The Last Kilometre Problem: AI Routing for Roads Without Addresses

The practical approach is to capture verified drop-point coordinates on every successful delivery, cluster them into a proprietary landmark graph, and let rider behaviour correct the map continuously. After a few thousand deliveries an operator has something no mapping vendor sells: a ground-truth address book for its own customer base.

This is a compounding asset. It gets more accurate with volume, it is specific to the areas the business actually serves, and it is very difficult for a new entrant to replicate quickly.

Pure distance minimisation produces routes riders refuse, long unpaid repositioning legs, drop sequences that end far from the next likely pickup, assignments through areas riders avoid after dark. A route a rider declines is not an optimisation; it is a failed dispatch and an unhappy customer.

Dispatch models therefore have to include earnings per hour, repositioning cost, and acceptance probability as first-class objectives. Operators that model rider behaviour explicitly see acceptance rates rise without raising per-trip pay.

For freight, delay rather than distance is the real cost. Document classification, pre-clearance preparation, and predicting inspection likelihood can remove days at Malaba, Busia, or Beitbridge. A truck sitting for three days at a border burns driver cost, capital, and customer patience in a way no route optimisation recovers.

This work is document-heavy and unglamorous, extracting and validating fields from certificates, manifests, and permits, and it has among the clearest returns in African logistics.

Most operators discover, when they finally attempt modelling, that they never recorded actual drop coordinates, actual arrival times, or the reason a delivery failed. Those three fields are the foundation of everything above, and adding them costs a sprint of app work rather than a data science programme.