Mine planning is one of those domains where AI vendors have been promising disruption for the better part of a decade. The reality has been slower and more incremental — partly because the underlying problem is constrained by physics, geology, and engineering tolerances, and partly because mine planners have a healthy scepticism about black-box systems telling them where to put a pit wall.

In May 2026, the situation is more interesting than the marketing suggests. AI components inside the major mine planning packages are being used in production on a non-trivial number of Australian sites. The headline use cases are not the ones the vendors talk about most.

What’s actually in production

Three categories have moved from pilot to standard practice across the Australian operations I’ve been talking to.

Block model interpolation with machine learning is now common where the geological dataset is rich enough to support it. The output is not a replacement for the geological model — it’s a faster way to test sensitivity and produce reasonable interpolations across drill spacing where traditional methods get conservative. Geologists I’ve talked to describe it as “another tool in the toolbox” rather than a revolution.

Schedule optimisation with AI-assisted heuristics is being used on operations with complex multi-pit, multi-product flowsheets. The classic mixed-integer optimisation problem is hard. The AI-assisted versions narrow the solution space and produce schedules that planners then refine. Two iron ore operations and one gold operation I know of have moved from “AI-assisted” being a research project to it being a standard part of the quarterly planning cycle.

Equipment dispatch optimisation is the most quietly successful category. The systems that route trucks, decide which face to load next, and respond to delays in real time are now sophisticated enough that the operations centre running them sees a measurable productivity improvement on the same fleet. The improvement is not always dramatic — typically two to five percent — but on a Pilbara operation that adds up to real money.

What’s still in pilot

Several categories remain in pilot or limited deployment, and the reasons are worth understanding.

Geotechnical AI is where the marketing is loudest and the production reality is thinnest. The data needed to train models that predict slope failure or wall behaviour is enormous, and most operations don’t have it in a clean form. The AI tools that exist tend to flag anomalies in monitoring data rather than predict actual failures with useful lead time. Engineers are using them. They’re not betting on them.

AI for blast design has had some traction but the value proposition is harder to demonstrate. Blast outcomes are influenced by enough confounding variables that attributing improvements specifically to AI versus other changes is contested.

Reserve estimation with AI remains controversial inside the geology community. The reserve number underpins the financial value of the operation. Replacing the traditional kriging-based approach with machine learning models has regulatory and technical implications that the industry is still working through.

The integration problem

The repeated theme across operations is that the AI components in the planning stack only work as well as the data plumbing around them. An AI schedule optimiser fed by stale survey data is less useful than a manual schedule built by an experienced planner with current information.

Operations that have invested in the underlying data infrastructure — the historian, the survey data pipeline, the equipment telemetry, the geotechnical monitoring stream — get more out of their AI tools. Operations that bolted AI onto a fragmented data stack get the level of value you’d expect.

For larger operations thinking about how to sequence the investment, this is the pattern I’d point to. Data foundations first, AI on top, then specific use cases. The operations that have flipped this order have generally regretted it.

For the actual delivery of the AI components, I’ve seen Australian operations work with both internal data science teams and external partners. The internal teams know the operation. The external partners bring delivery discipline. The combination works best when the external partner is engaged for delivery and capability uplift, not just for a one-off project. AI strategy support of this kind has helped a couple of operations sequence the investment more sensibly than they would have alone.

What the planners actually want

The AI tool wishlist from mine planners I’ve talked to is not what the vendors lead with in their marketing. The most-requested capabilities are mundane.

Better scenario testing. The ability to run a hundred variants of next year’s plan and look at the distribution of outcomes, not a single deterministic schedule.

Better integration with the geotechnical monitoring data. A plan that automatically flags when a scheduled cut intersects with a monitored zone of concern.

Better handling of equipment uncertainty. A schedule that knows the truck fleet has variable availability and produces a robust plan, not a brittle one.

These are not exciting demo material. They are what makes the day-to-day work of planning faster and more reliable.

Where this is heading

By the end of 2026, I expect the gap between the leading and lagging operations on AI-assisted planning to widen. The leading operations are compounding their data and process advantages. The lagging operations are still trying to consolidate their data infrastructure.

The honest answer to the “should we adopt AI in mine planning” question depends on where the operation sits on that data and process maturity curve. AI on a strong base is delivering. AI on a weak base is still mostly disappointing.

The vendors who tell you otherwise have something to sell.