Why couriers ignore the optimal route — and what 1.4 million deliveries taught us

Every dispatch system ships with an article of faith: the optimizer knows best. Feed it the stops, the traffic, the time windows, and it returns the route that minimizes total travel time. The courier’s job is to follow it. When they don’t, we call it deviation — a polite word for “the human got it wrong.”

A study of 1.4 million package deliveries in Shanghai suggests we’ve had the error backwards. The couriers aren’t failing to follow the optimum. They’re optimizing for something the algorithm wasn’t.

The finding

The researchers had the ground truth most route teams never see: not the route the engine recommended, but the route the courier actually drove, across more than a million real deliveries. The pattern was consistent and a little uncomfortable: couriers generally favor shorter distances over time-efficient routes. Given a choice between the stop that’s closer and the stop that the model says keeps the overall route faster, they go to the closer one.

The pull toward proximity gets stronger exactly where you’d hope the algorithm would take over — “under complex routing conditions or during peak traffic hours.” When the situation is hardest, couriers trust what they can see (the next stop is right there) over what the model asserts (trust me, the longer hop pays off later).

And here’s the part that complicates the easy conclusion: these deviations do cost something. Actual routes and travel times came out longer than the predicted optimum, and the gap was worst for inexperienced drivers and on long routes. So the human isn’t simply right and the machine wrong. Both are partly right, and the interesting engineering lives in the gap between them.

Why proximity wins

Step into the courier’s seat and the preference stops looking irrational. The time-optimal route is a claim about the future — “skip this near stop now, and the sequence nets out faster by the end of the shift.” It depends on traffic behaving as predicted, on no stop running long, on the model’s travel-time estimates holding. Every one of those is a small bet.

The near stop, by contrast, is a certainty. It’s there. Going to it banks guaranteed progress and removes one variable from a shift full of them. Couriers, especially under pressure, are doing what any sensible agent does under uncertainty: discounting the model’s confident prediction and taking the sure thing. That’s not a failure of discipline. It’s risk management with a steering wheel.

The cost data tells us the model isn’t wrong — its optimum really is faster when the assumptions hold. The behavior tells us the assumptions don’t hold often enough for couriers to trust them. Both facts are true at once.

The wrong fix and the right one

The reflex is to push harder on compliance: tighter routing, stricter adherence metrics, nudges to follow the line. That treats the deviation as the problem. But if a million couriers reliably make the same “mistake,” the cheaper hypothesis is that the route, not the courier, is mis-specified.

The study’s own conclusion points the other way — toward a need to realign routing algorithms with courier preferences. Not to surrender to the human, but to fold the human’s information back into the model. Concretely, that means a few things for anyone building dispatch:

Penalize fragile optima. A route whose advantage evaporates if one assumption slips is worth less than its predicted time suggests. Weighting toward robust routes — ones that stay good when traffic misbehaves — closes much of the gap to what couriers actually do, because robustness is exactly what they’re chasing.

Treat proximity as a feature, not a bug. If couriers consistently value the near stop, the model should price that preference in rather than fight it — and reserve its “trust me” hops for the cases where the payoff is large and the assumptions are solid.

Segment by experience. The deviation cost concentrated among inexperienced drivers and long routes. That’s a targeting signal: the people and routes where guidance helps most are identifiable in advance, so support and tighter sequencing can go where they actually pay off, instead of being sprayed across everyone.

The broader lesson

There’s a pattern here that goes well beyond last-mile routing. We tend to deploy optimization as an oracle — it computes the answer, humans execute it — and we treat any divergence as human error to be stamped out. But the human on the ground is running a model too, one trained on conditions the algorithm never fully sees. When their model and ours disagree a million times in the same direction, that’s not noise to suppress. It’s a free, labeled dataset telling us where our model is wrong.

The best dispatch systems won’t be the ones that force couriers onto the line. They’ll be the ones humble enough to ask why the line keeps getting ignored — and to update.