AI in Logistics: Routing, Load Planning and ETA Prediction

By Tara Minh · Updated Oct 11, 2026

AI in logistics is the use of optimization engines, machine learning and language models to decide how goods move: which route a vehicle takes, how freight is packed and consolidated, when a shipment will arrive and which carrier should haul it. People still own the promises made to customers and the trade-offs between cost, speed and risk. It isn't one product. It's a set of techniques pointed at the data transport already produces: orders, locations, vehicle capacity, driver hours, GPS pings, scan events and carrier performance.

This is the hub of the logistics, fulfillment and field operations group in the AI for Operations collection. It covers the whole movement of goods at a manager's level, then points to deeper pieces on AI route optimization, AI in warehouse operations, AI field service scheduling, AI in reverse logistics and AI fleet management. For the wider chain around transport, start with AI in supply chain management.

Key Facts

• UPS says its ORION routing system is expected to cut 100 million miles driven and 10 million gallons of fuel a year, and to save $300 million to $400 million annually in the United States (Transport Topics, August 2016). Those are company expectations, not audited results.
• A World Bank working paper that reviewed 40 years of research across 27 countries found an average empty-mile share of 29 percent in freight trucking, with wide variation between settings (World Bank Policy Research Working Paper 10775, May 2024).
• Google and DeepMind reported that a graph neural network improved the accuracy of real-time ETAs by up to 50 percent in cities where it was tested, with gains ranging from 16 percent in London and Copenhagen to 51 percent in Taichung (DeepMind, September 2020).
• UPS's Network Planning Tools, which combine machine learning with analytics to route volume across facilities, were expected to save $100 million to $200 million a year, and engineers still make the final routing calls (MIT Technology Review, November 2018).
• The pattern across all of it: AI pays off where the problem is a well-defined puzzle with clean data, and it disappoints where the real constraints live in people's heads.

What AI in Logistics Actually Does

Four jobs cover most of what you'll be sold.

Routing. Decide the order of stops and the path between them for each vehicle, given delivery windows, capacity, driver hours and traffic. This is an optimization problem, and machine learning mostly helps by feeding it better inputs such as realistic drive times and service times per stop.

Load planning. Decide what goes on which vehicle and how it's packed, and combine partial loads into fuller ones. The goal is fewer trips, fewer empty miles and less damaged freight.

ETA prediction. Estimate when a shipment will arrive and update the estimate as conditions change. This is a prediction problem, and the output is only useful if it's honest about uncertainty.

Carrier selection. Choose who should move each load, based on price, service history, lane coverage and capacity. Language models add a fourth skill here: reading rate confirmations, bills of lading and emails, and drafting exception messages.

In ACE Framework terms, those map to Ingest (documents and tracking events), Analyze, Predict (ETAs and delay risk), Generate (customer updates) and Execute (dispatching a load or tendering it to a carrier). The first four produce information a planner can check. Execute commits money and trucks, so that's where you set tight limits.

One useful distinction: a route plan is a decision, an ETA is a prediction. Teams that blur the two end up promising customers the optimistic number from the planning model.

Where It Shows Up, Function by Function

Routing: the best-documented example

UPS is the example most managers have heard of, and it's useful because the company has been unusually open about it. ORION (On-Road Integrated Optimization and Navigation) recalculates delivery routes for drivers. According to UPS, as reported in Transport Topics, it is expected to cut 100 million miles and 10 million gallons of fuel a year, with annual savings of $300 million to $400 million in the United States (Transport Topics). The same article reports that eliminating a single mile per driver per day would be worth $50 million a year to the company.

Two lessons carry over to a business ten times smaller. First, the savings come from small per-vehicle gains multiplied across a very large fleet, so your math depends on your volume. Second, the figures are the company's own, so don't assume a smaller carrier will see the same percentage. The deeper mechanics are in AI route optimization.

Load planning: attacking empty and half-empty miles

A truck that runs empty still costs money. A World Bank working paper reviewed 40 years of estimates across 27 countries and found an average empty-mile share of 29 percent, with significant variation across settings (World Bank). Treat that as a reason to look at your own number rather than a benchmark to hit, because private fleets, for-hire carriers and different freight types differ a lot.

AI helps in two ways. It consolidates shipments so fewer, fuller vehicles go out, and it schedules trips so vehicles come back with something on board. UPS describes the same idea in its Network Planning Tools, which group outbound packages into the fewest trailers or planes and schedule trips so vehicles don't return empty (MIT Technology Review). The article also notes the tool is built so engineers, not the software, make the final routing decisions. Inside the building, the same logic applies to staging and loading, covered in AI in warehouse operations.

ETA prediction: better than a distance divided by speed

A naive ETA divides remaining distance by average speed. A learned ETA uses historical trips on the same road at the same hour, current traffic, weather, and dwell time at stops. Google and DeepMind's published work on Google Maps is a clean public example: a graph neural network model improved real-time ETA accuracy by up to 50 percent in several cities, with results ranging from 16 percent in London and Copenhagen to 51 percent in Taichung (DeepMind). That's road travel time for consumer navigation, not a promise about your freight, so it shows what the technique can do and not what you'll get on a dock-to-dock shipment.

For freight, the hard part is everything that isn't driving: waiting to unload, hours-of-service breaks, customs holds, a trailer that wasn't ready. Models that include those events do better than ones that only watch GPS. The practical rule is to publish a window, not a single minute, and to widen it when the lane has little history.

Carrier selection: scoring, not autopilot

Carrier selection tools score each option on price, on-time performance, claims, lane coverage and current capacity, then suggest who should get the load. They're best at the routine majority of loads, leaving buyers time for the awkward ones. The risk is a score that rewards carriers with long histories and quietly shuts out newer ones, so keep a rule for trialing new carriers on low-risk lanes. For the sourcing side of carrier relationships, see AI in supply chain management.

Network planning: moving volume, not just trucks

The UPS Network Planning Tools story shows the system view. The platform forecasts package volume and weight from history, shows engineers each facility's activity, and can reroute packages around a trouble spot such as a snowstorm, notifying affected facilities. UPS expected it to save $100 million to $200 million a year, and one executive said clearing a backlog at an Illinois facility took minutes where it would previously have taken at least a week (MIT Technology Review). A control view that pulls these signals into one place is the natural next step.

Illustrative Example: A Regional Food Distributor

This scenario is hypothetical, to show how the pieces fit together.

A distributor runs 30 trucks delivering to 400 restaurants and stores. Dispatchers build routes each evening from a spreadsheet and their memory of which customers have tight docks. Complaints cluster around late arrivals and a few trucks that finish at noon while others run until nine.

The operations manager picks one question, "can we cut overtime on the five worst routes?", and one function, routing. The team records the actual service time at each stop for four weeks, because planned stop times are the weakest number in the system. They load those, plus delivery windows and truck capacities, into a routing tool and compare its plans with the dispatchers' for ten days without changing any dispatch. Each morning, dispatchers mark where the tool's plan was impossible and why: a loading dock closed at 10, a customer who needs the same driver.

Those notes become constraints in the tool. By the end, the tool matches or beats the manual plan on most days and loses on days with unusual customer rules, so dispatchers keep final approval. The team then adds a simple arrival window for customers, widened for stops with little history. Nothing here needed a research model. It needed measured stop times, written-down constraints and a person with authority to say no.

How to Implement AI in Logistics, Step by Step

1. Pick one number you already track. Overtime hours, miles per stop, on-time rate, empty-mile share, cost per shipment, or late-delivery complaints. No baseline, no proof.
2. Choose one function and one lane or region. Routing in a single depot, or ETA on a single lane. A contained pilot gives a clean read.
3. Fix the data before the model. You need clean addresses, real service times, accurate vehicle capacities and timestamped scan events. Our piece on operational data for AI covers what to repair first.
4. Write down the constraints that live in people's heads. Dock hours, driver preferences, customer rules, equipment limits. A tool can't respect what it was never told.
5. Start with recommendations. Let the system propose routes, loads or carriers while a dispatcher approves. Move to automatic only for low-risk, repetitive decisions.
6. Run it in parallel with the current method. Compare both on the same metric for several weeks. Log where the tool lost, not only where it won.
7. Define the fallback. Decide what happens when the data feed drops or a driver calls with a road closure. Freight still has to move.
8. Review weekly, then widen. Add the next depot, lane or function only when the last one's numbers hold. Give any software that acts across the chain authority only after its recommendations have earned trust.

Pitfalls to Avoid

Optimizing the vehicle and ignoring the dock. A perfect route is useless if the loading bay can't stage freight in time. Check the whole path, from pick to proof of delivery.

Using planned service times. Models fed with the time a stop is supposed to take produce plans that fall apart by mid-morning. Measure actuals.

Showing one ETA minute. A single timestamp hides uncertainty. Give a window, and say so plainly when the lane is new.

Letting a score decide alone. Carrier and driver scores built on thin history can penalize the wrong people. Keep a human review path.

Ignoring drivers. The people executing the route know about the low bridge and the customer's back entrance. If they feel overruled by a black box, they'll work around it. Collect their overrides as data.

Forgetting upkeep. Lanes, customers and traffic patterns change. Someone has to own retraining, monitoring and switching a model off.

The Limits of AI in Logistics

Disruptions it hasn't seen. A port closure, a new tariff or a bridge collapse isn't in the history. Keep a human scenario process for the unprecedented.

Thin data. New lanes, new customers and one-off deliveries are where models are weakest. A rule or an experienced dispatcher may do better there.

Hidden constraints. Informal agreements, driver relationships and site access quirks often live outside any system.

Trade-offs belong to managers. Cost, speed, emissions and service risk pull in different directions. A model can show the options. It can't be accountable for choosing among them.

Governance. Decide who owns each model, what it may do alone and how to switch it off. The US National Institute of Standards and Technology describes its AI Risk Management Framework as voluntary guidance, a reasonable checklist for ownership and monitoring.

What to Measure

Track service, cost and control together. Fewer miles that cause more late deliveries isn't progress.

• Service: on-time in-full rate, ETA accuracy (how often arrival falls inside the promised window), late-delivery complaints.
• Cost: cost per stop or per shipment, overtime hours, fuel per mile, premium freight spend.
• Utilization: empty-mile share, load fill rate, stops per route hour.
• Planner trust: share of AI plans overridden and the reasons.
• Control: incidents traced to an AI output and time to detect a disrupted shipment.

For the wider set of operating metrics, see AI in operations management.

Where to Go Next

If you do one thing this quarter, measure the real service time at every stop for one depot for a month, then test a routing or ETA tool against your current plan on that data. Then read the spoke that matches your pain: route optimization for overtime and miles, warehouse operations for dock and picking delays, field service scheduling for technician visits, reverse logistics for returns, and fleet management for vehicle uptime and cost.