Automatic detection of invalid fuelings is one of the features with the greatest impact on reducing a fleet’s fuel costs, but also one of the most poorly explained. Vendors tend to talk about “smart algorithms” or “AI fraud detection” without saying what the system actually does. This guide changes that: it explains in plain language how the VEC Fleet algorithm works, what it evaluates in each fueling, how it decides between NORMAL and REVIEW, and why that logic is defensible before drivers, internal audit and senior management.
The logic is neither magic nor a black box. It is a set of explicit rules based on observable data. Understanding them gives you control over the system and confidence in its results.
When a fueling enters the system, whether through the card integration or manual entry, it goes through four validations:
1. Tank capacity.
Compares the liters fueled against the tank capacity of the vehicle’s model. If the liters exceed capacity beyond a tolerance margin, the fueling is flagged. The tolerance exists to absorb the differences between rated and actual capacity.
2. Expected fuel efficiency.
Calculates the fueling’s actual fuel efficiency —kilometers traveled since the previous fueling divided by liters consumed— and compares it against the model’s expected fuel efficiency. If the implied fuel efficiency falls well below, something doesn’t add up: either not all of the fuel went into that tank, or the vehicle has a mechanical problem.
3. Fuel card.
Verifies that the card used to pay for the fueling is the one registered on the platform for that vehicle. If they don’t match, the fueling is flagged.
This check exists for a specific reason: drivers swapping cards. A card assigned to one unit that shows up paying for another unit’s fueling is the clearest sign that something is being diverted, and it is one of the few checks that does not depend on estimates or tolerances.
4. Geolocation.
Cross-references the location of the station where the fueling took place with the vehicle’s position at that moment. If the vehicle was far from the station, the fueling is suspicious. The tolerance absorbs GPS inaccuracy and a vehicle parked around the corner.
Every fueling goes through all four. If it passes them all, it is marked NORMAL. If it fails at least one, it is marked REVIEW, with a clear indication of which check was triggered.
And something that is not a check: price per liter. The system shows the price of each fueling compared with the previous fueling, with an up or down arrow. It is a visual indicator for the manager to look at, not a rule that triggers review status. It is worth being clear about this: a high price does not flag the fueling, but it does help spot at a glance a station that has become expensive or an area where it makes sense to renegotiate.
The decision to use explicit rules instead of an opaque model is deliberate, and it rests on four reasons.
Defensibility with drivers. When a fueling is flagged and needs to be discussed, it has to be clear why. “The algorithm said so” is not a productive conversation. “The liters fueled exceed this model’s tank capacity by eight” is.
Internal audit. The audit team needs to understand how each decision was reached. Explicit rules are auditable; opaque models are not.
No overfitting to one context. Explicit rules work the same in Argentina or Mexico, in a small fleet or a large one. A model trained on data from one context can fail badly in another.
Direct adjustment by the customer. Each company can adjust the thresholds to its own reality: capacity tolerance, fuel efficiency threshold and distance tolerance.
There are other places in the system where AI does make sense. For fueling validation, explicit rules are still the best tool.
A fueling can fail one, two, three or all four checks. The combination matters:
Only one fails. Marked REVIEW with the rule identified. A one-off investigation by the person in charge.
Two related checks fail. For example, capacity exceeded plus low fuel efficiency, or a card mismatch plus geolocation out of place. A stronger signal, higher priority. That second pair in particular is the classic pattern of a fueling made into another unit.
Three or four fail. Very likely fraud or a gross error. It doesn’t just stay in the responsible person’s inbox: it escalates.
No system is perfect. Typical false positives and how to handle them:
A vehicle with a real mechanical problem that lowers fuel efficiency. The fuel efficiency check is triggered even though there is no fraud. The solution: cross-reference flagged fuelings with the vehicle’s open Corrective tickets. If there is a documented mechanical problem, the flag is expected and is closed as such.
A vehicle fueled with another unit’s card for a legitimate operational reason. This happens when a unit goes out on the road with the wrong card, or when a temporary replacement uses the card of the vehicle it is replacing. The solution is not to loosen the check but to correct the assignment on the platform, which is precisely what the check forces you to do.
A fueling for a vehicle parked around the corner from the station. Geolocation can be triggered if the vehicle was away from the exact pump location. Adjustment: widen the tolerance in dense urban areas, where parking in front of the pump is not always possible, and keep it tighter in open areas.
A vehicle with an auxiliary tank or jerrycan fills. The capacity check is triggered when the driver fills a jerrycan in addition to the main tank, which is common in remote operations. The solution is to reflect that extended capacity in the vehicle’s record.
False positives are reduced through iterative calibration during the first 60 to 90 days. After that they stabilize at a low level.
Urban operation with employed drivers.
Standard thresholds. Operational discipline and the usual controls are enough.
Mixed operation with contracted drivers.
Tighten the geolocation tolerance, because the risk of card misuse is higher. And pay special attention to the card check: it is the one that detects swapping between units.
Long-haul operation with planned stops along the route.
Expected fuel efficiency needs to be calibrated against actual highway operation, which is very different from urban operation. A threshold designed for delivery generates noise in long-haul.
Mining or remote operation.
Widen the geolocation tolerance, because GPS can fail in areas without coverage. It is best to define the adjustment per base rather than globally.
The workflow after detection is as important as the detection itself:
Those classifications are what later make it possible to see whether a threshold is set wrong: when the same pattern keeps recurring and always ends up classified as legitimate, the threshold needs adjusting.
