Route planning that respects the driver

An early version of a route planner we shipped was measurably better than the drivers at minimising distance. Within a fortnight most of them had stopped opening it. The optimiser was correct about the thing it was measuring, and the thing it was measuring was not the job.

The objective function is never distance

A driver is optimising for finishing at a time that gets them home, avoiding the stretch that floods after rain, avoiding a right turn across a dual carriageway with a loaded van, parking somewhere that exists, a customer who will not see anyone before eleven, a site where the gate is locked over lunch, and somewhere to eat that is not a service station. Some of that is on the map. Most of it is not, and all of it is real.

An optimiser without those constraints produces a plan that is shorter and unusable. Rejecting it is the rational response, and the drivers who reject it fastest are usually the best ones, because they have the most local knowledge to contradict it with.

Suggest, and record the disagreement

The change that worked was demoting the plan to a suggestion. The order is reorderable, and when a driver reorders it they pick a reason from a short list: access, customer availability, traffic they know about, a delivery window, something else.

Those reasons became the most useful dataset in the programme. Interviewing drivers about their constraints up front had produced very little, because people are poor at listing constraints in the abstract and excellent at naming one the moment it has just cost them twenty minutes. After a few months most of the constraints were in the model, put there by the people who knew them, at the moment they were relevant.

Show the arithmetic

A reordered list with no explanation is an instruction from a system to a professional who has been doing the route for six years. "This order saves twenty-two minutes and still makes the eleven o'clock" is an argument. People accept arguments from software and resent instructions from it, and the engineering cost of showing the reasoning is small compared with the cost of the feature being ignored.

One related rule: never silently re-optimise mid-route. If conditions change enough to justify a different order, ask. A plan that rearranges itself while somebody is driving to a stop is worse than a plan that is slightly wrong.

Do not grade adherence

The moment adherence to the plan becomes a number drivers are measured on, adherence goes up and the information stops. Nobody records a reason for a deviation they are going to be marked down for, so the model stops learning, the plans get worse in ways that never surface, and it takes about a year for anyone to connect the two. The metric worth watching is whether the day finished on time, which is what the business wanted in the first place.

The same tension runs through geolocation-verified visit tracking, which we built into the Würth field CRM. Verification is a legitimate requirement. It is also, from the seat of the person being verified, a system that watches them. Whether it produces good data or elaborate workarounds is decided almost entirely by whether the same platform gives them something back on the same screen.

It gets harder when the driver is not the beneficiary

In field sales the representative gains directly from a better route: more calls, less driving, an earlier finish. In distribution and logistics, which is the world our clients in that sector operate in, Mahindra Logistics among them, the person holding the device is often not the person whose costs the optimisation reduces. The incentive to cooperate with it has to come from somewhere else. That is a design problem rather than a training problem, and it is generally solved by making the plan carry things the driver wants: accurate site access notes, contact numbers that work, a realistic break.

A plan followed most days beats an optimal plan followed for a fortnight. It is an unsatisfying thing to tell an operations research team, and it is most of what we have learned about this.