AI in Logistics: Route and Fleet Optimization

In many distribution companies the next day's plan is settled the evening before, on paper: the dispatcher looks at the orders, splits them across a few vehicles from experience, and fixes the order of stops. The method works and has worked for years; the problem lies elsewhere. With twenty orders and three vehicles, the number of possible combinations exceeds what anyone can compare mentally, and what the mind selects is an acceptable plan, not the best one.

The gap between acceptable and best is where the margin hides; and because that gap appears under no accounting heading, nobody counts it as a cost. Extra kilometres, idle vehicle hours and a failed delivery that must be repeated all dissolve into general expense lines.

Three layers that get confused

• Stop sequencing: for a given set of destinations, which order is shortest or cheapest.
• Assignment and loading: which order goes on which vehicle so that weight and volume capacity fill better.
• The delivery promise: what time window to give the customer so that it is both precise and achievable.

Each layer is solved with different tools and their economics differ. For small fleets the second layer usually yields the largest saving and the third the largest effect on customer satisfaction.

The data you need to start

The obstacle here is not the algorithm; it is basic data. The first step needs three things: destination addresses as coordinates rather than text, the real dwell time at each stop, and the actual capacity of each vehicle. The hardest of the three is dwell time; most companies never record it, even though in urban distribution it consumes more time than driving. A plan that underestimates dwell time looks elegant on paper and is useless on the street.

The person who executes the plan

No optimal plan survives without the driver's cooperation. A driver who finds the proposed sequence illogical reverts to their own method, and you never learn why the outcome missed the forecast. Two things fix this: explaining the basis of the proposed sequence, and building a channel for the driver to object. Most of the time the objection turns out to be sound and a constraint was missing from the plan; adding it improves accuracy next time.

The reality of logistics in Iran

Several features of the Iranian market make the calculation differ from foreign examples. Addresses are inconsistent, and converting them to coordinates is a work stage in itself. Shifting fuel prices and vehicle maintenance costs move the weight of the cost components. Traffic restrictions in city centres and permitted freight hours are constraints that, if left out of the plan, make the output unusable. In exchange, these same difficulties mean the gap between a manual plan and a computed one is wider in Iran than average.

A practical example

Consider a company distributing daily with a few vehicles in one city, whose customers complained less about late arrival than about not knowing when the goods would arrive. Before any route optimization, only real dwell times and delivery sequences were recorded for several weeks. That data alone showed two things: two routes were always loaded beyond a day's capacity, and the time window announced to customers had been optimistic from the outset. Correcting those two removed a large share of complaints before any new tool was bought.

A 90-day starting path

• First month: choose one distribution hub and a few fixed routes, and record destination coordinates, dwell times and vehicle capacities.
• Second month: run the computed plan beside the manual one; keep both for several days and judge the difference on cost per delivery.
• Third month: if the number improved, add the constraints the drivers raised to the model and extend the scope to further routes.

Common mistakes

• Buying a comprehensive fleet system before holding basic data for a single route; the result is a subscription for a capability nothing feeds.
• Optimizing for the shortest distance when the real cost sits in time and stops.
• Ignoring real constraints such as permitted freight hours or a customer's receiving window; an unusable plan is worse than a weak one.
• Measuring success in kilometres driven instead of cost per delivery and first-attempt delivery rate.

Frequently asked questions

• How many vehicles justify it?
  The few you already run. The issue is not vehicle count but daily stop count; beyond roughly twenty stops, mental calculation exceeds even the best planner.
• Can we start without trackers?
  Yes, by logging delivery times by hand for a few weeks. Trackers make the work more precise but are not a prerequisite.
• Does it replace the human planner?
  No; that person moves from splitting orders to adjudicating exceptions, which is where their experience is worth most.

Takeaway

Route optimization is not a technology project; it is the disciplined measurement of something that has gone unmeasured. Start with one hub, a few routes and three simple data items, and know your cost per delivery before buying any system.

Glossary

• Route optimization: finding the cheapest travel sequence for a set of destinations.
• Geocoding: converting a text address into latitude and longitude; a prerequisite for any route calculation.
• Time window: the interval within which a delivery must be made.
• Cost per delivery: total fuel, wages and depreciation divided by the number of successful deliveries.
• First-attempt delivery: the share of orders delivered on the first visit.