The substrate problem
substrate underneath it.
The vendor pitches are confident, the board pressure is intense, and the proofs-of-concept are multiplying. Underneath, a contradiction is forming: agent budgets are rising and agent ROI is not. Of four AI agents we observed in production at LATAM mining operations through 2025, the two with clean telemetry foundations paid back within a quarter; the two without became compliance liabilities — one disabled after recommending setpoints from a calibration record seven years stale. The difference was never the model. It was the substrate. This deep dive argues the position vendors dislike and boards find uncomfortable: most natural-resources operations are not ready to deploy autonomous agents — and shows exactly what to fix first.
The four failure modes of agents on bad substrate — and why the most discussed (hallucination) is the least dangerous, while the most expensive one stays invisible for months
The five foundation layers every high-stakes agent depends on, from master data to human-in-the- loop telemetry — simultaneously necessary, not phased
The five diagnostic questions to ask before signing the next agent pilot — the cost of asking is one conversation; the cost of not asking compounds silently
CIOs, CDOs and VPs of Operations in LATAM mining, cement, pulp & paper and agribusiness deciding whether their operation is ready to give an AI agent decision authority
