The last mile is a people problem
If you instrument a last-mile operation properly, the distribution of lost time is rarely where people expect.
The routing is usually fine. Modern engines produce sensible sequences. The losses cluster in two places: the loading bay before departure, and the last hundred metres at each stop.
Before the van leaves
Loading is where plans meet physical reality. The route says twenty-eight drops in a sequence; the van gets loaded in whatever order the parcels came off the sorter. Every misordered parcel is a driver climbing over stock at some point in the day. Nobody measures this because it doesn’t appear in any system — it shows up later, diffused across the day, as “traffic”.
The last hundred metres
The other loss is access. Gate codes, service entrances, which of three buildings actually takes deliveries, whether reception signs for parcels after four. Experienced drivers carry this in their heads. It’s the single most valuable dataset in the operation and it’s almost never captured, which is why route quality collapses when a regular driver is off.
Neither problem is a routing problem, and neither is solved by a better solver. The first is a sequencing handoff between warehouse and transport. The second is a data capture problem that most operations could fix in a quarter if anyone owned it.
Optimise the plan all you like. If the van is loaded wrong and the driver doesn’t know the gate code, you’ve optimised the wrong thing.