The challenge
The organisation faced a routing economics problem with a structural cause: the legacy routing engine had been built for a delivery network that no longer existed. When 80% of volume was business deliveries on predictable rounds, the system worked. Now the same network ran majority residential, with cash-on-delivery windows, customer reschedule requests, and a mixed fleet of vans, two-wheelers, and gig drivers. Each operational change had been patched onto the previous one until no one could explain why a particular route had been chosen or how to improve it systematically.
Discovery
We spent the first four weeks across three depots — Riyadh, Jeddah, Dammam — working alongside drivers and watching dispatch in detail, then reconstructing what actually happened on each route against what the system had planned. The variance was the finding: drivers were silently rerouting around 31% of stops because the system had no knowledge of parking constraints, building access requirements, or the customer time windows the call centre had renegotiated that morning. The routing problem was an information problem.
Our approach
Routing architecture redesign: replaced the static routing engine with a learning optimiser that ingests live traffic, parking-availability data, customer reschedule events, and per-driver historical performance
Dispatch intelligence build: designed and built a single dispatch surface letting the depot supervisor see the optimiser's plan, driver-suggested deviations, and the cost-of-change for any manual override
Continuous learning programme: built a model-evaluation pipeline that replays each completed day against the plan and feeds the variance back into the next training cycle — closing the feedback loop the legacy system had never had
Phased rollout with change management: rolled out depot by depot with a 21-day stabilisation window in each, so depot managers could develop confidence in the system before committing to it operationally
The Outcome
The engagement delivered a routing platform that improves week over week from the depot's own delivery data, with a dispatch surface that incorporates the operational knowledge of drivers and supervisors rather than overruling it. Adoption was built through visibility, not mandate — supervisors could see the cost consequence of every change they considered making.
The results
“The old system told the drivers what to do. The new one works with them. That is what changed — and the operating numbers followed.”
