Route optimisation in road freight: more deliveries, fewer kilometres

A delivery round with 15 stops can be driven as 653,837,184,000 different round trips. Every morning, a planner picks one, usually with experience, a map and a spreadsheet. Good planners know their roads, their clients and their drivers, and their routes are rarely bad. They are also rarely the shortest, because nobody can compare more than a handful of the alternatives.

Route optimisation is software that searches those alternatives for the plan with the fewest wasted kilometres, inside the rules a real day has to respect: delivery windows, vehicle capacity and driving time. This article goes deeper into what we showed in AI Xplained EP.02. It starts with what a kilometre costs a European fleet, explains why the problem is too big to solve by hand and how the software solves it, then looks at the evidence, what a project needs, where projects fail, what the law says and how to start with one depot and one week.

What a kilometre costs a European fleet

Most of the cost of a delivery truck moves with the hours it spends on the road and the distance it drives. In the French reference cost index for regional rigid trucks, the kind of vehicle that runs many delivery rounds, the driver accounts for 36.8% of the total, fuel for 16.3%, maintenance for 7.1%, travel allowances for 3.6% and tolls and road taxes for 1.4% (CNR, 2026). Together, that is 65% of the cost. The vehicle itself and the company’s overheads make up the rest.

The order matters. In distribution, the driver costs more than twice as much as the diesel, so a shorter round saves hours first and fuel second. Both are getting harder to buy:

• Drivers. Europe has more than 502,000 unfilled truck driver jobs, 13 in every 100 positions, and 660,500 drivers are due to retire within five years (IRU, 2026). An hour a route wastes is an hour of a driver the sector does not have.
• Fuel. By mid-March 2026, IRU’s monitoring showed EU diesel prices up 20%. The second EU emissions trading system (ETS2), which puts a carbon price on road fuels from 2028, could add about 10% to the price of diesel at €50 to €100 per tonne of CO2 (IRU, 2026).
• Emissions. Heavy-duty vehicles produce more than a quarter of the EU’s greenhouse gas emissions from road transport and over 6% of all EU emissions (European Commission, 2026).

Then there is the distance driven with nothing on board. Around a fifth of the kilometres driven by EU-registered goods vehicles are run empty, 21.6% in 2024 (Eurostat). Most of that is return legs and repositioning, which is a freight-matching problem. Route optimisation attacks a different waste: the extra kilometres driven inside delivery rounds, with the goods on board, because the stops are in the wrong order.

A simple formula puts a number on the routing part:

Kilometres driven per year × share of kilometres saved × variable cost per kilometre = annual saving

As an illustration, a depot running 15 vehicles that each drive 60,000 km a year covers 900,000 km. Saving 5% of that distance, a conservative assumption next to the published results below, removes 45,000 km, three quarters of what one vehicle drives in a year. At a variable cost of €0.80 per kilometre, that is €36,000 a year. The figures are an example, not a benchmark. Use your own variable cost, not the full cost per kilometre: the vehicle and the overheads are paid whether the route is short or long.

Why nobody can check every route

The number of possible routes explodes with every stop. For a single vehicle leaving a depot, visiting every stop once and coming back, the count of different round trips is n!/2, where n is the number of stops, counting each round and its reverse as one:

• 5 stops: 60 possible routes.
• 10 stops: 1,814,400.
• 15 stops: 653,837,184,000.
• 20 stops: about 1.2 × 10¹⁸, or 1.2 billion billion.

When direction matters, as it does with delivery windows, the count doubles.

A computer checking a billion routes per second would need about 11 minutes to try every round of 15 stops, and almost 39 years for 20. A real depot does not plan one vehicle with 20 stops. It plans several vehicles with hundreds of stops, and every stop can go into any of them.

Then come the rules that make a plan usable: a client who only receives between 8:00 and 10:00, a rigid truck that is full at 18 pallets, a street closed to trucks over 12 tonnes, a low-emission zone. And in Europe, the driving-time rules: daily driving may not exceed 9 hours, extendable to 10 hours twice a week; weekly driving is capped at 56 hours and fortnightly driving at 90; a break of at least 45 minutes is due after four and a half hours, and the daily rest is normally at least 11 hours (Regulation (EC) No 561/2006). Each rule removes routes from the list, but it also makes the remaining ones harder to find.

This is why exact methods, which prove that a plan is the best possible one, have long been limited to small cases. In 2007, researchers at SINTEF, the Norwegian research institute, wrote that exact methods could not consistently solve problems with more than 50 to 100 customers in reasonable time, and that approximation methods were the only viable approach at industrial scale (Hasle and Kloster, 2007). Exact methods have improved since, but daily planning with time windows and driving rules still relies mainly on heuristic search.

How route optimisation works

The software never tries every route. It searches in four steps.

1. Read the day. Orders, addresses, delivery windows, quantities, vehicles and their capacity, drivers and their shifts. It also needs travel times between every pair of stops, ideally by time of day, because 20 minutes at 6:00 can be 45 at 8:30.

2. Build a first plan. A fast method assigns each stop to a vehicle and an order, for example by inserting each stop where it adds the least distance. The first plan is feasible but rarely good.

3. Improve it, move by move. The software tries small changes and keeps the ones that shorten the plan without breaking a rule: reverse a stretch of a route, move a stop to another vehicle, swap two stops between routes. Better methods also accept a worse plan now and then, to escape a solution that only looks good because every small change makes it worse. It stops when the time limit runs out or nothing improves. The result is a very good plan, without proof that it is the best possible one, and it arrives in seconds or minutes.

4. Re-plan when the day changes. When traffic, a new order or a cancellation changes the day, the software re-plans what is left of it, from where each truck is.

The search itself comes from operations research, the mathematics of planning. Machine learning earns its place in the inputs: it learns real travel times by road and hour from the GPS history the trucks already log, and how long each client really takes to unload. A perfect search on wrong travel times produces a plan that fails by 10:00. Transport also lags on AI in general: in 2025, 11% of EU transport and storage companies with 10 or more employees used at least one AI technology, against 20% of all companies (Eurostat, 2025).

The last step is human. The planner checks the plan, overrides it where they know something the data does not, and signs off. We wrote about designing those approvals without slowing the operation in Human in the loop.

What the evidence says

The research and the case studies point the same way, with one caveat: the gain depends on how good the current plan already is.

• The reference range. In the reference book on the subject, Paolo Toth and Daniele Vigo of the University of Bologna report that computerised planning of distribution generally saves 5% to 20% of total transport costs (Toth and Vigo, 2002).
• A European distributor. A Croatian distributor of fast-moving consumer goods, with 50 to 60 vehicles and around 1,400 orders a day, used a delivery optimisation system in daily planning for more than four years. It lowered overall distribution costs, replaced four of its eight depots with satellite points that hold no stock and took on more customer orders without adding vehicles. When it moved to fixed weekly delivery days by area, a simulation on a month of real orders projected 26% less distance than the routes actually driven, 15% fewer vehicle tours (596 to 506) and 78% less overtime, the last partly because the dispatcher relaxed working-time limits (Galić, Carić and Fosin, 2013). Those figures come from a simulation, not from measured results, and they combine a new delivery model with the optimisation.
• Newspaper distribution. SINTEF researchers reported 2% to 25% shorter total route duration on industrial newspaper-distribution cases, in conference slides that do not state the baseline (Hasle, Kloster and Smedsrud, 2010).

In our experience, dense rounds with many stops and tight windows gain the most, because that is where human planning hits its limits. A truck with two drops a day gains little.

What it needs: three things

Your orders. Stops, addresses, windows and quantities, from the order system or the TMS the team already uses.

The GPS history your trucks already log. Telematics and tachograph data give real travel times by road and by hour, and real service times by client. European law now helps on both counts. Since 19 August 2025, heavy-duty vehicles registered in the EU and operating in another member state must carry the second version of the smart tachograph, which also records border crossings and loading and unloading operations; light commercial vehicles over 2.5 tonnes in international transport follow from 1 July 2026 (European Commission, 2025). And since 12 September 2025, the Data Act gives users of connected products, trucks included, the right to access the data their vehicles generate and to share it with a third party of their choice (Regulation (EU) 2023/2854). Truck makers such as Volvo Trucks now publish how fleets can request it, through their telematics platforms or the standard FMS interface (Volvo Trucks, 2025). A fleet can get its own GPS history even when it sits on a supplier’s platform. Where that data identifies drivers, the GDPR still applies.

A planner who signs off every route. The optimiser proposes, the planner decides. Their knowledge of clients and roads becomes the rules that make the next plan better.

Where it goes wrong

In our experience, when projects fail, the cause is one of those three inputs, rarely the algorithm.

• Wrong addresses. A stop placed 2 km from the real gate turns a good plan into a bad one, and many order systems were never built for geocoding.
• Guessed service times. If every client is planned at 10 minutes, the client who always takes 40 breaks every route after it.
• Rules nobody wrote down. The dock that does not fit an articulated truck, the client who only opens the back gate. Planners keep this knowledge in their heads, and the optimiser ignores what it is not told.
• Overrides that teach nothing. A planner who changes a route without recording why fixes today and repeats the problem tomorrow. Every override should become a rule or a correction in the data.

The data does not need to be perfect to start, as we explained in Your data doesn’t need to be perfect, but these four gaps need closing early.

The legal side

Driving time is a hard limit inside the plan, as described above. The other rule to check is the AI Act. Planning routes, stops and vehicles is not on its list of high-risk uses. The line is the driver. An AI system that allocates tasks based on individual behaviour or personal traits, or that monitors and evaluates the performance and behaviour of workers, is high-risk (Regulation (EU) 2024/1689, Annex III). A route optimiser that plans stops within legal driving limits sits outside that category. A system that also scores drivers and assigns them work based on that score sits inside it. In between there is a grey zone: an optimiser that assigns routes to named drivers using what it has learned about each of them, such as their speed or their service times, may count as allocating tasks based on individual behaviour. Under the changes to the AI Act agreed in 2026, the obligations for these high-risk systems are due to apply from 2 December 2027. GPS monitoring of drivers is also limited by the GDPR and by national labour law. Specific cases should still be checked with legal advisers. Our article on the AI Act for operations leaders covers the rest.

How to start: one depot, one week

Day 0. Pick the depot and set the baseline. Choose the one with the most stops per day. Record the current kilometres per stop, stops per route, on-time deliveries and overtime hours. Without a baseline there is no result. Setting that measure before the project starts is the subject of How to measure the return on an AI project before you start.

Week 1. Plan in parallel. The planners plan as usual and the trucks run their normal routes. The optimiser plans the same orders on the side. At the end of each day, compare the two plans: kilometres, vehicles, hours and missed windows. Nothing changes on the road yet.

Weeks 2 to 12. Run it and measure. The planners start from the optimised plan and adjust it. Every adjustment becomes a rule or a correction to the data. Compare the results with the baseline every week.

Back to the example: a depot driving 900,000 km a year that saves 5% removes 45,000 km and €36,000 in variable costs a year. One week in parallel will not prove the full year. Its job is to show, with the depot’s own orders, how far the current plan is from the best one the software can find.

More deliveries, fewer kilometres

Route optimisation is a mature discipline. What has changed is everything around it: drivers are harder to find, fuel and carbon are getting more expensive, and European law now records where trucks in international transport load and unload and gives fleets the right to their own data. The hours and kilometres a better plan removes can be cut from the bill or used for more deliveries with the same fleet.