PROOF POINT · EV CHARGING
Fifty candidate sites. Fifteen worth a visit.
Choosing EV charging sites is expensive guesswork. A site-selection model built with AI reasoning applied offline to research and public data was validated against live occupancy from France’s QualiCharge open-data API.
The problem
Operators of EV charging networks shortlist dozens of candidate locations and then evaluate them on the ground. Each site visit costs time and money, and the shortlist itself is built on judgement and rules of thumb. The result is expensive guesswork at scale.
The operator wanted a better way to decide which sites deserved a visit—one that could be trusted, and checked, before capital was committed.
The intervention
We built a site-selection model in which AI reasoning was applied offline to research and public data, producing a rating for every candidate location. Then we did the part that most AI analyses skip: we checked the ratings against reality.
Location characteristics, surroundings, traffic and demand signals assembled from research and open sources for each candidate site.
The model reasons over the evidence for each site and produces a rating—offline, reproducibly, and without the AI sitting in any live decision path.
Ratings were compared with live occupancy from France’s QualiCharge open-data API at existing chargers. Top-quartile-rated locations were measurably busier than bottom-quartile ones.
The 50-site search narrowed to the 15 locations worth a site visit—so effort and capital go where the evidence points.
The outcome
- Top-quartile-rated locations are measurably busier than bottom-quartile ones
- A 50-site search narrows to the 15 worth a site visit
- The model was validated against live, independent data, not against its own assumptions
- A repeatable process for the next search, not a one-off study
A model that cannot be checked is an opinion. This one was checked against what actually happens at real chargers.
The full validation is published as a report you can read for yourself: Saint-Denis validation report
Which decision is still made by shortlist and site visit?
If the data exists to rank the options and the outcome can be checked, it is a candidate for the same approach. Tell us about it.