The cheapest line in your AI budget is the intelligence
The 95% of enterprise GenAI initiatives with no measurable return (MIT NANDA) and the 72% of mining organizations reporting only partial benefits (NTT DATA × MIT Technology Review) tell half the story. The other half: the initiatives that did pay are public, and they share one design choice — the AI tied to the metallurgical variable that reconciles to margin, not to a department’s local KPI. This deep dive reads Huang’s definition of intelligence as the build order of an orchestration layer designed for the global optimum, and the yardstick for auditing any pilot. The arithmetic is public: one point of recovery at world scale is worth on the order of US$ 100 million a year.
Huang’s 2022→2026 arc read as an architecture specification, and why the commoditization of intelligence moves the margin to the design of the decision space.
The public cases with verified figures (Freeport Bagdad: ~10% more throughput, ~1 pp more recovery; Antofagasta Minerals: +0.75 pp at Centinela) and what separates them from the stalled-pilot pattern.
The four questions that make any AI proposal auditable, a concentrator’s decision catalogue, the three look-alikes of automation — and the one-afternoon test to classify your portfolio before the next board meeting.
Boards, COOs, operations directors, and AI program owners in copper mining and adjacent heavy industry across LATAM.
