Real-time ore grade prediction has quietly moved out of research and into production at three Australian mines that I can speak to with confidence. The tools have matured enough that the conversation has shifted from “can we make this work” to “what does the operating model around it look like.”

What the technology actually does

The core capability is integrating real-time sensor data — XRF, NIR, hyperspectral imaging, sometimes acoustic — with the geological block model to produce a continuously-updated picture of grade at the working face. The historic block model is informed and corrected by live data.

The mine knows what it is mining with higher confidence and less lag than it used to.

The operational impact

Three categories of benefit appear consistently. Blend variability into the processing plant has reduced. Ore loss to waste stockpiles has fallen. Dilution at the high-grade-low-grade contacts has tightened. Average impact across these three operations: a 2% to 4% improvement in metal recovery and a 6% to 10% reduction in misrouted material.

What it took to make it work

None of these implementations went smoothly. The first attempt at one operation stalled for about eighteen months because sensor calibration drifted faster than the model could compensate for. Trust took a year to rebuild after the calibration process was overhauled.

Another operation had to substantially rebuild its data infrastructure to handle the volume and latency the prediction model needed. The compute and networking work was bigger than the AI work.

The team structure that is working

A consistent pattern is team composition. None of these operations are running this with just a data science team. The configurations that work all have a geologist, a mining engineer, a metallurgist, and a machine learning engineer working as a permanent unit.

That is not the structure most AI consulting engagements deliver. It is the structure these operations have built deliberately.

Where external help has earned its place

The early model-building work is one place outside specialists have proved valuable. For mining companies thinking through how to build this capability, custom AI development firms that understand the production environment, not just the model architecture, are the ones to talk to. The conversation about MLOps in an operations-critical context is different, and the firms that get that build models that survive the transition to operations.

What is next

The next step that two of the three operations are pursuing is integration with the autonomous haulage routing. Today, the ore grade prediction informs human-mediated routing decisions. Tomorrow it informs autonomous routing decisions directly, with the human operator overseeing rather than executing.

That transition is the work of the next eighteen months. The operations that have built the team structure through the current generation of the technology will move quickly. The operations starting now will be playing catch-up for several years.