What problem does it solve?
A delivery fleet has more possible routes than a planner can compare. Every extra mile costs fuel, driver time and vehicle wear, and a route that runs late means missed delivery windows, failed first attempts and extra stops the next day. Manual planning also depends on the person who knows the area, which makes results uneven between depots.
Two operators on this page describe their starting point. ArrowXL's chief information officer says the company outgrew its previous planning technology because its outputs no longer gave logical, valuable delivery insights to support growth. New Castle Building Products' chief operating officer says route planning was manual and decentralized, which made it difficult to optimize deliveries consistently across locations. At UPS scale the stakes are larger: a UPS executive told investors in 2021 that every mile saved per driver in a year, across all its US drivers, saves the company more than USD 60 million a year.
How does it work?
- Collect the day's work. Orders, stop locations, time windows, service times, vehicle types, capacities and driver hours come from the order or transport management system.
- Estimate travel and stop times. The optimizer needs a drive time for each road and hour and a service time for each stop. Many systems take these from mapping data and fixed settings. Amazon says its AI based routing adjusts for conditions such as reduced visibility after sunset and the extra time a downtown delivery needs to find parking at lunchtime. Other teams train a model on the fleet's own telematics, a design choice no operator on this page documents.
- Build the routes. An optimizer assigns stops to vehicles and orders them to minimize miles and lateness within the constraints. Walmart describes its tool as considering time, location and store delivery windows, packing trailers efficiently and planning backhauls so trailers are not empty on the return trip.
- Review and release. Planners check exceptions, such as stops the optimizer could not fit, and release the plan to drivers' devices. ArrowXL says its system now plans 95% of routes overnight.
- Replan during the day. Newer systems, such as UPS's Dynamic ORION, reoptimize a route as conditions change and give the driver turn by turn directions. Amazon says its driver navigation now adjusts routes automatically in bad weather or when hazards are detected.
- Review execution. Actual drive and stop times, early route terminations and failed deliveries are compared with the plan and used to correct the travel and service time settings.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Mainstream
- Channels
- Internal tools, Mobile app, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Lower cost to serve, Customer experience, Employee productivity, Risk and loss reduction.
Indicative value
A regional delivery fleet with 250 vehicles
USD 60,000 to USD 675,000
Mileage related cost avoided per year
How this is calculated
Formula: vehicles * milesPerVehicle * mileageReduction * costPerMile. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Delivery vehicles vehicles, vehicles | 250 | 250 | The reference fleet. ArrowXL by JP Home describes its fleet as approximately 250 vehicles. Source |
| Miles driven per vehicle per year milesPerVehicle, miles per vehicle per year | 20,000 | 30,000 | Editorial assumption for urban and regional delivery vans and trucks. Replace with your own telematics data. |
| Reduction in miles driven mileageReduction, fraction of miles | 0.03 | 0.1 | Editorial assumption, below the approximately 13% fleet mileage reduction that Descartes reports for ArrowXL, which it credits to better planning and less manual work together. |
| Mileage related cost per mile costPerMile, USD per mile | 0.4 | 0.9 | Editorial assumption in US dollars for vans and light trucks, covering fuel, tires and maintenance only. Heavy trucks cost more per mile. Replace with your own cost per mile in your own currency. |
What it leaves out: Counts fuel, tire and maintenance cost avoided on the miles removed. It leaves out driver hours saved, more stops per route without more vehicles, fewer failed first attempts, lower emissions and the license, integration and change costs of the routing system.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Amazon
United States · Retail and ecommerce · 2026
Amazon's vice president of transportation says in September 2026 that Amazon is investing in AI based routing technology for the drivers of its Delivery Service Partner network. She says the routing adjusts for real world conditions, such as reduced visibility after sunset or the extra time a downtown delivery needs to find parking at lunchtime, and that driver navigation now adjusts routes automatically in bad weather or when hazards are detected.
A September 2025 post from the same executive says Amazon uses AI for delivery quality and safety through routing technology that analyzes traffic patterns and road conditions to design safer routes. Neither post gives a mileage, time or cost result for the routing, says how many routes it covers, or describes the models.
No outcome disclosed.
Walmart
United States · Retail and ecommerce · 2024
Walmart built an AI driven route optimization system for its own trucks and trailers. It maps multi stop journeys against delivery windows, packs trailers, and plans backhauls so trailers return loaded. In March 2024 Walmart said it avoided 94 million pounds of CO2 by removing 30 million unnecessary miles and routing around 110,000 inefficient paths, and that it won the 2023 Franz Edelman Award for deploying the technology at scale. The same release announced that Walmart Commerce Technologies would offer the software to other businesses as a service. The release gives mileage and emissions figures but no cost, time or service level result.
No outcome disclosed.
United Parcel Service (UPS)
United States · Logistics and transportation · 2021
UPS runs ORION (On Road Integrated Optimization and Navigation), its in house route optimization system for delivery drivers. At its June 2021 Investor Day, UPS's chief information and engineering officer described Dynamic ORION, a newer version that reoptimizes a driver's route as conditions change and gives turn by turn directions.
He said initial results pointed to a further reduction of 2 to 4 miles per driver per day, in addition to the 8 miles per driver per day from the original ORION deployment. He then said the combined effect amounts to a reduction of more than 130 million miles driven and 10 million gallons of fuel saved each year.
Another UPS executive said that every mile saved per driver across all US drivers saves UPS more than USD 60 million a year. UPS does not call ORION AI in this transcript, so this record shows route optimization only.
No outcome disclosed.
ArrowXL by JP Home
United Kingdom · Logistics and transportation · 2026
ArrowXL by JP Home is a UK home delivery and warehousing specialist that delivers large items such as washing machines and cookers with two person crews from a fleet of about 250 vehicles. Descartes says ArrowXL cut fleet mileage by approximately 13% using the route planning and execution tools in its fleet performance management suite. ArrowXL's chief information officer says the system now plans 95% of routes overnight, that early route terminations fell from 5% to 1%, and that deliveries outside customers' preferred time windows fell by 4%. A separate Descartes case study page lists the same 13% mileage cut alongside a 5.6% drop in average miles per route.
No outcome disclosed.
New Castle Building Products
United States · Logistics and transportation · 2026
New Castle Building Products is a family owned distributor of roofing and exterior building materials with 24 locations in six northeastern US states. Its chief operating officer says route planning used to be manual and decentralized, which made it hard to optimize deliveries consistently across locations. With Descartes' route planning and execution platform the company centralized dispatch and says it has taken approximately 25,000 miles out of its routes each year while improving delivery reliability. Descartes describes the platform as optimizing routes against delivery windows, vehicle capacity and traffic.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Geocoded stop addresses with delivery windows and access constraints
- Vehicle types, capacities, driver hours rules and depot locations
- History of actual drive times and stop durations from telematics or driver devices
- Order data with weight, volume and special handling needs
Systems to integrate
- Order management or transport management system
- Telematics and driver mobile apps
- Mapping and traffic data provider
- Customer notification tools for delivery windows
- Warehouse or yard system for loading sequence
Complexity: Medium
Routing engines are bought as software or built from solvers. The work for the fleet is clean address and geocode data, accurate service times and vehicle constraints, an order feed from the order or transport management system, and driver devices that return the actual route. Planners also have to accept plans they did not build by hand.
- 1
Measure the baseline on real routes
Record miles, route duration, stops per route, early route terminations and first attempt delivery rate for a few weeks per depot before changing anything. Without it, the gain cannot be separated from seasonal volume.
- 2
Fix the data the optimizer depends on
Clean addresses and geocodes, set realistic service times by stop type, and record vehicle and driver constraints. An optimizer working from wrong service times produces routes that look efficient and run late.
- 3
Run in parallel before cutover
Plan the same days with the tool and with the current method, compare miles and lateness, and let planners explain the differences. Move one depot at a time.
- 4
Put planners on exceptions
Release plans automatically when they pass checks on lateness, hours and capacity, and send only stops the optimizer could not place to a planner. ArrowXL says its system plans 95% of routes overnight, which leaves a small share for manual work.
- 5
Add same day replanning
Once the overnight plan holds, allow reoptimization for traffic and new orders within limits drivers accept, and keep the driver able to override a sequence.
- 6
Close the loop with execution data
Feed actual drive times, stop times and failed deliveries back into the estimates every week, and review routes drivers repeatedly deviate from.
Guardrails
- Driver hours, rest and vehicle load rules are hard constraints, never soft costs the optimizer can trade away
- Drivers can override a sequence for safety or access reasons, and overrides are logged and reviewed
- Delivery windows promised to customers are checked before a plan is released
- Location and performance data of drivers is used for routing and safety, with a documented purpose and retention period
KPIs to instrument
- Miles per route and per stop, against the baseline for the same depot and season
- Deliveries per route and per driver hour
- Early route terminations and routes finishing late
- First attempt delivery success and share of deliveries inside the customer's window
- Share of routes released without manual edits
Human in the loop
Planners own exceptions: stops that do not fit, special handling and new depots. Drivers keep the final say on the road. Operations managers review lateness, mileage and driver feedback each week and approve changes to service time and constraint settings.
Common failure modes
- Wrong service times
- Stops take longer than the model assumes, so routes overrun. Measure service time by stop type from driver data and update it often.
- Plans drivers do not follow
- Drivers who see a sequence as illogical reorder it, and the gain disappears without anyone noticing. Compare planned and driven routes and talk to drivers about patterns.
- Savings credited to the wrong cause
- A mileage drop can come from consolidating depots or lower volume. Compare against a baseline for the same depots and period, and note when manual processes were removed at the same time.
- Constraints left out
- Low bridges, vehicle access limits or customer rules missing from the data produce plans that cannot be driven. Collect constraints from drivers and add them to the data.
What are the risks and rules?
EU AI Act
Depends on design
Planning routes for a fleet is not an Annex III use. It becomes high risk under Annex III point 4(b) when the system allocates tasks to drivers on the basis of individual behaviour or personal traits, or monitors and evaluates their performance and behaviour, for example by assigning harder routes or scoring drivers from telematics and using that for pay or discipline. A system that only optimizes stops, vehicles and loads, with no per driver profiling, is usually minimal risk.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 4(b) covers AI that allocates tasks based on individual behaviour or traits, or monitors and evaluates workers, which is the point to check for driver routing.
- Regulation (EC) No 561/2006 on driving time and rest periods in road transport (European Union, Europe). Sets driving time limits and rest rules for drivers of goods vehicles over 3.5 tonnes in the EU, so most last mile vans fall outside it, but a plan for heavier trucks must respect it as a hard constraint. The EUR-Lex page could not be text verified because it answers automated requests with a bot challenge.
Controls to put in place
- Documented list of constraints the optimizer treats as hard, owned by operations and compliance
- Records of manual overrides by planners and drivers with reasons
- Data protection assessment for vehicle location and driver performance data
- Review of any use of routing or telematics data in driver performance management
Frequently asked questions
- How many miles can AI route optimization remove?
Reported results vary and depend on the starting point. Descartes says ArrowXL by JP Home reduced fleet mileage by approximately 13%, and the company's chief information officer credits better planning and less manual work together. New Castle Building Products says it takes approximately 25,000 miles out of its routes each year, without a baseline. Walmart says its system avoided 30 million unnecessary miles, and a UPS executive said the original ORION deployment cut 8 miles per driver per day. Plan on the low end of these figures for a first depot, and measure your own baseline.
- Is route optimization really AI?
The public evidence is mixed. Walmart describes its route mapping as AI driven, and Amazon says it is investing in AI based routing technology for its delivery drivers, without describing the models or giving a result. Descartes calls its suite AI powered but its ArrowXL release does not say which planning steps use machine learning, and its New Castle Building Products release calls the routing data driven. UPS describes ORION as a technology and an optimization system, and the 2021 Investor Day transcript does not call it AI. Most of the mileage results on this page come from classic optimization software. Ask a vendor which steps use machine learning, such as travel and service time estimates, and what they add.
- Is AI route planning high risk under the EU AI Act?
Usually not. Optimizing stops, vehicles and loads is not an Annex III use. It can become high risk under Annex III point 4(b) if the system allocates tasks to drivers on the basis of individual behaviour or traits, or monitors and evaluates driver performance. Keep routing separate from driver scoring, or assess the combined system as high risk.
How to cite this page
Blits.ai AI Use Case Library, "AI route optimization for delivery fleets and trucks", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/delivery-route-optimization. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published