Last-mile delivery is not trade delivery with smaller boxes. The economics are driven by different variables, the failure modes are different, and planning approaches borrowed from wholesale distribution consistently underperform.
The density equation
Last-mile cost per parcel is dominated by one relationship: how far you drive between consecutive drops.
`` Cost per stop ≈ (driver cost + vehicle cost) / stops per hour Stops per hour ≈ 60 / (service time + inter-stop travel time) ``
At 30 stops per hour, an extra minute of average inter-stop travel costs you roughly 15% of your productivity. In dense urban routes, inter-stop travel can be under 90 seconds; in rural delivery it can exceed eight minutes. That single variable explains most of the cost difference between operations that otherwise look identical.
Everything else in last-mile planning is a means of increasing effective density:
- Delivery day consolidation — offering fewer delivery days per postcode area, so drops cluster.
- Time slot shaping — steering customers towards slots that suit the emerging route.
- Pickup points and lockers — converting many low-density drops into one high-density one.
- Reattempt policy — an aggressive redelivery policy manufactures low-density work for tomorrow.
Every routing product demos beautifully on a suburban grid. Ask for a plan over your densest urban postcode and your sparsest rural round.
Why first-time delivery rate dominates everything
A failed delivery costs the whole stop again, plus handling, plus a customer contact, plus often a refund risk. In consumer delivery the difference between 92% and 97% first-time success is usually worth more than any routing improvement available.
Levers, in rough order of effectiveness:
- Reliable, narrowing ETAs — day-before window, morning refinement, live tracking on approach.
- Safe place and neighbour authorisation captured at order time, not renegotiated at the door.
- Access instructions as structured data — building entry, concierge hours, gate codes.
- Photo proof with a defined standard, which cuts disputes as well as failures.
- Customer-initiated reschedule before the vehicle is loaded, not after it has failed.
What last-mile routing software must do differently
| Requirement | Why |
|---|---|
| Very fast re-optimisation | Order cut-offs are late and volumes are large |
| Slot-aware order capture | The routing engine must be callable at checkout |
| Dynamic ETA publication | Customer expectation, and it drives first-time success |
| Walking-sequence support | In dense urban areas the driver parks once and walks several drops |
| Parcel-level tracking | Not just stop-level: multiple parcels per stop with independent statuses |
| Exception workflows | Refused, damaged, absent, access failed — each with a different next action |
| Reattempt scheduling | Automatic re-injection of failed drops into subsequent plans |
Walking sequences deserve emphasis. In apartment-dense areas, the vehicle stops once and the driver serves six addresses on foot. Software that models each address as a separate vehicle stop will produce plans that are wrong by a wide margin, and driver-reported times that never match the plan.
Urban constraints that break naive plans
- Vehicle access restrictions — low emission zones, weight limits, timed pedestrian zones, width restrictions.
- Parking reality — the nearest legal stopping place may be 200 metres from the address, and it varies by time of day.
- One-way systems and turn restrictions — a plan built on straight-line distance will be wildly optimistic.
- Building access — concierge hours, secure entry, service lifts with booking systems.
- Kerbside competition — in commercial districts, loading bay availability is a genuine constraint that almost no software models.
Where routing software cannot represent these, the standard workaround is enriched service times by micro-zone: build the constraint into the time model rather than pretending it does not exist.
Rural last mile
The opposite problem. Drive time dominates, density is unachievable, and the levers are different:
- Fixed weekly delivery days by area to force clustering.
- Longer service windows in exchange for reliability.
- Local collection points — shops, petrol stations, community hubs.
- Bigger drop sizes through consolidated ordering incentives.
- Genuinely offline-capable driver apps, since coverage will fail.
Applying urban KPIs to rural routes produces demoralised drivers and meaningless reports. Segment your targets by density band or you will manage the wrong thing.
Metrics worth managing
| Metric | Why it matters |
|---|---|
| Stops per hour, by density band | The core productivity measure |
| First-time delivery rate | Dominates cost per successful delivery |
| Cost per successful delivery | The true unit cost, including failures |
| Plan-versus-actual time variance | Early warning of model decay |
| ETA accuracy (within promised window) | Drives customer contacts and failures |
| Parcels per stop | Consolidation effectiveness |
Cost per successful delivery is the one to put on the wall. Cost per attempted delivery flatters an operation with a poor first-time rate.
Common questions
What stops-per-hour should we expect?
It depends almost entirely on density and drop type: dense residential parcel work commonly runs 15–25 stops per hour, suburban 10–18, rural 5–10. Benchmarks from a different density band are actively misleading — build your own by band.
Are lockers and pickup points worth it?
Where volumes justify them, yes: they convert many low-density stops into one dense stop and remove failure risk almost entirely. The economics hinge on the share of customers who will accept them, which varies enormously by market and by how the option is presented at checkout.
How late can we accept orders for next-day delivery?
Set by your slowest dependency, usually warehouse pick capacity rather than routing runtime. Modern engines can re-plan a large day in minutes; picking cannot. Measure the whole chain before promising a later cut-off.
Should drivers be able to resequence their own route?
Allowing limited resequencing within a time-window envelope generally improves both morale and performance, because drivers hold knowledge the model does not. Allowing unlimited resequencing destroys ETA accuracy. Most operations settle on constrained flexibility.
How do we handle the peak season?
Plan capacity in advance with scenario modelling on last year's peak volumes, recruit and train before the ramp, relax time-window offers during peak weeks, and expect service time to rise as inexperienced drivers join. Peak failures are almost always planned in October, not caused in December.
