If you listen to the marketing departments of EV charging networks, you’d think their chargers were sentient. Every charging station is now branded as intelligent, and every grid management tool is supposedly AI-powered. However, for engineers, fleet operators, and infrastructure planners, the gap between what is sold in glossy brochures and what is shipping in production environments is wider than ever.
To build a reliable green future, we must move past the buzzwords. We need to distinguish between speculative sells, practical ships, and the significant technical challenges that arise when we treat a statistical engine as an infallible oracle.
The most prominent product being sold in the EV space today is utilization forecasting. The pitch is seductive: an AI that can predict with near-perfect accuracy when a charger will be free, how much load the grid will take on a rainy Tuesday, and where a company should break ground on their next hub to maximize ROI.
However, the idea of perfect foresight remains a myth. In the real world, these models often collapse because they are built on a statistical mirage. In reality (so far), an AI cannot account for a local road closure, a broken connector, or a sudden change in electricity subsidies that shifts consumer habits overnight. Investing in a charging hub is a decade-plus commitment; usually a 10–15 year horizon, and no level of AI can predict a future shaped by new housing developments, shifts in local traffic patterns, nearby competitors entering the same zone, or evolving electricity tariffs. Treating such probabilistic estimates as absolute certainties is a recipe for operational disappointment.
The primary hurdle for these forecasting models is the early enthusiast problem. Machine learning requires representative historical data. Currently, the bulk of our historical charging data comes from early EV adopters, a niche group of affluent early enthusiasts who often have private driveways, high-end home chargers, and very specific driving patterns.
For example urban hubs like Amsterdam and Rotterdam achieved high occupancy rates in recent years because they served a niche group of wealthy, tech-oriented enthusiasts with private driveways and high-end chargers. Machine learning models trained on this data systematically over-predict utilization by 40–62% when applied to secondary markets like Warsaw or Porto, where actual rates average 25–35%.
As we pivot to the early adopters, then early majority, and so on, the average driver may live in a high-rise apartment without home charging and have vastly different economic motivations; this old data does not represent this. When you train an AI on small data from a niche group, the resulting forecasts are often just digital echoes of a past that won't look like the future. We are currently selling a crystal ball, while the data used to polish it is still profoundly skewed and limited in scale.
While the industry sells the vision of an all-knowing oracle, what is shipping in real-world production environments is far more pragmatic. The most successful applications of AI emphasize integration, traceability, and human collaboration.
Generative AI delivers massive value in specific back-office workflows when properly constrained. Converting GIS satellite imagery into permit-ready architectural renders can accelerate site approvals by 40%.
However, we must respect the Hallucination Challenge. In 2025, domain-specific evaluations showed that large language models can have a hallucination rate of up to 27% when referencing complex energy directives, frequently fabricating non-existent regulations. In an infrastructure context involving multi-year commitments, the safe pattern is to use AI for drafting and ideation while keeping final technical decisions firmly tied to human verification.
The real power of shipping AI lies in data integration. A modern charging hub sits at the nexus of weather patterns, fluctuating energy spot prices, and vehicle telemetry. AI excels at identifying correlations that a human analyst might miss, such as how a sudden drop in temperature, combined with a local festival, drives a specific surge in DC fast-charging demand.
The importance here is staying in control; the AI provides the correlation, but humans use it to make better-informed tactical decisions about energy procurement.
The most sophisticated AI applications shipping today are not standalone super brains, but parts of integrated, collaborative ecosystems. A useful metaphor, popularized by AI pioneer Patrick Winston, is the raisin bread model. In this analogy, the raisins represent the intelligent AI modules, while the bread is the robust operational system that holds everything together.
The raisins may be small, but they define the flavor, and they make the loaf worth baking. Likewise, AI adds intelligence and adaptability, yet it depends on the much larger foundation beneath it:
If the bread is missing or poorly mixed, no number of raisins will make it rise. In EV infrastructure, success comes from integrating these ingredients so the system behaves as one coherent loaf. We are now seeing this principle reflected in Agentic AI; networks where vehicles, chargers, and local transformers communicate to dynamically balance load. When the grid is stressed (the room is too cold), these agents can collectively slow charging rates to preserve stability. It’s not about one all-knowing intelligence, but about many small, smart parts working harmoniously within a strong system.
As we integrate these tools, we must address the pros and cons with a healthy dose of skepticism regarding the current AI narrative.
A recurring challenge is the tendency to use an AI hammer to crack a nut. This refers to the over-engineering of simple problems. In many instances, a basic if/then script or a standard linear regression can manage load-balancing more efficiently than a massive, power-hungry neural network.
If we use a high-compute AI that consumes significant energy just to figure out how to save 1 kWh on the grid, we are defeating the purpose of sustainable infrastructure. We must ask: Is the complexity of this AI justified by the output, or are we just adding layers of intelligence because it sounds better to investors?
Perhaps the most dangerous challenge is the cultural assumption that AI is all-knowing and always correct. In reality, AI is a statistical engine that generates the most probable answer, not necessarily the true one.
The hallucination challenge: In a technical permit or a safety manual, a hallucination, where the AI confidently asserts a non-existent regulation, can lead to catastrophic failures or legal liability.
Fact-checking with sources: We cannot allow the black box of AI to dictate infrastructure policy without a transparent trail. Every AI-generated output on grid safety or technical specs must be treated as a draft and verified in a human-in-the-loop against primary sources.
AI in the EV sector is not a magic wand but a sophisticated set of tools that require careful calibration. As we scale toward the 2030 targets, the winners will be those who use AI to solve specific, high-value problems while maintaining rigorous human oversight.
AI is the raisin, not the bread. It’s the small but defining ingredient that gives the system its distinct edge, yet it depends on the broader network of hardware, energy, and data to rise together. We must decide what kind of loaf we’re baking, ensuring that intelligence enhances, rather than replaces, human design and vision.
At Dodona, we help build the bread that today’s AI raisins need to matter: a decision platform that combines 50+ market, grid, and mobility data sources with configurable models to answer three practical questions for CPOs and investors:
Could we deploy?
Should we invest?
How do we deliver at scale?
If you want your next wave of charging sites to be evidence-based rather than guesswork, explore how Dodona can support your network planning and investment decisions at thedodona.com.
Stefan Furlan, CEO, Dodona
(Book a meeting with me here)