AI system for reducing fuel costs: 8–15% savings

AI fuel consumption optimization: where accounting breaks down

Overconsumption of fuel in a fleet almost never looks like one large leak. It is the sum of small episodes: drains and shortfills at fueling and parking, trips at excessive RPM and with sharp accelerations, long idling during loading with the engine running.

Each episode separately looks harmless in a report, but together they produce that very share that is then sought for months.

The second, less obvious loss is fuel consumption accounting itself. The tank level sensor lies on grades, on uneven profiles and in frost; measurement error accumulates, and fueling documents and tracks arrive at different intervals.

When the discrepancy between standard write-off and actual level becomes regular, it is closed by manual adjustment in the log — and the data finally stops reconciling.

After that there is nothing to argue with: not with the driver, not with the mechanic, not with the fuel supplier.

It is telling that no single source by itself provides an answer. Tank level does not distinguish a drain from honest fueling, a track does not see a shortfill, and a waybill does not know about idling during loading.

The solution appears when several independent channels are consolidated into one model: AI system for reducing fuel costs matches them by time and place, suppresses measurement error, and separates a real drain from a false alarm on a slope.

Next we break down how this is closed in practice: where to place control points, which slices to consider normal, and how to turn this into sustained 8–15% cost reduction without manual investigations.

Measurable effect: consumption, downtime, equipment life

Effect is measured not by feelings, but by reporting that is already maintained in the fleet: liters per kilometer and per engine hour, operating hours, number of repair requests.

The first thing visible after launch is a reduction in fuel consumption of 8–15% on comparable routes.

Part of it is delivered by a trained recommendation model: it analyzes acceleration, braking, gear selection, and idling for a specific vehicle and shift, and sends a hint to the driver before he burns extra fuel.

Then equipment downtime goes down. The system monitors deviations in the operating parameters of units and warns about them in advance, so repairs are planned in the window between trips, not at the moment of failure on the line.

In parallel, the maintenance schedule moves from the calendar to actual operating hours: consumables are replaced when they have used up their life, not by the date in the log.

Engine life grows as a consequence of the same regime: fewer cold starts under load, less operation at elevated RPM without need, more stable thermal regime. For a fleet, this means a longer service life before major overhaul and fewer unscheduled stops for the same unit.

A separate effect is reduced load on dispatchers. Recommendations on route, load, and fueling arrive during the shift, rather than being analyzed retrospectively at month end, so decisions are made where something can still be changed.

The result is verified for each vehicle before and after: fleet fuel savings become a number in the report, not a promise.

Operating modes: analytics, hints, auto-restrictions

Operating modes differ in degree of intervention, so they are enabled in stages: first analytics by routes, then driver hints, and only then automatic restrictions and fueling control.

The calculation part is common: only whether the conclusion reaches a person or becomes a control action on the vehicle changes.

Switching is done by vehicle groups, the mode profile is stored in the configuration, so returning to the previous option and comparing periods takes minutes, not a separate implementation project.

In practice, modes are combined: analytics and fueling control work continuously, while restrictions are enabled where statistics on the route have already been accumulated.

Such a set gives fuel cost reduction within 8–15% without stopping the fleet and without the risk of "clamping" a vehicle on a difficult section.

Implementation: from data audit to handover of regulations

An AI fuel consumption optimization project begins not with a model, but with data: as long as it is unknown what telemetry writes and at what frequency, any forecast is built on guesses.

Therefore the first step is an audit of sources and assessment of their suitability for the specific task.

1. Telemetry data audit. We analyze onboard logs, level sensors, trips, and fueling: polling frequency, gaps, sensor drift, discrepancies between systems. The output is a data quality map with a list of gaps that we close before training.
2. Integration scheme. We determine where data comes from and where the result goes: gateways, queues, intermediate storage, accounting and dispatch systems. We record delivery delay and behavior during communication loss.
3. Model configuration. We collect features: route, road profile, load, seasonality, driving mode, idling with the engine running. We select the forecast horizon and trigger thresholds so that recommendations are actionable for the driver.
4. Historical validation. We run the model on closed periods and compare it with actual consumption, including trips with repairs and emergency situations. We look not at average error, but at the stability of the effect across the entire fleet.
5. Pilot operation. We launch on a limited group of vehicles in recommendation mode, without automatic intervention. We verify forecast against fact every shift, adjust thresholds, and collect dispatcher reaction.
6. Handover of regulations. We provide the procedure for working with reports: what to do on deviation, how to replenish data, when to retrain the model. Dispatcher training and an incident contact point close the project.

What the customer retains: models, schemes, regulations

The system for fuel consumption control and reduction lives not in reports, but in configurations, schemes, and regulations that remain within the customer's environment.

We hand over the set so that operation can continue without us: the dispatcher shift works by instruction, the analyst sees data sources and quality, and model retraining does not require finding the project authors.

• Configured models — working configurations with a feature set, trigger thresholds, and applicability boundaries by equipment types and routes, plus a retraining procedure on new data.
• Data source map — a list of fuel consumption and level sensors, fueling and drain data, positioning, fleet reference data, with arrival frequency and the owner responsible for each source.
• Integration documentation — exchange schemes with telematics, fuel, and accounting systems: formats, polling frequency, behavior during communication loss and repeated export.
• Access policies — roles of dispatcher, analyst, and fleet manager: what is visible in the interface, what is exported, who is allowed to change trigger thresholds.
• Anomaly response regulation — drain, shortfill, overconsumption on a trip, long idling with the engine running: who receives the signal, within what time, and what action is prescribed.
• Dispatcher instructions — how to parse a signal, confirm or reject it, leave a comment on the trip so that feedback enters the data for retraining.
• Model quality log — share of false positives, trip coverage by calculation, discrepancy between calculated and actual consumption by control groups.

Acceptance is based on these artifacts, not on a demonstration: customer engineers check the source map and access policies, dispatchers are trained on their own fleet. Describe the task — we will return with an assessment and a handover plan.

Fleet case study: metrics before and after implementation

A fleet of 60 units: long-haul tractors, medium-duty trucks, and delivery vans. Before the start, consumption was calculated from waybills and fuel receipts — the discrepancy with actual fill reached 9%, and overconsumption was written off to season and traffic jams.

There was no single standard per route: the same model on different legs showed a spread of up to one third.

What changed was not the vehicles, but the accounting loop. Tank level sensors and onboard bus reading were installed, and consumption, engine hours, and load for each trip were consolidated into one report — Wialon.

Then the standard was calculated per route, not per model: load, terrain, number of stops, season. A deviation of more than 10% on a leg raised a signal to the mechanic on the same day, not at the end of the month.

The effect was distributed unevenly, and this is important for planning. On long-haul legs, savings reached 15% — there the section is long and even and the load is predictable.

On short delivery routes, it was 6–8%: the leg is shorter, and idling during loading accounts for a larger share of the shift. It was its reduction by almost half that gave the main part of the result — and it required less than replacing vehicles or fuel.

How does consumption control differ from consumption forecasting?

Fuel consumption control works on fact: tank level telemetry, standard flow meter, mileage, and coordinates.

The system compares actual consumption with the norm for the route and marks the discrepancy — a drain, a shortfill during fueling, overconsumption due to driving style or a malfunction. A management decision is made after the trip, when the fuel has already been consumed.

Consumption forecasting is built on the same data, but answers a different question — how much fuel will be needed for a trip before it starts. Predictive analytics keeps route history, season, cargo weight, terrain, and idling during loading in the model. The output is a fueling and loading plan, not an analysis of deviations.

The difference between the approaches is in horizon and cost of error. Control provides feedback: the deviation is visible almost immediately, but the money has already been spent.

Forecasting helps prevent overconsumption, but it errs more the less clean history has been accumulated — the model learns from sensor noise and one-off trips.

Therefore, in practice they are not separated. First, order is established in the data: calibration of level sensors, cutting off false jumps on slopes and warm-ups, linking fueling to primary documents.

On such a basis, control stops generating false signals, and the forecast fits within a few percent per trip. Then "how much to fill" and "where it went" answer one question from two sides — we examine how this is addressed.

Technical doubts: data, integrations, security

Technical doubts usually come down to three things: whether data quality is sufficient, how the solution will fit into the accounting loop, and who will get access to sensitive figures.

Integration with the accounting system 1С and the dispatch system is built on regular exchange, not manual exports: trips, fueling, and routes are pulled automatically, and discrepancies are reconciled automatically.

Will equipment need to be retrofitted?

There is no mandatory fleet replacement. The system relies on the data that the dispatch system already has; if it is insufficient for accuracy, standard level and flow sensors are added — this is retrofitting individual vehicles, not re-equipping the entire fleet.

Who will see what in the system?

Access is granted by roles: the dispatcher sees trips and deviations, the mechanic sees consumption and equipment condition, the manager sees the summary picture.

Data is protected during transmission and storage, exports are recorded in the log, and rights are revoked with one action.

Let's discuss the task: what data is needed for assessment

Project assessment relies not on general considerations, but on an export for a specific fleet: the model calculates consumption by driving modes and compares it with the standard, so data completeness directly determines what share of deviations can actually be caught.

Below is what is worth preparing for the conversation. If some data is missing, this is not a stop: we will suggest what to replace it with and how to collect what is missing.

• Telemetry for several months: consumption, RPM, engine load, time under load and at idle — linked to vehicle and time.
• Data on fueling and drains: statements, fuel cards, level sensor readings, to compare actual consumption with calculated.
• Routes and trips: waybills or exports from the monitoring system — length, road type, terrain, share of urban sections.
• Fleet composition: vehicle types, their purpose, age, and standard consumption by each group.
• Load: cargo weight, number of trips, share of empty runs, seasonal peaks and downtime.
• Reference data and constraints: maintenance regulations, driver work rules, who approves process changes.
• Contact on your side: an engineer or dispatcher who knows how the data was collected — will help see gaps and distortions.

Using this set, we calculate the scope of work, choose a baseline for comparison, and define the pilot boundaries on part of the fleet — usually one group of vehicles with comparable routes. Describe the task — we will return with an assessment and a plan.