Pulse Intelligence Co-founder/Director Will Coetzer and mining advisor Scott North on the failure modes behind unreliable AI outputs, and the discipline required to produce numbers you can defend.
Last week Will Coetzer and Scott North ran through something we get asked about constantly: why generic AI keeps producing unreliable results on mining data, and what it actually takes to fix that. If you missed it, the full replay is embedded below.
Want to see it run on your own data? Book a demo →
They walked through the specific failure modes: sources that can't be verified, documents that don't get read fully, and numbers that can't be defended in a meeting. The theme throughout — a general-purpose model is a powerful tool, but it is not a source of truth. Ask it to build a comp table or find the leading producers in a commodity and it will pull from whatever it happens to find, in whatever currency, with a citation that may already be dead by the time you open it in the boardroom.
Scott North gave a live demonstration of the M&A monitoring workflow he runs every day — a single prompt that scans the last day's market announcements and returns a structured, source-linked report in minutes. Will walked through the data integrity framework Pulse provides to clients: the guardrails and validation that sit between raw filings and a number you can actually use.
One figure worth holding onto: querying pre-extracted, validated data is anywhere from 100 to 300 times less compute-intensive than sending the same question to a raw document — and that cost difference compounds fast at scale.
The session closed with a simple checklist for judging any AI-generated number before you trust it: where it came from, when it's from, whether the currency and entity are comparable, whether it was reconciled or just retrieved, whether the model actually finished reading, and whether you could defend it in a meeting.
Less searching. More strategising.™