In short: A routing solver optimises against a snapshot, so a plan evaluated once on one week of orders then gets executed many times by drivers holding information the model never saw. Most of a modelled saving leaks away through service times that were assumed rather than measured, access restrictions the data does not carry, and orders arriving in a different pattern from the week that was solved. A master route trades some theoretical efficiency for consistency, which buys driver familiarity, predictable arrival times for customers, and a plan people will follow. Adherence is measurable from telematics by comparing the planned stop sequence against the actual one, and it separates a planning problem from an execution problem before anyone buys software.
A distribution business runs its first proper routing study. The software takes a week of real orders, the real fleet, the real depot, and comes back with a plan that runs eleven percent fewer miles than what is on the road today. The number survives a review. The board signs it off.
Nine months later the mileage report is showing about three percent. Two of the longest-serving drivers have gone. Customer service is fielding calls from accounts that used to be quiet, mostly about deliveries arriving at times nobody expected.
Nothing was wrong with the optimiser. The eleven percent was real inside the model. What the model did not price is that a plan evaluated once against a snapshot of one week gets executed fifty-two times by people who hold information the model never saw.
What the solver is actually solving
The underlying problem has been named since Dantzig and Ramser posed the truck dispatching problem in 1959: assign stops to vehicles and sequence them, minimising cost, subject to capacity. Everything since has been the business of adding the constraints that make it resemble an operation.
The ones that bite in practice are:
Time windows. Each stop is only servable inside an interval, and the intervals rarely nest neatly. This is the constraint that most changes the shape of the answer, and it is a large enough topic to belong elsewhere (X4).
Vehicle capacity in more than one dimension. Weight, pallet positions, cube, and often a chilled compartment that only holds so much. A plan that respects weight and violates pallet count is infeasible in the yard, not in the report.
Driver hours. Regulated rest and maximum duty are hard constraints with legal consequences, and they interact badly with time windows because waiting at a closed customer still consumes duty in most regimes.
Multi-day patterns. Most delivery businesses are not solving a single day. A customer visited twice a week has a set of admissible day pairs, and choosing the pattern is a decision that sits above sequencing. This is the periodic routing formulation, and it is the model that a master route actually corresponds to.
Compatibility. Vehicle size against site access, tail lift against no dock, driver certification against hazardous goods.
There is a second thing worth being clear about before looking at any percentage. The cost function most solvers report is distance or drive time, and the real cost function has a large step in it. A driver day and a vehicle day are close to fixed. Shaving nine percent off the miles without removing a vehicle from the plan saves fuel and tyres, which on a typical multi-drop operation is a minority of route cost. The savings that matter are the ones that drop a vehicle, and those are lumpy rather than proportional.
Where the eleven percent goes
Drivers who run the same territory every week get faster at it, and the mechanism is specific rather than sentimental. They know which gate the goods entrance is behind, which sites want the paperwork before the pallets, that the receiving clerk at one account takes lunch at half past twelve, and that a particular high street is impossible after nine. Zhong, Hall and Dessouky modelled this directly in 2007, treating driver learning within a territory as a first-class term rather than noise, and the routing literature has a whole branch on consistency: Groër, Golden and Wasil formalised the consistent vehicle routing problem in 2009, and Kovacs, Golden, Hartl and Parragh surveyed the family in 2014.
The arithmetic is worth doing on your own operation, because it usually settles the argument. Take a route with 40 stops, two hours ten minutes of driving, and eight minutes of average service time per stop. That is 130 minutes on the road and 320 minutes at the kerb. An eleven percent mileage saving is worth about fourteen minutes. If unfamiliarity adds ninety seconds per stop, and ninety seconds is a modest estimate for a driver working an area for the first time, that is sixty minutes. The saving is gone four times over before you count the errors.
The second leak is the customer side. A customer who knows their delivery is Tuesday morning staffs the goods-in bay on Tuesday morning, has the returns stacked, and has the previous delivery cleared off the dock. Take that away and service time rises for reasons that never appear in a routing model, because routing models treat service time as an attribute of the stop rather than a function of how predictable the arrival was.
The third leak is informational. Drivers on stable routes report things: a store that has quietly halved its order, a competitor's vehicle on the dock, a site being rebuilt. That channel closes when the route changes every day, and nobody notices it closing because there is no report for it.
What the master route is buying you
Set against the mileage, a fixed weekly route buys a set of things that are real and mostly unmeasured.
Customers get a known day, which lets them plan receiving and lets your sales team commit to a day without a routing check. Costs become forecastable at route level, so a variance against plan means something and can be investigated. Planner workload drops to maintenance rather than daily production, defence and re-explanation. Holiday cover and driver substitution get easier, because a documented route can be handed over and an emergent one cannot.
The honest counterweight is that a master route decays. Customers open, close, grow, shrink and change frequency, and a route built eighteen months ago is carrying the shape of a customer base that no longer exists. Left alone long enough, the familiarity advantage stops covering the accumulated inefficiency.
The pattern that survives contact
What works in most delivery operations is a fixed master route with a dynamic exception layer on top, and the design question is where the boundary sits.
The master layer gets rebuilt on a cadence rather than continuously. Quarterly is common, or on a trigger when a drift measure crosses a threshold you set in advance. The rebuild solves the periodic problem: visit frequencies, day patterns, and the sequence within each day, with the current master route supplied as a warm start so the solver has a reason to keep what already works.
The daily layer handles what the master cannot anticipate. New accounts that are not yet on a route, one-off volumes that will not fit, cancellations, a vehicle off the road, a site closed for stocktake. The design rule that keeps this from turning into daily dynamic routing by accident is that exceptions are inserted into the master route rather than triggering a fresh solve of the whole day. Cheapest-insertion against the existing sequence, with a feasibility check, gets you most of the value and leaves the rest of the route untouched.
Then a change threshold. Re-sequencing an existing route is only permitted when the projected saving clears a stated bar, something like thirty minutes of drive time or one vehicle removed. Below the bar, the plan stands. Without an explicit threshold, a solver will happily reorder six stops to save ninety seconds and hand you all the disruption cost for none of the benefit.
Measuring whether the plan is being followed
Most businesses that own a routing system cannot answer whether their routes are being run as planned, which makes every argument about routing quality unresolvable.
The measurement is not difficult. For each route, take the planned stop sequence and the actual sequence from telematics or the handheld, and compute:
Sequence adherence. The share of stops served in their planned position, or better, a rank correlation between planned and actual order so that a driver who shifts everything by one is not scored the same as one who inverted the route.
Unplanned events. Stops served that were not on the plan, and planned stops not served. These are usually the loudest signal and the easiest to explain.
Planned against actual drive time and distance, per route, with the two reported separately rather than netted.
Arrival deviation as a distribution, not a mean. The mean is close to useless here because early and late cancel.
The diagnosis comes from reading adherence against actual cost together. Low adherence with actual mileage below plan means the drivers are finding something the model does not know, and the model inputs are wrong. Low adherence with actual mileage above plan is an execution or discipline problem. High adherence with actual times consistently above plan points at service time estimates. Each of those leads somewhere different, and without the split you get an argument instead of a diagnosis.
Related but separate: if the pattern of deviation is geographic rather than sequential, the issue is likely to be how accounts are grouped in the first place, which is territory design (X2) rather than sequencing.
The limit: service times nobody has measured
Everything above assumes the optimiser has decent inputs. The input that is almost always weakest is service time, and it is also the one that dominates the day.
In the example above, service was 320 of 450 working minutes. A twenty percent error in the service time estimate is 64 minutes, roughly half the total driving time, and about four times any plausible sequencing gain. Most businesses carry a single flat service time, or a per-customer-type figure someone set when the system was installed, and have never validated either. An optimiser fed that number produces a plan that is precise about the wrong quantity, and the resulting timings are wrong in a way that looks like driver non-compliance.
Measuring it is available to anyone with telematics. Take door-open to door-close, or arrival geofence to departure geofence, join it to the order, and you have a service time observation per stop. Fit it against drop size, delivery type, and site characteristics rather than assuming a constant, and validate the fit on held-out time periods rather than in-sample, because a service time model tuned on the same weeks it was fitted on will flatter itself.
There is a harder version of the limit underneath that. Even a well-measured service time is a distribution with a long right tail, and a routing plan built on the mean will run late on the days the tail shows up. That is tolerable when the promise to the customer is a day, and it becomes the central problem when the promise is a narrow window, which is why window width is a commercial decision before it is a routing one.
If you want one thing to do this week, pull last month's telematics and compare the planned stop sequence against the actual one on your twenty busiest routes. The adherence number tells you whether you have a planning problem or an execution problem, and it costs nothing to find out before anyone buys software to fix the wrong one.