How AI Stopped Gold Mines From Sending Ore to the Waste Dump

For decades, some of the most sophisticated gold operations on the planet have been quietly making an expensive mistake. They have sent valuable ore to the waste dump and fed barren rock into their mills, not through carelessness, but because the human-drawn structural models guiding those decisions could not see the invisible chemical halos that actually govern where the gold sits.

That is the problem artificial intelligence in gold mining now solves. The shift is not futuristic. It is already an operational reality on major assets, where machine learning has quietly rewritten the economics of established mines.

By replacing visual structural mapping with multi-element geochemical modelling, operators are pulling more gold from the exact same volume of blasted rock. No extra digging. No extra blasting. Just better sorting.

What comes next gives you a clear framework for understanding how machine learning actually works at the pit face, which metrics prove it is working, and how to separate genuine technological change from marketing gloss when you assess a mining asset.

Why traditional structural domain models fail in complex gold systems

To understand why AI matters here, you first need to see what the old approach gets wrong. Conventional grade control leans on geostatistical methods such as kriging, a technique that estimates the grade of unsampled rock by averaging nearby known values. Kriging assumes relatively simple, linear relationships between geology and grade.

That assumption breaks down in Archean gold systems, some of the oldest and most geologically complex deposits on Earth. Here, the gold is not neatly controlled by faults, veins or lithological boundaries. It is governed by subtle geochemical halos and alteration fronts, with grade shifting sharply over very short distances.

Manual structural mapping routinely misses this fine-scale variability. The result is systematic misclassification: ore mistakenly routed to waste, waste mistakenly fed to the mill. Both errors cost money on every truckload.

Multi-element geochemical domain models take a different route. Instead of drawing boundaries from mapped structures, machine learning recognises complex, non-linear patterns across vast assay and sensor datasets, defining domains by geochemical and mineralogical signature rather than visible geometry.

The data density required is demanding. Accurately modelling grade distribution in Archean vein systems calls for drill spacing of 5 by 10 metres, far tighter than the coarse grids used at many exploration-stage projects.

Where that data exists, the accuracy gains are measurable. Maptek’s DomainMCF, an AI engine that builds domain boundaries directly from multi-variable sample data, reconciled within approximately 2% of actual volumes in tested cases. Operations running real-time machine-learning grade control report reductions in ore misclassification of 30% to 50% compared with conventional blast-hole sampling and kriging.

Here is the distinction that matters:

• Traditional structural domaining: defines ore by faults, veins and lithological boundaries; relies on kriging’s linear assumptions; misses short-range geochemical variability; produces systematic misrouting of ore and waste.
• AI-driven geochemical domaining: defines ore by multi-element geochemical and mineralogical signatures; learns non-linear grade relationships; captures fine-scale variability; sharply reduces misclassification.

The core problem AI solves is not digging rock faster. It is stopping the costly habit of sending ore to the waste pile and waste to the mill. That makes AI adoption a fundamental upgrade to a mine’s baseline profitability, not a marginal efficiency tweak.

The Detour Lake case study and the 20 percent throughput breakthrough

The theory becomes concrete at Detour Lake, the mine where this approach was first proven in a live production environment. John Florek, a geologist with roughly 35 years of experience across BHP, Placer Dome and Barrick Gold, served as Chief Geologist there and pioneered AI-based grade control at the operation.

His methodology traced back to Australian mining practices he encountered while working with BHP in the early 1990s, which handled Archean-hosted deposits more effectively than the North American approaches of the time. The team he worked with became active around 2016 and spent roughly 18 months collecting data before deploying the new geochemical models in Q4 2017.

The deployment did two remarkable things. It compressed three-dimensional deposit modelling from a process that traditionally took about two years down to just one to two months. And it increased gold mill throughput by 20% without moving a single additional tonne of rock.

That 20% gain is the number to understand. It came entirely from correcting misclassification, sending more genuine ore to the mill and less to the waste dump, rather than from mining more material. When you read a mining company’s operational updates, this is the distinction to hunt for: throughput gains achieved through better sorting accuracy, not just heavier blasting.

Current digital integration at Detour Lake

Detour Lake is now 100% owned by Agnico Eagle, and it has become the company’s flagship property. The 2025 full-year results confirm the operation’s scale: 692,675 ounces of gold milled from 27.9 million tonnes at an average head grade of 0.86 g/t, with overall recovery of 89.9%. Mill throughput averaged roughly 76,353 tonnes per day, and 2026 guidance targets 700,000 to 730,000 ounces.

The digital footprint has widened well beyond the original grade control work. According to a January 2025 report from Symboticware, the site has implemented the 4-Sight.ai platform to monitor mobile equipment, fuel use, operator behaviour and mechanical condition, adding predictive maintenance and inefficiency detection.

Machine learning and geochemistry now also support fragmentation and blending to sustain higher mill throughput. This is what mature AI integration looks like on a balance sheet, and it gives you a benchmark against which to test similar claims from other operators.

Mine-to-mill optimisation frameworks extend this logic beyond the dig line, connecting blast design, fragmentation outcomes and plant feed characteristics into a single data loop where grade control decisions upstream have measurable effects on recovery rates downstream.

How AI grade control is performing across global mining operations

The Detour Lake approach has not stayed put. Geologists trained under Florek’s methodology have carried it to other major Canadian operations, including Greenstone, Côté Lake and Canadian Malartic, spreading a technique that began at a single mine across a growing portfolio of assets.

The financial outcomes elsewhere follow a strikingly consistent pattern, and they are showing up on more than one continent. Consider three deployments where the numbers are documented:

1. IntelliSense.io, large Australian gold mine. Between December 2023 and February 2024, the site reclaimed approximately 400,000 tonnes from digitally modelled stockpiles using IntelliSense.io’s Stockpile and Inventory Optimization solution. Average plant feed grade rose from 0.80 g/t to 0.84 g/t, generating more than AUD $2 million in additional value over the campaign period, according to Mining Technology’s Excellence Awards coverage in October 2025.
2. Éléonore underground mine, Quebec. Goldcorp, now part of Newmont, implemented a machine-learning ore-body knowledge system that integrated underground mapping, grade-control drilling and production reconciliation. The system is reported to have reduced the gap between planned and actual ore grades by approximately 30%.
3. Touquoy mine, Nova Scotia. St Barbara’s operation used OREPro3D blast-movement modelling to predict ore displacement and refine dig lines, an approach credited with improving grade by roughly 7% through better ore-versus-waste classification.

These systems operate dynamically. Conveyor-mounted analysers, hyperspectral scanners and drone imagery feed high-frequency grade data into models that update block estimates in near real time, at the scale of individual dig blocks.

Broader adoption of computational intelligence in resource extraction is reshaping cost curves across the mining sector, with operators reporting that the same pattern-recognition capability applied to grade control is now being extended to processing plant optimisation, fleet dispatch and energy management.

What this tells you is that the financial upside of AI grade control does not depend on geography. Whether the mine sits in Canada or Australia, the improvement in feed grade and classification accuracy follows a measurable pattern you can track in quarterly production reports. That gives you a practical way to judge which operators are genuinely extracting maximum value from their stockpiles and dig lines.

The hidden infrastructure costs of making mining AI work

Now for the cold reality. None of these systems work without dense, high-quality geochemical data, which means the algorithms are only ever as good as the drilling and sampling that feed them.

That dependency is not negotiable. Accurate Archean vein models demand drill spacing as tight as 5 by 10 metres, a data density that is expensive to achieve and operationally disruptive to acquire. A company touting AI capability while running coarse, cheap drill grids is promising an outcome its data cannot support.

The integration challenge runs deeper still. Extracting real value requires aligning information technology with operational technology on the ground, connecting physics-informed digital twins, machine learning and material tracking across run-of-mine stockpiles, flotation and plant operations. Isolated algorithms deployed without this backbone routinely disappoint.

Analysts are candid about the risk. S&P Global Market Intelligence framed the state of play plainly in its February 2025 assessment.

A peek at AI revolution in mining: promise meets peril.

That framing captures the analytical caution you should apply. AI grade control delivers where data quality, change management and domain expertise are strong, and it falters where those foundations are weak.

For readers wanting to evaluate whether a specific operation has the physical and organisational foundations to make these systems work, our dedicated guide to AI readiness in mining walks through the infrastructure checklist, covering data architecture, sensor networks, and the change-management requirements that separate successful deployments from expensive pilots.

For you, this is the protective layer. When a producer or explorer touts AI capability, check whether the drilling budget and sensor networks exist to support the algorithms. Their absence is a red flag that the technology claims may be running ahead of the physical data acquisition needed to make them real.

Valuing mining assets in a data-driven production era

The move to AI-driven grade control changes more than throughput. It reshapes the risk profile and cash-flow timing of a gold operation, because bringing misrouted ounces forward and stabilising mill feed pulls value into the present rather than deferring it.

It also changes who does the work. As these tools automate routine tasks such as dig-line design and classification, geologists and engineers shift toward model validation and complex exception handling, a change in skill profile you should expect to see reflected in how modern operations are staffed.

The forward view links this to compliance. The same real-time monitoring systems that optimise grade are increasingly used to minimise cyanide use and manage environmental performance, tying ESG obligations directly to the digital infrastructure. The operators who build that data backbone well are the ones best positioned for the next production era.

The connection between digital mining and environmental performance is tightening as real-time monitoring systems originally built for grade optimisation are repurposed to track cyanide consumption, water use and emissions, creating an integrated compliance layer that ESG-focused investors are beginning to price into their assessments.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results, and several performance figures cited here are reported by their respective sources and have not been independently verified.