Catching Fuel Fraud With ML: How Centrix Identifies the 4–7% of Transactions That Don't Add Up

The cost line you can't afford to take on faith

Fuel runs $42,000–$48,000 per truck per year on a long-haul dry-van fleet at current diesel prices. Across 100 trucks, that's $4.2M–$4.8M of annual spend. A 4% fraud rate — which is the median we measure on new managed fleets — is $168K–$192K of cash walking out the door, and almost none of it lands in a fleet manager's morning report because the fuel-card statement looks normal at first glance.

Fraud isn't usually a single dramatic event. It's a steady drip of small things:

• A driver fills 105 gallons into a 100-gallon tank
• A fuel card swiped at a station 60 miles off the route
• A 60-gallon transaction at 2:14 AM in a state the truck wasn't in
• A driver with a 6.2 mpg pattern who suddenly buys fuel like he's running 4.8 mpg
• Two cards fueled within 2 minutes at the same pump

Each of these has a plausible explanation. The pattern across thousands of transactions is what tells you fraud is happening. That's what the model catches.

Why Relay Fuel, not Alvys, is the source of truth

Centrix's fuel data comes from Relay Fuel, not from Alvys. Alvys has a fuel sync, but it's redundant — Relay is the system of record for fleet fuel transactions on most managed carriers, and the data is more complete and more real-time. The fleet rule (documented in project memory): fuel transactions come from Relay Fuel, not Alvys; Alvys fuel sync is redundant.

If your fleet uses a different fuel-card provider (Comdata, EFS, WEX, RTS, Pilot Direct), Centrix has connectors at varying stages — bring it up in a demo.

The fraud signals the model watches

The fraud model is a stacked classifier (gradient boosting + a feature-derived rule layer) trained on labeled fraud examples from prior managed fleets. The top signals by feature importance:

Tank-cap violations

The truck's tank capacity is on file. A transaction that exceeds the tank cap by more than the safety margin is either a misread, a card swipe at a different truck, or a driver pumping into a personal vehicle. Tank-cap violations alone catch ~30% of all fraud cases in the first month.

Location vs route mismatch

The truck's GPS position from Samsara at the time of the fuel transaction should be at the fueling station — within 0.5 miles. A transaction that fires when the truck is 30 miles away is a card-sharing event. A transaction that fires in a state the truck never entered is a card-skimming event.

MPG outliers

Each truck has a baseline MPG, derived from prior 60 days of fuel and miles. A driver whose MPG suddenly drops 18% — without a fault code, without a route change, without a season change — is either pumping less fuel than logged (skim), pumping fuel into a non-truck vehicle, or pumping in someone else's truck.

Time-of-day anomalies

Truck-stop fuel transactions cluster around shift boundaries. Off-cycle transactions (3 AM at a station the truck has never used, on a day the truck is off-duty) are heavily weighted.

Velocity pairs

Two transactions on the same card within 2 minutes at the same pump = double fueling, common with kickback schemes.

Station optimization deviation

The fuel optimizer recommends the cheapest in-route station within the truck's range. Drivers consistently fueling at $0.30+ above the recommended station — when the recommended one is on the route — is a kickback signal.

The optimizer side: $/mile improvement

Fraud is one half of the fuel story. The other half is per-mile cost optimization — making sure the gallons you do buy are bought at the cheapest in-route price.

Centrix maintains a continuous national fuel price feed and routes it through the planned trip. For every truck on an active load, the model surfaces:

• The cheapest station within 1 mile of the route, by gallon
• The cheapest station within 5 miles, weighted by detour cost
• The truck's required gallons to reach the next mandatory stop
• Any driver-preference constraints (truck-stop chain, shower availability, etc.)

The optimizer is wired into the dispatcher view and the driver Telegram bot. The driver sees the recommendation in their language. The dispatcher sees the adherence rate in the morning brief.

Median savings on per-gallon cost: $0.12 to $0.18 per gallon. On 7,500 gallons per truck per year, that's $900–$1,350 per truck per year × 100 trucks = $90K–$135K per year, on top of the fraud savings.

The fuel_optimization role

Centrix has a dedicated `fuel_optimization` role in its RBAC system. The role sees fraud alerts, station optimization adherence, per-truck fuel cost trends, and the IFTA tax filing flow. It's typically held by an analyst or operations person, not the fleet manager — fuel optimization is a full-time discipline at fleet scale, and the role lets the work happen without granting access to unrelated systems.

What the workflow looks like

The fraud model runs nightly. Alerts route into the fuel_optimization role's queue, sorted by confidence and dollar magnitude. The analyst reviews each alert:

• Confirmed fraud — card flagged, driver flagged, recovery process

initiated, training (or termination) workflow started

• False positive — flagged in the model so similar patterns score lower

next time (continuous learning)

• Pending — needs more data; sits in the queue until the next sync cycle

Recovery on confirmed fraud averages 60–80% of the underlying loss when caught within 30 days, dropping to 20–30% after 90 days. Speed matters — which is why the model runs nightly, not weekly.

ROI on a 100-truck fleet

• Fraud catch: 4% × $4.5M = $180K. Median 70% prevention/recovery rate =

$126K/yr saved.

• Per-gallon optimization: $0.15/gal × 750K gallons = $112K/yr saved.
• Total fuel program impact: ~$238K/yr on a 100-truck fleet.

For comparison, the fully loaded cost of a Centrix fuel optimization role (software + analyst time) lands around $40K–$60K — so the program runs 4–6x ROI on the fuel layer alone, before you count any other module.

See the fuel program — bring 90 days of Relay Fuel transactions and Samsara position data and we'll score your fleet's fraud exposure same-day.