Route optimization leverages AI: But not how you think

Carriers and fleets are increasingly turning to AI-powered tools for route optimization. According to Trimble's 2026 Transportation Pulse Report, 31% of carriers and LSPs use AI for driver scheduling and route planning. In addition, the top AI capabilities they prioritize include real-time rerouting (54%) and predictive route and load planning (53%). When asked which logistics activities are most suitable for AI agents, 40% say route and fuel optimization.

Neil Fernandes, Founder of EnrouteAI, built his company to leverage AI for route optimization, but he says it is not used the way you might think.

Meeting Delivery Expectations

“In the mid-2010s, Amazon and Walmart reset delivery time expectations for everyone,” Fernandes recalls. “Delivery times went from weeks to days, and in some cases to hours. What struck me was that even Fortune 500 companies running state-of-the-art routing software struggled, because the routes those systems produced didn't reflect reality: real traffic, real parking, how long a stop actually takes.”

A route that looks perfect on a screen falls apart on the street, and most carriers never close that gap, Fernandes explains. First, the map address is not necessarily where the delivery happens. Google Maps may take a driver to the front of the building, but the loading dock is in back, and in a dense city during rush hour, going around the block can take 15 to 20 minutes.

The second mistake is planning with averages, says Fernandes. “A city delivery, where the driver hunts for parking and then hunts for the right entrance, can take 20 minutes. The same delivery in the suburbs takes five. Yet most routing software assumes a single fixed service time at every stop. Traffic gets the same treatment: one flat multiplier applied to the whole route, as if rush hour and mid-afternoon were the same road. Multiply those errors across a hundred stops a day, and your plan is fiction by 10 am.”

The third mistake is being reactive about disruptions that are actually predictable, Fernandes advises. “A ball game schedule is published months in advance. A presidential visit is announced. Bad weather is forecast. All of these will wreck a route plan, yet most carriers treat them as surprises and scramble on the day. The ability to plan is what separates the best operators from the rest.”

All three problems have the same root cause: carriers don't capture what their own drivers learn. An experienced driver knows the best place to park, which entrance to take, and where the package needs to be dropped off.

The Fingerprint of the City

“AI helps routing, but not in the way most people assume,” Fernandes explains. “You have to separate two things: machine learning, which has been quietly powering logistics for a decade, and large language models (LLMs) like ChatGPT or Claude, which is what most people mean by AI today.”

The real work happens in machine learning, upstream of optimization, he continues. “Routing lives or dies on how well your plan matches reality, and ML is excellent at learning reality from messy operational data: how traffic actually behaves on this street at 8 am versus 2 pm; where drivers actually park at this address; how long this customer's dock really takes. That learned layer is what I call the fingerprint of the city. Feed it back into the planner, and the system improves with every delivery. That's the part people miss. A static route is obsolete the day you plan it. A learner gets better every single day, and the ETAs get better with it.”

Fernandes says LLMs are a different story. Language models are bad at math, sometimes even basic arithmetic – and route optimization is a mathematics problem, best left to specialized algorithms that are more accurate, faster, and can explain their choices. “Routing analysis is usually done best by traditional, non-AI algorithms, and it's where most of the differentiation in routing actually lives. So, the division of labor is simple: the AI learns the world, and the algorithm solves the puzzle. The companies achieving real results treat AI as the conductor and the specialized algorithms as the orchestra, rather than asking a chatbot to plan routes.”

The challenge is separating real capability from AI-washing, Fernandes warns. “The AI label is on everything now, and it's hard for a buyer to distinguish a learning system from a rebranded rules engine. My advice is simple: judge vendors on live results with your own data, not on demos.”

Enroute Final Mile is the company's route optimization and dispatch platform for delivery fleets. Final Mile plans each day's routes, learns how deliveries actually happen at every address, and keeps ETAs accurate for dispatchers, drivers, and end customers. “Measured across our customers' operations, routes can be made 15% to 30% more efficient compared to manual planning,” Fernandes reports. “Traditional optimization alone gets you the first 15%. The rest comes from the self-improving layer, the fingerprint of the city feeding real-world learning back into the plan. And efficiency is only part of the payoff. The same learning loop is what makes ETAs accurate enough to share with customers, which is often worth as much as the saved miles.”

The Next Frontier

Fernandes believes the next frontier in route optimization is customer-facing. “The math of planning routes is largely solved; the next wins come from what surrounds the route. An AI agent that calls the customer when their couch delivery is running two hours late. Software that reads 'leave my package at the neighboring business' and actually translates that instruction into the driver's workflow. Language is where modern AI genuinely shines, and logistics is full of language problems we've been treating as afterthoughts.”