Why AI Fails in Mining Investment: Origination, Comps, and Diligence — and What Decision-Grade Actually Takes.

On 20 July 2026, Scott North and I are running a webinar for corporate development, PE, fund and royalty teams on why AI fails in mining investment origination and assessment.

We will cover where AI genuinely helps across origination, comps and diligence, where it breaks, and how to verify what it gives you.

We also cover what is required to enable using the latest LLMs like Claude and ChatGPT's best models to achieve 8–20x token cost reduction, and use on average 200x less compute whilst simultaneously achieving enterprise-grade outputs 10x faster.

You'll leave with a 90-second test you can run on any AI vendor tomorrow.

The problem we will address

Every deal you screen comes down to a handful of derived numbers (any combination of more than 400 individual metrics), be it grade, reserve, recovery ratios, AISC, risk metrics, etc.

Whilst a general AI confidently hands you a result in seconds, it never shows its working. No audit trail. No system of record.

For a capital allocator, that breaks in three predictable places.

Origination

You ask what's out there, and the model names 38 lithium companies. The real universe is 283. You're not screening a market, you're screening whatever it happened to remember. You can't automate anything, and you have to validate just about everything yourself anyway.

Comps and benchmarking

You build a peer set and two reserve figures come back both "correct" and 2.6x apart, because nobody stated the accounting basis. A comp you can't source is a comp you can't defend. You use a legacy platform, but there's no way to easily confirm whether the metric is correct or even current.

Due diligence

Non-standardised data. Not AI-ready. Manual. The number the whole deal rests on traces back to nowhere. No filing, no page, no lineage. Fine for a first look. Not fine for an IC submission.

And we'll close with a live demo — the same mining question, the same data, put to a vanilla model and to Pulse-connected AI, side by side.

If AI has been fast but never quite trustworthy, this is the session that shows you why, and what decision-grade actually takes.

15:00 UK · 20 July 2026 · Online. RSVP link in the comments.

Less searching. More strategising.™