AI in geological modelling has stopped being a slide in a vendor deck and started showing up in actual workflow. Talked to half a dozen geologists across iron ore, gold and copper operations in the last fortnight to get a read on where the tooling is working and where it still falls over.

The short version: the easy work is mostly automated. The hard work is mostly not.

What’s working

Drillhole data ingestion and QA is the single biggest productivity gain. What used to take a junior geologist two days a fortnight — checking assay data against logging notes, flagging duplicates, reconciling lithology codes — now runs as an overnight job and produces a cleaner output. The reduction in obvious data errors has been substantial. The gain isn’t in fancy machine learning; it’s in disciplined automation of work that was always tedious.

Variogram modelling is a mixed picture. AI-assisted variography produces reasonable starting points and saves hours on the obvious cases. The cases where it matters most — anisotropy detection in structurally complex deposits, nugget effect estimation in nuggety gold, sill identification in transitional ore types — are still mostly done by human geostatisticians. Two of the geologists I spoke with said the AI tools quietly increase their workload because the rough drafts they produce need careful auditing before they’re trusted in a resource model.

Implicit modelling with AI assistance is where the genuinely useful work is happening. The tools can produce surface meshes from point data faster and more consistently than humans can. They handle complex topology better than legacy software. The gain is in iteration speed: a geologist can run ten model variations in the time it used to take to produce one. Whether you trust any of them is still a human judgement.

For the more complex implementations — site-specific model integration, custom data pipelines into operational systems, governance frameworks for AI-driven resource models — most of the operations I spoke with brought in a consultancy to handle the engineering work. Building this in-house takes a team mining houses generally don’t have.

What’s not working

Resource classification is still a human call. AI tools can suggest classification boundaries based on drilling density, geological continuity and grade variability, but the JORC code and the auditors haven’t caught up to AI-generated classifications and probably won’t for a while yet. The Competent Person sign-off remains a personal accountability that doesn’t transfer to a machine learning model. None of the geologists I spoke with thought that should change soon.

Mine planning integration is rough. Resource models produced with heavy AI assistance often don’t slot cleanly into legacy mine planning software. The data formats, the metadata expectations, the validation checks all break in subtle ways. Operations running modern integrated platforms cope. Operations on older toolchains find the AI-assisted work creates more downstream problems than it solves.

The most awkward case is when AI models contradict legacy hand-modelled interpretations. The AI is sometimes right. Sometimes it’s confidently wrong because the training data didn’t represent the local geological setting. Telling the difference is genuinely hard, and the cases where it matters most are the cases where confidence is least justified.

How teams are actually using it

The pattern that works is using AI tools for the high-volume, low-judgement work — data QA, basic statistics, surface generation — and reserving human time for the calls that have real consequences. The pattern that doesn’t work is treating AI as a productivity replacement for senior geological judgement.

The mines I spoke with that are getting genuine value have invested in three things: clean data infrastructure that the AI tools can consume, training programmes that bring senior geologists up to speed on what the models are actually doing, and governance frameworks that record what was AI-generated versus human-generated in the resource model. None of that is exciting. All of it is the work.

In 12 months we’ll know whether this is the start of a real productivity step-change or whether the gains plateau early. Right now the honest answer is: it’s helping, but less dramatically than the marketing suggests, and only for teams that did the unglamorous foundational work first.