In three months, miners profiting from AI went from 4 to 15 — and only one plant application has reached the bottom line

McKinsey tracked 19 major miners and built a maturity ladder for each artificial intelligence application. The result shows where AI already pays its way, where it is still a promise — and what separates one from the other.

At Mining Forum Americas, held at the end of September 2026 in Colorado Springs, in the United States, McKinsey presented a review of how artificial intelligence is reaching the financial results of major miners. The consultancy tracks 19 companies in the sector. In the third quarter, through September 11, 15 of them had reported financial gains attributed to AI. In the previous quarter, there were 4. Mineral processing leads: 10 of the 19 companies report bottom-line impact from applications in their plants. To organize what works and what does not yet, McKinsey classified each type of application on four rungs of evidence. At the top, "proven in P&L", are tools in production with disclosed financial figures. Just below, "proven", are deployed tools with measured effects. Next, "emerging", applications that meet only part of the criteria. At the bottom, "aspirational", what is not yet in production. Ferran Pujol, a Santiago-based McKinsey partner, summed up the message in one sentence: a pile of AI initiatives is not a transformation. The test, he said, is whether the area starts to run differently, not whether it has better tools. The panel included Ravi Malladi, senior adviser for data and AI at Freeport-McMoRan, and Akilan Kapilan, Microsoft's director for energy and resources.

McKinsey's ladder is useful because it dismantles the idea that "AI in mining" is a single thing. Each application is at a different stage, and the pattern has a technical logic. In processing, only automatic tuning of grinding and flotation parameters has reached the top rung. That is no coincidence: the plant is the most instrumented environment in a mine, with continuous sensors for flow, density, particle size and grade, and every percentage point of metallurgical recovery becomes immediately measurable revenue. Computer vision, ore blending and stockpile optimization sit one rung below, with measured gains but no financial figure disclosed yet. Leaching and water and reagent optimization remain emerging, because they involve slow chemical kinetics and less data per hour of operation. In maintenance, shutdown planning has already reached the top, and predictive maintenance of fleets and plants is on the "proven" rung. In mine planning and operations, however — scheduling, drill and blast, dispatch, collision avoidance, electric fleet charging, mixed-fleet autonomy — almost everything is still between emerging and aspirational. And the three exploration applications tracked (target generation, automated core logging and grade control) are still emerging. The practical lesson is that AI pays off first where data is already dense, continuous and reliable. Ravi Malladi, from Freeport, gave the panel's most concrete example: if an equipment oil sample does not reach the lab, or the result does not get into the system, no predictive maintenance model will make a difference. The bottleneck is data discipline on the mine floor, not the algorithm. McKinsey also described the organizational profile of companies that scale, which it calls the 70-70-70 rule: about 70% of technology talent kept in-house, 70% of the team building solutions rather than coordinating vendors, and 70% drawn from senior ranks. Add to that technology teams linked directly to operating teams and a single executive who controls both the budget and the financial outcome of the initiative. Akilan Kapilan, from Microsoft, completed the picture with a limit on use: models should recommend options, not make unchecked decisions, and it is up to the operations specialist to set the boundaries, because models without reliable data produce wrong answers that look certain.

For people working in mining, the ladder works as a priority map. If a company is going to start or reorganize its AI program, the higher rungs show where return has already been demonstrated by other companies: advanced plant control and shutdown planning. The lower rungs show where investment in the data foundation is still needed before expecting results. The 10% to 15% EBITDA gain estimate also helps size the conversation with the board — and shows that the gain comes more from cost than from volume. For professionals, the message is just as direct: the most valued skills are not just data science, but the combination of process knowledge, governance of operational data and the ability to turn a model recommendation into a change in operating routine. The survey itself notes that, in energy and materials, only 7% of companies have scaled AI agents in manufacturing — and that mining is probably below that average. There is plenty of room for those who can bridge the plant and the algorithm.