Can AI Optimise Haulage Costs?

The problem, in numbers

In an open-pit mine, most of the money goes into moving material. A 2026 peer-reviewed review of autonomous haulage puts haulage at 50% to 60% of total mining costs.^{Source 1} Every tonne of ore and waste rides a truck from the face to the crusher or the dump, and the trucks burn diesel, wear tyres and need a driver in the cab around the clock. In our assessment, that makes haulage the first place to look for a cost saving in an open pit.

Australia's iron ore miners have spent more than a decade testing whether software can take that cost down. By January 2018 Rio Tinto had more than 80 autonomous trucks across five Pilbara sites, which had moved more than a billion tonnes.^{Source 2} By November 2020 Fortescue's autonomous fleet had reached 183 trucks.^{Source 3} In 2019 BHP said it was assessing automation of 500 haul trucks across Western Australia and Queensland, with a capital budget its chief financial officer put at under US$800 million.^{Source 4}

What improves, and why

AI reaches haulage costs in three ways. The first is autonomy: the truck drives itself on a planned course, controlled from a remote centre. The second is dispatch: software decides which truck goes to which shovel and dump, in real time, to cut queues and empty running. The third is energy: models tune speed, payload and road condition to burn less fuel per tonne.

Autonomy does the heavy lifting in the evidence, because its savings are structural. A truck without a driver does not stop for a shift change or a meal break. Rio Tinto estimated that each of its autonomous trucks operated about 700 hours more than a conventional truck in 2017, at around 15% lower load and haul unit costs.^{Source 2} A peer-reviewed assessment of driverless trucks at a typical large remote Australian open-pit mine pointed the same way in 2011: automation would cut labour costs and add productive hours.^{Source 7}

The site matters as much as the software. The 2026 review of 50 studies found that turning radius is the primary constraint on where autonomous trucks can safely go, that gradient limits of 8% to 12% are still recommended, and that GIS-based and integrated optimisation approaches report productivity gains of 5% to 20%.^{Source 1} That is what it means for a site to be engineered for autonomy: haul roads, intersections and dump layouts are built around what the trucks can do.

What the evidence actually says

The cost figures that matter are company-reported. Rio Tinto's release gives its 15% load and haul saving for 2017 as an estimate and states no method, and in 2015 the company had been reported as citing 12% higher productivity and 13% lower load and haul costs.^{Source 2} Fortescue reported "a greater than 30 per cent increase in productivity levels" in 2020 without stating the baseline, metric or period.^{Source 3} In 2017 an executive at the parent of WesTrac, which supplied Fortescue's autonomous trucks, put Fortescue's productivity gain at 20% and its pit to port costs 43% lower.^{Source 3} That is a vendor-side statement, and the report does not say how much of the cost fall came from the trucks. In our reading the figures point the same way, but they move between years and definitions, and we found no independent audit of any of them.

The optimisation layer has thinner evidence still. One peer-reviewed AI model, built on data from a Queensland coal mine and an Iranian iron mine, reported fuel reductions of 9% to 12%, but those are the model's outputs, not savings measured on trucks in service.^{Source 5} Dispatch research is mostly simulation: one study compared dispatching strategies in a stochastic simulation built from a real mine's data, and noted that such comparisons had been missing from the literature.^{Source 6} The 2026 review named field validation as a gap.^{Source 1}

The economics also depend on wages. A peer-reviewed comparison of China and Australia attributes Australia's lead to high labour costs and remote operations, and finds adoption in China constrained by lower wages and a large mining workforce.^{Source 8} In our reading, the saving depends on drivers being expensive and hard to house. It is not a universal per-tonne discount.

Where it falls short

• Capital first, savings later. Retrofits, remote centres, communications and redesigned roads come before any saving, and BHP put the capital budget for its proposed program at under US$800 million.^{Source 4}

• Self-reported results. Every operator figure here comes from the operator or its supplier, and the metrics change between announcements.^{Source 2,Source 3}

• Optimisation is mostly modelled. Fuel and dispatch gains in the literature rest on simulation and models more than measured fleet trials.^{Source 5,Source 6}

• Labour costs drive the case. The saving shrinks where wages are lower and the workforce is large.^{Source 8}

Where governance decides the outcome

Two cost pressures need a check. The first is the boundary with manned equipment: the saving must not come from relaxing separation rules, skipping commissioning steps or pushing software changes without change control, because that is where autonomous fleets meet people. Our companion review covers that risk in detail. The second is the workforce. Driverless haulage removes driving jobs and shifts work to remote centres in cities, with effects on remote towns.^{Source 7} Operators have promised redeployment: BHP said half of a 400-person operations team had retrained from heavy vehicle roles, and Fortescue said it had trained about 3,000 people to work with autonomous haulage.^{Source 3,Source 4} Those are company claims. Publish the transition plan, consult workers before the trucks arrive, and report what happened to the people who used to drive them.

How to think about adoption

Start with your own cost structure. If haulage is a large share of cost, drivers are expensive and the mine has many years of life, autonomy has a case worth studying; if not, dispatch and fuel optimisation on a manned fleet may be the better first step. Price the site work, not just the trucks: roads, intersections, communications and the control room. Set a baseline for load and haul unit cost, operating hours and fuel per tonne before anything changes, so the saving is measured on your fleet rather than taken from an announcement. Treat every vendor and operator figure, including those here, as a hypothesis to test.