could see it.

A global FMCG distribution network. Ten Asian countries, 1,000+ distributors and 5,000+ riders, with 12,000+ trips a month eliminated against $4B+ in optimized orders, at 3X ROI. Removing trips is the only intervention that reduces the fixed side of the cost base rather than spreading it further, which is why it shows up in the cost per order rather than in the cost per kilometer.

Common Mistakes in Pricing Incremental Delivery Work

Quoting from a fixed-fleet estimate. It is the easiest number to produce and it understated true incremental cost by 7.8 times here. It answers a question about absorption, not about cost, and the two get confused because both are expressed per stop.

Assuming distance is the main driver. Every geographic variable tested moved incremental cost by less than 1.5 times, while the accounting treatment moved it by 7.8. A costing debate about which postcodes are expensive is usually arguing about the smaller variable.

Pricing each account at its own increment. Increments do not sum to total cost, so a book priced this way under-recovers by the density dividend, and the shortfall only becomes visible once every contract is live.

Treating the average as a floor. Average cost is not a floor either. It is a different number for a different purpose, and using it to reject incremental work turns away business the operation could genuinely absorb at a profit. Both errors are live in most commercial teams at the same time, on different deals, because neither number is wrong so much as misapplied.

How Locus Helps Cost Incremental Work Properly

Locus, the world’s first Decision-Intelligent, Agentic TMS, produces the second number rather than the first, because it plans the incremental work against the full constraint set rather than dropping it onto an existing plan. The route planning and dispatch layer solves routing, load allocation and vehicle assignment together across more than 250 real-world operating constraints including capacity, driver hours, time windows and access rules, so a scenario that needs another vehicle returns a plan with another vehicle in it rather than a plan that quietly breaches a shift limit. That is the difference between an absorption estimate and a cost. The Control Tower then carries the executed record against the plan, which closes the loop that most costing exercises never close: what the account actually cost, compared with what it was quoted at. Six governance mechanisms covering explainability, traceability, evaluation, autonomy levels, execution sandbox and human-in-the-loop keep each of those decisions traceable to the inputs that produced it.

The platform reasons across those constraints over 1.5B+ deliveries for 360+ enterprise customers in 30+ countries at 99.99% uptime, with $320M+ in aggregate logistics cost savings, 800M+ miles reduced and 17M+ kg of CO2 avoided. Locus has been recognized by Gartner for seven consecutive years, including the 2026 Gartner Hype Cycle for Supply Chain Execution and Logistics Technologies and the 2026 Gartner Market Guide for Multicarrier Parcel Management Solutions, where ShipFlex is featured as a Representative Vendor. Locus holds Leader designation in the QKS SPARK Matrix for Transportation Management Systems 2025 and the #1 position for Route Planning in G2’s 2026 Best Software Awards. In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

Two deployments show the cost structure moving rather than being re-divided. A beverage distributor running vans, trucks and motorbikes from depots to thousands of small retail points cut fuel consumption 37% and raised orders per delivery trip 22%, with route planning time down 35% and end-of-day reconciliation down 60%, after address validation pinned each retail point and vehicle types were matched to routes. A Fortune 50 parcel and logistics network running more than a million freight shipments a year across 51 sites lifted weekly execution from 75% to 92% and exposed more than $14M in unused capacity at 99.99% uptime, once planning was decided across the network rather than inside each depot. In the first case the fixed cost base was spread across more orders; in the second, capacity that was already paid for became visible enough to sell.

Average last-mile delivery cost is the right number for reporting and the wrong one for pricing new business, but the incremental number most operations produce is worse. Measured against today’s routes, a 20-stop account cost $0.70 a stop in this model. Measured against the fleet the work actually requires, the same account cost $5.51, and the gap between those two treatments was 7.8 times, larger than every geographic variable tested combined. Getting it right means letting the fleet size float in the model, pricing against the stepped cost curve rather than the flat part, and reconciling at the book level because increments never sum to total. Locus plans incremental work against the constraints and the capacity it genuinely needs, so the number that reaches the quote is the number that reaches the road. Request a Locus cost-to-serve review to see what your last account actually cost you.

Frequently Asked Questions

What is the average last-mile delivery cost per stop? In this model it was $7.93 a stop, on an operation of roughly 118 deliveries a day served by four to five vans. The figure is highly specific to drop density, service time and labor cost, which is why published benchmarks rarely transfer between operations.

Why is average cost the wrong basis for pricing new business? Because accepting an account is an incremental decision, not an average one. The relevant question is what the work adds to total cost, and that can be well below average when capacity absorbs it or well above when it forces another vehicle.

How do you calculate the incremental cost of a new delivery account? Re-plan the whole operation with the new stops included and let the fleet size adjust to whatever the combined work requires, then compare total cost with and without. Holding the vehicle count fixed produces an absorption figure, which understated true incremental cost by 7.8 times here.

Does it matter where the new account is located? Less than most costing debates assume. Across distance bands from the core to 45 km out, and across tight and dispersed clusters, incremental cost per stop stayed between 0.60 and 0.89 times average. The accounting treatment moved the answer far more than the geography did.

Can you price several accounts at their incremental cost? No. Each account’s increment is measured against a network the others also share, so the increments sum to less than total cost and a book priced that way under-recovers. Incremental cost informs whether to take work; the rate card has to reconcile at portfolio level.

What does it cost when an account forces another vehicle? A full vehicle-day, which in this model is an illustrative $175 of fixed cost before any distance is driven. That is why the cost curve steps rather than slopes, and why an estimate taken from the flat part misprices work that will be served from the step.