Density is not a penalty to absorb. It is a modelling problem with a specific answer — parking, transport mode, and the last hundred feet on foot.
Routing models are built on assumptions that hold almost everywhere: the van can stop near the door, the driver walks a short distance, the next stop is a drive away. Tokyo and New York violate all three. There is nowhere to park, the last hundred feet are on foot or by bike, and one badly sequenced stop costs twenty minutes.
The industry response is to treat those markets as structurally expensive and move on. That is a defensible read of the constraint and a bad read of the opportunity: these are also the highest-volume, highest-density postal codes in the network, so a small per-route gain compounds faster here than anywhere else.
The bar was not a percentage. It was whether operations would run the routes the model produced in cities where drivers already knew better than the software.
Where a vehicle can actually stop — legally, at that hour — became part of route construction rather than something the driver resolves alone at the curb.
Bikes and on-foot legs stopped being exceptions handled offline and became modelled segments the optimizer could choose, with their own speed and capacity behaviour.
In a tower or a dense block the expensive decision is the order of doors, not the order of streets. The model sequences the last hundred feet explicitly.
Operations will not run a route they cannot interrogate. We shipped the reasoning alongside the recommendation and treated override rate as a product metric rather than a nuisance.
The mechanism was never the model. It was refusing to accept that a hard segment is structurally hard, naming the specific constraint the existing system ignores, and shipping the reasoning alongside the recommendation so the people doing the work would trust it.
That is the same sequence we run in a three-week thesis sprint and a ninety-day build pod — at a company with a fraction of the scale and none of the internal AI bench.