Mining is the canary for whether AI lands in heavy industry or stays in the slide deck. The work is brutal, the cycles are long, the data is messy, and the people who run mines have heard a thousand vendors promise to revolutionise their operations. None of that has stopped Rio Tinto and BHP from quietly building the most ambitious industrial AI stacks on the planet.
What's interesting now is that the headline numbers are no longer projections. They're operating metrics — auditable, in production, on real ore.
For technology leaders watching from outside the sector, that statistic does the same thing the OpenAI launch did for software: it ends the abstract debate about whether AI works at scale and forces the conversation onto what to copy and what to skip.
The numbers that ended the debate
Three operating data points side by side. Rio Tinto's 300th Komatsu AHS truck was delivered in August 2024; the fleet now spans 10 of its Australian sites and contributes ~80% of daily production capacity. BHP has converted 65% of its Pilbara haul fleet to autonomous operation as of FY2024-25, building from its first fully autonomous deployment at Jimblebar in 2017 — a programme that has reduced significant truck-related events at Jimblebar by close to 90%. And AutoHaul — Rio Tinto's autonomous heavy-haul rail network — runs 220 trains across more than 1,866 km of track in the Pilbara, the largest deployment of autonomous rail anywhere in the world. Since July 2018, AutoHaul has logged more than 33 million km of operation with over 98% of missions arriving at their destination on schedule.
These aren't pilots. They run continuously, in 50°C heat, across thousands of kilometres of track, hauling tens of millions of tonnes of ore a year. The downtime numbers, safety numbers, and throughput numbers all moved in the right direction simultaneously — which historically is the hardest test for any industrial automation programme, because most efforts trade one for the other.
The second wave is operating-layer AI inside the equipment itself. Sandvik and Epiroc embed computer-vision models in their crushers and sorters to assess ore quality in real time, letting operators adjust grade control as the rock changes — the kind of decision that used to wait for an assay lab and a shift hand-off. BHP went further on the chemistry side: in May 2023 it partnered with Microsoft to put Azure-based ML into the Escondida copper concentrator in Chile, letting operators tune recovery variables against live process data rather than weekly reports. Escondida is the world's largest copper mine, and the deployment was later expanded to a second concentrator at the same site.
Exploration: heat maps over hunches
The flashier story is in greenfield exploration, where AI has changed the unit economics of finding deposits in the first place. Kobold Metals (battery metals) and Earth AI (multi-element targeting) are the two reference cases — both running models that ingest geological surveys, satellite imagery, geophysical readings, and historical drilling, then output ranked target maps that focus the drill on the most prospective ground. The commercial validation is still emerging — these are early-stage exploration plays and discoveries take years — but the change in workflow is real today.
BHP's exploration alliance with Ivanhoe Electric, signed on May 8, 2024, is the more interesting industrial-scale signal. The deal commits $15 million of BHP investment over three years across six areas in Arizona, New Mexico, and Utah, with the option to set up 50/50 joint ventures on any prospects that emerge. Ivanhoe brings its proprietary Typhoon geophysical survey system plus the machine-learning algorithmic software and data-inversion services of its subsidiary Computational Geosciences. Two majors with deep historical datasets, building shared ML for copper and critical-minerals detection — explicitly framed as a way to compress exploration timelines and reduce ecosystem impact per discovery dollar.
The lesson for any AI agents and autonomous systems programme outside mining is the same one BHP and Rio Tinto learned the hard way over a decade: the model is downstream of the data plumbing. Greenfield exploration AI works when you have decades of clean drilling logs, unified survey data, and well-labelled rock types. The companies investing now in their data substrate are the ones that will find their AI delivering returns in three years; the ones treating it as a procurement decision will not.
The operational layer: throughput, planning, maintenance
The value pool from operational AI in mining is real but not magical. BCG's published mining-and-metals research puts the throughput uplift from mine-to-mill optimisation at 2-5% for mature deployments — small as a percentage, large in absolute terms when you're moving hundreds of millions of tonnes a year. The bigger gains come from compounding: better throughput on top of better predictive maintenance on top of better planning, all running against the same instrumented dataset.
Predictive maintenance is the most boring part of this stack and the one with the cleanest payback. Equipment-health models running on sensor streams from haul trucks, conveyors, and crushers catch failures before they cascade. Once the model is calibrated to a specific fleet, the unplanned-downtime curve flattens. The next maturity step BCG and others flag is agentic maintenance — agents that don't just predict failures but schedule fixes, re-route work, and order parts. The early movers running this are typically two to three years into their data programmes, not new entrants.
The third operational layer is precision drilling. Petra and Novamera deploy AI-guided robotic systems that drill with surgical accuracy in narrow vein deposits without blasting — which sounds incremental until you account for the environmental and recovery economics. For deposits that previously weren't viable to mine, this changes the answer.
The data problem nobody talks about
The thing that separates the firms running production AI from the firms running pilots is unglamorous: clean block models, well-instrumented fleet telemetry, and unified data lakes that don't fragment by site. BCG's research is unusually direct on this point — without clean data foundations, GenAI simulations of bottlenecks across the exploration-to-production chain don't work, no matter how capable the model.
This is the part of mining AI that maps most directly to other industries. Every operational AI programme we've shipped at Axccelerate, in mining-adjacent and non-mining contexts, has run into the same rate-limiting step: the data plumbing has to come first. That's why our work on the unglamorous data and infrastructure layer — model routing, eval harnesses, drift detectors — pays for itself many times over against the more visible model-and-prompt layer above it. The miners that figured this out a decade ago are the ones with the operating numbers above. The ones that didn't are still piloting.
Talent is the second constraint. The cross-section of geologists, mining engineers, and ML practitioners is a small group, and most of them already work for the majors. For mid-tier operators, the realistic path is hybrid teams — internal domain experts paired with specialist external partners — rather than trying to hire a full ML capability from cold start.
What this means outside mining
Mining is unusual in the speed and visibility of its AI deployment, but the playbook generalises. Three observations apply to any heavy-industry, infrastructure, or operations-heavy CTO reading this:
The cost-of-doing-nothing is now measurable. Five years ago you could defend a wait-and-see posture on industrial AI on the basis that the technology wasn't proven. Rio Tinto and BHP took that question off the table. The metric that matters now is the productivity gap between firms that have shipped operational AI and firms that haven't — which compounds each cycle.
Mature use cases compound, early-stage ones don't. Autonomous haulage and ML-driven throughput optimisation are mature and shippable today. Agentic exploration, GenAI mine simulation, and asset-design AI are still maturing. Treat the mature ones as line items; treat the early ones as research bets.
Data infrastructure is the rate limit. No model is going to outperform poor block-model data, fragmented fleet telemetry, or siloed sensor streams. This is why we keep returning to the same conclusion in client work: spend on the unglamorous data layer first, then layer agents on top. The reverse order doesn't work.
The most interesting agentic-AI programme in heavy industry isn't a chatbot or a copilot. It's a system that integrates exploration models, predictive maintenance, mine planning, and procurement automation into one coordinated stack, with humans approving the high-stakes decisions and the agents handling the deterministic high-volume work. That's the same pattern we see across every AI workflow automation build that holds up at scale, in or out of the mine.
The ten-year head start that Rio Tinto and BHP have built is real, but the playbook isn't proprietary. The question for any operations-heavy business is which year of that ten-year curve you start on.
