OES Puts 30 AI Agents on Mining Haul Trucks and Lifts Haulage by 10%

A coal open-pit mine runs more than 240 pieces of equipment, and over 10,000 deviations from the norm pile up in a single shift. A dispatcher with a radio physically cannot react to all of them. Engineers at Operational Excellence (OES) did not replace the monitoring system. Instead, they built a layer of 40+ models and 30+ AI agents on top of it. On the same fleet, rock mass haulage rose 10%, excavator productivity rose 17%, and the dispatcher's workload fell 60–70%.

A Dispatcher with a Radio vs. 10,000 Deviations

The fleet consists of BelAZ 75131 (130-tonne) and 75306 (220-tonne) mining dump trucks and more than 25 excavators. Before the rollout, traffic was balanced by hand: the dispatcher watched a screen and phoned drivers. By OES's estimate, a dispatcher handles less than 0.1% of deviations per shift, and conventional reporting captures only 10–20% of real losses.

The problems were concrete. On average, an excavator lost one trip per shift because no truck was available, and the dispatcher learned about it 10–15 minutes later, after the losses had already built up. Equipment left the line 30–40 minutes before the end of a shift and came back late, with up to 25% early departures and late returns. An unplanned refuel cost 15 extra minutes on average, because nobody staggered the machines in time. The customer asked OES to cut excavator downtime and raise haulage without touching the existing system.

Two Modules on Top of Someone Else's Dispatch System

The first module, OES Truck Balance, is an AI balancer. It contains 40+ models of different types:

• a neural-network ETA algorithm that calculates each vehicle's arrival time;
• a balancing algorithm that evaluates dozens of allocation options every 3 minutes;
• text2speech and speech2text for voice communication with operators.

The system forecasts queues 30 minutes ahead and manages refueling departures: two or more trucks from the same unit cannot leave at the same time. When a vehicle breaks down, it redistributes the fleet instantly. Directives arrive on a tablet in the cab.

The second module, OES MAS, is a multi-agent system of 30+ specialized agents, each with its own area of responsibility:

• Refuel handles the standard refueling time for each equipment model;
• LoadCycle tracks loading cycles and deviations from the plan;
• ShiftChange manages shift handovers and the positioning of equipment at the start of a shift;
• Speed monitors technical speed on each segment of the route;
• Quality checks loading quality;
• Idle records downtime;
• Navigation monitors the health of navigation equipment.

Every deviation has a life cycle: event, warning, violation, directive. Escalation has three levels: driver, dispatcher, shift supervisor. Each deviation is immediately converted into rubles, so the loss is visible in monetary terms.

Telemetry as the Bottleneck

All the "smart" parts live in the cloud, but they are fed by Galileosky 7x and 10 terminals installed across the entire active fleet. The terminals collect GPS position, fuel level, engine parameters, speed, and bucket and dump body status, with data arriving every 1–5 seconds. Fuel and pressure sensors also register loading and unloading events and monitor refueling.

The data path runs from the terminal to the OES IIoT platform, then to the AI modules, the driver's tablet and the dispatcher interface. The platform ingests the telemetry and converts data from BelAZ, Komatsu and Caterpillar machines into a single format. Everything depends on the accuracy of the input: if a vehicle "disappears" for 30 seconds or a fuel sensor produces noise, the queue forecast drifts. That is why the terminals were chosen for their resistance to vibration, dust and temperature swings.

The rollout began in 2024. From February to April, a prototype ran on a single unit of 3–5 excavators with manual balancing by a coordinator. Commercial operation began in May across the entire fleet of 220-tonne trucks and 20-cubic-meter excavators. June went to experiments with balancing and ETA accuracy. In July, the team added the 130-tonne trucks, Komatsu PC2000 excavators and directives on tablets. Starting in October 2024, the system was synchronized with equipment positioning, and refueling dispatch was automated.

What They Got and What to Consider Before Copying It

OES metrics:

• rock mass haulage +10% on the same fleet;
• excavator productivity +17%;
• truck queue time at loading −30%;
• excavator downtime −15–25%;
• excavators left without a truck during a shift −20%;
• dispatcher workload −60–70%;
• 95% of directives are carried out without a call to the dispatcher;
• 100% of losses are recorded in real time, versus 10–20% in conventional reporting;
• 15 minutes saved on each refuel (Refuel agent);
• early departures and late returns at shift changes down 25%;
• economic effect exceeded 20 million rubles in the first month of MAS operation.

An example from a master shift in August 2025: of 64 recommendations, 61 were carried out; the plan was 210 thousand m³, the actual result 233 thousand m³ (+11%).

For those who want to repeat the approach, three things matter. First, an AI layer can be built on top of existing monitoring without replacing hardware, but forecast quality is limited by the frequency and cleanliness of the telemetry. Second, narrow agent specialization simplifies debugging: for each deviation, it is clear who owns it and how much it costs. Third, the authors themselves stress that the result came from several factors: processes changed in parallel, from manual balancing by a coordinator to drivers switching to working with directives. Without this organizational part, the numbers would most likely have been more modest.