From Fuel Receipts to Operational Intelligence

Fuel looks like a procurement line item until operations maps its full path.

It is ordered from multiple suppliers, delivered into tanks, measured at uncertain temperatures, transferred, dispensed to vehicles or generators, consumed during productive work and idling, and reconciled after the fact. Each step may live in a different system. Small variances accumulate into stock loss, emergency purchases, idle equipment, and disputes.

This is a strong bespoke-model opportunity because the task is narrow, the data is proprietary, and the outcome can be measured in liters, hours, cost, and production.

The economic signal

Two mining cases illustrate the operational leverage of refueling decisions.

Mine Tech Services reports that optimized refuel assignments across 30 haul trucks saved 1,242 hours per year and enabled an additional 348,000 tonnes of movement, valued at nearly $7 million in revenue. [Veridapt](https://www.veridapt.com/news-and-press/case-studies/pitstop-performance) estimates that cutting average waiting by one minute recovered 990 truck-hours per year and$1.7 million in production profit.

Both are supplier-reported cases, so the exact return should not be generalized. They reveal the mechanism: the cost of fuel operations is often dominated by lost equipment time, not only the price per liter.

Build the digital fuel ledger

Before modeling, create a reconciled event model:

• purchase order and contracted quantity
• supplier, terminal, and batch
• dispatch and arrival time
• bill of lading, invoice, and receipt
• tank level before and after delivery
• temperature and density adjustment where relevant
• meter or dispenser reading
• vehicle or asset identity
• operator identity and authorization
• issue quantity and timestamp
• telematics fuel level and consumption
• engine hours, distance, payload, route, and idle time
• maintenance state and production output

Every record should retain its source. Document AI may extract an invoice quantity, but the system should link the value to the exact page and field. An SLM can normalize a supplier description, but it should not erase the original.

Four task-specific models

Document extraction and matching

A document model extracts supplier, order, product, quantity, unit, price, tax, vehicle, seal, and time from receipts and invoices. A matching service compares them with the purchase order, tank movement, and payment record.

The core metric is field-level precision and recall on local document layouts. Operational metrics are straight-through match rate, manual review minutes, duplicate detection, and dispute cycle time.

Consumption forecasting

A forecast should operate by site, asset class, and time horizon. Inputs can include production plan, engine hours, payload, route, season, maintenance, and historical consumption. Quantile forecasts are more useful than a single number because procurement must plan uncertainty.

Measure weighted absolute percentage error, bias, coverage of prediction intervals, stockout events, and emergency-buy volume.

Anomaly detection

Not every variance is theft or leakage. It may be a late posting, meter calibration, temperature effect, transfer loss, idling, changed payload, or maintenance degradation. Use a layered detector:

1. deterministic balance checks
2. time-series anomaly detection
3. peer-group comparison by asset and duty cycle
4. an explanation model that assembles evidence without declaring guilt

Rank cases by expected recoverable value and evidence quality. Send each case to a human investigator with receipts, meter readings, telemetry, and comparable periods attached.

Refueling optimization

The optimizer decides when and where an asset should refuel, subject to tank capacity, queue, route, service level, production plan, and safety constraints. It should produce a feasible recommendation and explain which constraint drove it.

Measure queue minutes, productive hours lost, runout risk, tanker utilization, schedule adherence, and production effect.

Why a bespoke model beats a generic dashboard

Fuel behavior differs by asset, shift, operator pattern, product mix, supplier, route, temperature, and site. A generic threshold such as "flag usage 10% above average" will flood the team with false positives.

A bespoke system learns expected consumption conditional on work performed. It can compare two loaders only when payload, haul profile, idle policy, and maintenance condition are comparable. It also understands local aliases for sites, tanks, assets, and suppliers.

The language model is not the calculator. It is the interface that explains a reconciled numerical model, retrieves supporting policy, and collects investigator feedback.

A variance waterfall

For each period, calculate:

Opening stock + verified receipts - verified issues - closing stock = unexplained variance

Then split verified issues into productive consumption, idle consumption, test or maintenance, transfer, and unresolved. Do not ask a model to reason over raw totals without this accounting spine.

The target is not zero variance. It is a stable, explained variance within measurement tolerance, with high-value exceptions resolved quickly.

KPI framework

Segment all KPIs by site, supplier, tank, asset class, shift, and duty cycle. A fleet-wide mean hides the operational cause.

A 90-day implementation

Days 1 to 20: Map the physical and financial fuel flow. Reconcile units, clocks, asset IDs, tank IDs, and supplier names. Quantify missing data.

Days 21 to 45: Build the document extractor, balance rules, and baseline forecast. Back-test anomalies against known events. Define the human investigation workflow.

Days 46 to 70: Add telemetry-conditioned consumption models and a daily exception queue. Operate in shadow mode. Measure precision at the top 10 and top 20 cases.

Days 71 to 90: Pilot procurement forecasts or refuel recommendations at one site. Require supervisor approval. Compare against a matched baseline for queue time, stockouts, variance, and production loss.

The business case formula

Annual value can be separated into:

• avoided price premium from emergency purchasing
• reduced unexplained fuel loss
• reduced inventory buffer
• productive hours recovered from refueling
• reduced idle consumption
• avoided stockout downtime
• lower manual reconciliation labor

Subtract sensors, integration, model operations, review labor, and change management. Apply a confidence factor to vendor or modeled benefits. Never combine savings that overlap, such as idle-fuel reduction and total consumption reduction.

The conclusion

Fuel intelligence is a good enterprise AI wedge because it connects documents, sensors, operations, and money. It is large enough to matter and narrow enough to validate.

The winning system is not a chatbot that discusses fuel. It is a digital ledger with specialist extraction, forecasting, anomaly, and optimization models, plus an evidence-backed interface for the people who act.

Research note

Research is current through September 5, 2026. Mining examples are used to show the economics of mobile industrial fleets and are supplier-reported. Each site should validate measurement tolerances, safety constraints, and realized financial value.

Continue the research

• Manufacturing AI solutions beyond computer vision
• Frontline voice and data capture for manufacturing
• A 90-day path from AI pilot to production
• The hidden economics of enterprise AI
• AI fuel and haulage optimization for mining

Building a Production-Ready System

Conscious Engines builds AI fuel consumption optimization systems that reconcile purchases, storage, transfers, dispensing, asset telemetry, and production demand. We model expected use, identify explainable variance, and measure savings against output so lower consumption is not confused with lower production.