How AI fits into mining diligence without replacing expert judgment
How AI fits into mining diligence without replacing expert judgment
A mining investment decision is consequential. It involves capital allocation, geological uncertainty, operational constraints, and regulatory accountability. The people who make these decisions are qualified professionals: geologists, mining engineers, metallurgists, commercial analysts, and investment leads.
AI is useful in this workflow. But the line between what the machine does and what the human does has to be explicit, not blurred.
What the machine can do
AI is a set of replaceable capabilities attached to the workflow:
- Extract. Read a 400-page technical report and propose the material claims: grade, tonnage, recovery, ownership, access, infrastructure.
- Compare. Flag where two sources disagree on the same figure. Two resource tables reporting different tonnage. A financial model assuming 88 percent recovery when the lab data shows 71 percent.
- Retrieve. When someone asks where a number came from, find the exact page, table, and document.
- Draft. Propose a screening memorandum from the reviewed application state: a reproducible view of the information already inside the system, not a piece of independent writing.
- Monitor. Watch for new evidence (new drill results, new permits, new metallurgical data) and flag which claims and decisions may need review.
All of these write provisional outputs into the same review and provenance model. The machine does not make the final call.
What the human does
People interpret geology, judge materiality, resolve ambiguity, approve conclusions, and retain accountability.
- The geologist reviews the resource estimate and exploration potential.
- The mining engineer reviews the proposed mining method.
- The metallurgist reviews the recovery assumptions.
- The commercial analyst reviews ownership, capital requirements, and project economics.
- The investment lead approves the decision.
SubsurfaceOS records who reviewed what, what they concluded, and where they disagreed. Two experts may interpret the same evidence differently. The system records both interpretations, their assumptions, and the authority that resolved or accepted the disagreement.
Why the line matters
If the system produces an opaque score, the decision cannot be defended. When a committee, a regulator, or a buyer in diligence asks you to stand behind a number, you need the trail: which source it came from, who reviewed it, what they disagreed on, and what has changed since.
A model can regenerate the answer. It cannot reconstruct that trail afterwards. The trail has to be built while the work happens.
Uncertainty must remain decomposed
Missing evidence, poor evidence, contradictory evidence, geological uncertainty, commercial uncertainty, and unresolved review are different conditions. They must not be collapsed into one confidence score.
The 88 percent recovery claim is not "70 percent confident." It is supported by two recent tests, contradicted by one older test, and missing variability data across the full deposit. Those are three different states, and the decision depends on understanding all three.
The principle
AI prepares and checks the work. People make and approve the decision. The system preserves disagreement rather than forcing consensus. Every material conclusion exposes its sources, contradictions, assumptions, reviewer, approval status, and revision history.
Generated text without this chain is not decision-grade.