The useful question about AI in exploration is where it changes the next decision. Which ground should an explorer acquire? Which survey should it run? Where should the next hole go? In our view, treating every company answering one of those questions as the same business is an excellent way to buy the wrong tool.
There is no independently comparable league table in the sources reviewed for this piece. Here is a more useful shortlist: companies with different approaches, followed by the public evidence that deserves attention.
| Company | What its public material describes | Evidence to distinguish from the pitch |
|---|---|---|
| Octavia | An early-stage platform combining geophysics and proprietary AI for discovery, expansion and drill planning, focused on gold, copper and uranium. | A named, publicly documented drill-confirmed outcome attributable to its platform: NOT FOUND in the reviewed material. Watchlist, rather than a demonstrated discovery ranking. |
| Esper | Hyperspectral satellite imagery and mineral mapping, including a Northern Territory map offering. | A mapping product is a different milestone from an economic deposit. A named drill-confirmed discovery attributable to that product: NOT FOUND in the reviewed material. |
| Atomionics | GRAVIO gravity sensing and ORE-O AI-assisted three-dimensional interpretation. | New measurements plus modelling. Its website's performance multipliers are vendor claims; the sources reviewed do not provide an independently comparable discovery benchmark. |
| KoBold Metals | Mineral exploration, with Mingomba in Zambia its prominent copper project. | Project-partner documentation provides a development milestone, but also a history predating the later AI narrative. |
| ALS Geoanalytics | Geoscience with data science and machine learning; ALS renamed its GoldSpot service business in 2024. | Historical GoldSpot case studies connect target-generation work with named holes and published assays. |
| EARTH AI | AI targeting followed by its own field validation and drilling. | Legacy Minerals' ASX reporting documents work and exploration results at Fontenoy. |
Those distinctions matter commercially. Ask a sensing company about field repeatability and survey cost. Ask an interpreter about data quality and geological assumptions. Ask a targeting company whether the prediction was recorded before the hole, how many targets failed, and whether the improvement survived outside its training ground. These are diligence questions, not findings against any company.
Has old data actually helped?
Yes, in reported programmes combining archived information with new work. GoldSpot's 4 May 2020 Patents release explicitly describes integrating filed assessment reports and public geophysics with its geologists' 2019 field data. It also says fieldwork was essential to checking historical records and selecting useful model inputs. The reported result was 12.8 grams of gold per tonne over 0.5 metres in hole MTU-20-14. That is a narrow drill intersection, not evidence of an economic mine. The release itself cautioned against assuming further drilling would repeat the result.
A more striking historical example is GoldSpot's work with New Found Gold at Queensway. Its 29 January 2020 release describes compiling historical information, adding field mapping and using traditional geology alongside machine learning. It linked that programme to 19 metres grading 92.86 grams of gold per tonne in NFGC-19-01.
That is a reported downhole length, not a true-width resource calculation. GoldSpot was commercially interested in New Found Gold. The assay is evidence of mineralisation; the release is the vendor's account of AI's contribution, not a controlled experiment isolating an algorithm from the geologists.
Our view: the filing cabinet can contain information worth revisiting, but these cases do not show that uploading old reports alone creates a deposit. The old maps got another shift. So did the people in boots.
Bigger successes, with the right labels
For development scale, Mingomba belongs in the discussion. ZCCM-IH reported the launch of shaft sinking on 29 April 2026. Its Lubambe annual-report extracts also document the Extension Project and its transfer into Mingomba. Describing the whole story as an AI discovery from untouched ground would erase that history. Shaft development is a meaningful milestone; it does not by itself establish operating production or financial returns.
For an Australian drilling example, Legacy Minerals' 25 November 2025 ASX release reported that EARTH AI had earned a 51% Fontenoy project interest after meeting the first-stage spending and drilling conditions. It reported 120 metres at 0.30 grams per tonne combined palladium, platinum and gold from 298 metres in hole EFO7D, including 10 metres at 1.2 grams per tonne from 388 metres. The broad interval used no grade cut-off. These are dated downhole exploration results, not a palladium-only grade, resource estimate or operating mine. They demonstrate field validation, without proving economic extraction. The reported ownership milestone is dated November 2025; it is not a freshly verified October 2026 ownership statement.
The strongest counterargument to our caution is straightforward: a tool can save meaningful time or money without discovering a giant mine. Better survey coverage, cleaner historical data and fewer poorly chosen holes could be valuable. The problem is proving that improvement against a fair baseline, including unsuccessful work, rather than displaying the best hole in the deck.
Opinion: The most persuasive leaders will join prediction to evidence: dated targets, disclosed drilling, repeatable results and understandable costs. KoBold, the historical GoldSpot case studies and EARTH AI offer public project evidence worth examining. Octavia, Esper and Atomionics deserve attention for different capabilities, without assuming their promises have already become mines.
Commercial company claims are attributed; absence of an outcome in this review is not proof that none exists.