The situation
The business runs 12 distribution centres across the UK, moving stock between suppliers, regional hubs and a long tail of trade and retail customers. By the time we were brought in, every site was planning its routes the same way it had for years. A small team at each centre worked through the next day's drops in a spreadsheet, checked them against the transport management system, then handed the result to the dispatchers. On a normal day this took about six hours per site. On a bad day, when a major customer reshuffled an order or a vehicle went off the road, it took longer and other work simply did not happen.
Inventory was a similar story. Stock decisions were made when something ran short, not before. Buyers reacted to depot calls and sales spikes rather than working from a forward view, which meant the network carried a lot of safety stock in the wrong places and still ran out in the right ones. The ops director had been tracking two numbers for about eighteen months. The first was planner overtime, which was climbing every quarter. The second was the cost of expedited inter-depot transfers, which had quietly become one of the largest controllable lines on the operations P and L. Neither number was going in a direction anyone was happy about, and headcount was not an option.
What we did
We started with the data the business already had. Three years of order history, depot-level stock movements, vehicle telematics and customer delivery windows were all sitting in different systems and nobody had ever pulled them into one place. We built a forecasting model on top of that combined dataset, tuned per depot rather than per network, because demand patterns in the south west looked nothing like demand patterns in the central belt. The model produced a rolling fourteen-day view of expected order volume by SKU and region, which the buying team used to pre-position stock instead of chasing it.
The route optimisation piece was layered onto the existing transport management system rather than replacing it. The business had years of configuration in there and ripping it out would have been reckless. Instead, the optimiser pulls the day's orders, vehicle availability and driver hours, generates a planned set of routes, and pushes them back into the TMS for the planners to review. We were deliberate about where the humans stayed in the loop. Planners still sign off every route before it goes to a driver, still handle the awkward customers the model does not know about, and still make the call when something goes wrong on the road. The model does the arithmetic. The planners do the judgement.
We also spent a fair amount of time on the unglamorous part, which was getting the data clean enough to trust. About a third of the project was reconciling SKU codes between depots and fixing addresses that had been wrong in the customer master for years.
The result
Daily planning time per depot dropped from around six hours to about forty-five minutes. Across 12 sites that gave the planning team back the equivalent of several full-time roles, which were redeployed onto customer service work that had been getting dropped. Annualised savings across the network came to roughly £840,000, with the largest contributions coming from reduced expedited transfers, lower overtime and a meaningful drop in empty running miles. Planning efficiency, measured against the operations team's own internal benchmark, improved by 42 per cent.
The honest operational observation is this. The numbers are real, but the bigger change was cultural. The planners stopped firefighting and started looking a week ahead, and the ops director stopped getting calls at seven in the evening about a depot that had run out of something. The model is not perfect and it never will be. It is wrong often enough that the planners still need to pay attention, which is exactly how the business wanted it.