How we choose between generative AI, machine learning, and optimization to solve the problems that matter most for our customers

Every logistics vendor seems to be announcing an AI feature right now. Some of them will matter. Many will make for a good demo and change nothing about how a truck gets loaded, charged or how goods are delivered.

At Einride, we try to avoid the trap of picking a technology first and then hunting for a problem to apply it to. We work the other way around: find the problem with the biggest impact for the customer, understand it properly, and only then decide which technology fits. Sometimes that is generative AI (GenAI), but often it isn’t, and the best solutions usually combine several technologies.

What makes a problem worth solving

Before we pick any technology, we ask four questions:

• Does it move cost, emissions or vehicle utilization in a meaningful way?

• Does it come up often enough that small improvements add up?

• Do we have the data, or can we get it?

• Can we measure that what we implement has the intended impact?

Getting more out of every vehicle

Vehicle utilization is one of the highest-impact challenges in freight, directly affecting both profitability and emissions. In Europe, one in five truck kilometers was driven empty in 2024, while in the U.S., empty miles rose from 16.3% to 16.7% in 2025¹. For electric fleets, effective planning is even more crucial: a depleted battery is far more disruptive than needing a quick diesel refill. Moreover, because electric trucks currently carry higher upfront costs due to developing supply chains, maximizing their operational utilization is essential to ensure commercial viability.

Planning a fleet well is not one problem but three connected ones, and each is best solved by a different technology.

1. Defining the goal: GenAI. 

A planner knows what they know (fleet size, driver availability, contracted loads) and what they want: “run these loads, cut emissions by this much, stay above 99% on-time.” Explaining that in plain language is what GenAI is good at. It turns the planner’s description into precise parameters for the planning engine, so the planner doesn’t have to configure anything manually. The human still decides what a good plan looks like.

2. Predicting what will happen: machine learning. 

A plan only works if the vehicle can actually complete it. How much battery a route uses depends on traffic, weather and what’s on the truck. A cold winter day in Chicago with a heavy load of beer drains the battery far faster than the same route in summer with a load of lightbulbs. Planners used to estimate this by gut feeling. Einride uses machine learning, models that have been trained on seven years of historical operational data to predict energy consumption for each specific vehicle on every route. This removes the guesswork; a planner is able to predict at over 90% accuracy exactly what that vehicle will be able to run in a day. This level of precision planning enables them to add that extra order to the daily shift to maximize revenue, or cut time from the charging time to make a tight delivery, without running out of battery.

3. Making the best decision: optimization. 

Finally, the plan has to be built. Optimization algorithms take the planner’s goal (for example, minimizing total cost or emissions), the fixed constraints (driver hours, vehicle range, delivery windows, charger availability) and the predictions (expected demand, travel times, charging times) as input to find the best plan. Because Einride’s optimization runs on accurate vehicle-specific energy predictions, it can adapt on the fly when disruptions occur, such as unexpected traffic delays or an out-of-service charger, rerouting or swapping trucks so deliveries stay on schedule without wasting capacity.

Why not use GenAI for everything?

It’s tempting to hand the whole problem to a large language model (LLM), but that would produce a worse plan at a higher cost. LLMs are not built to guarantee a plan that respects every constraint, they can give different answers to the same question, and running them at scale is expensive. Machine learning predicts better, and optimization decides better. Each technology has a job it does best, and the results improve when they work together.

None of this is an argument against AI. It’s an argument for fit.

What this means for you as a transport buyer

When a vendor tells you about a new AI feature, a few questions will tell you whether it adds value:

1. Which specific problem was this technology chosen to solve, and why is that problem worth solving?

2. How is the result measured, in cost reductions, emissions savings, on-time delivery score?

3. What would the result be without it?

We start with those questions ourselves, and we’ve been building and benchmarking our planning technology on real customer operations since 2018 with more than 2 million days of transport data. The result is freight capacity planned to maximize every vehicle, with fewer empty kilometers and lower emissions.

To see what this looks like for your network, get in touch at sales@einride.tech.