On 12 November Nick Carreras and I ran the third Pulse webinar, for the people who finance and buy mining assets: royalty and streaming companies, private equity and credit funds, banks, brokers and corporate development teams. The subject was deal origination, and the argument was that most of the work around it is done in isolation, one stage at a time, often by different teams. The full recording is below, with the three client case studies we walked through and the questions the room asked.
Why does mining deal origination run in isolation?
Whatever the mandate, mining finance does the same five jobs. Origination produces a target list. Comps and benchmarking produce a peer table. Diligence produces an investment committee paper. Execution produces a signed deal. Portfolio and risk management produce a watch list. Administration, from board packs to investor reporting, sits around all five.
The timelines and risk appetites differ between a royalty company, a credit fund and a producer's corporate development team. The jobs do not. What they share is that each stage is usually run on its own. A mandate lands and a fresh search starts. The comp set is rebuilt by hand. The diligence work is dropped when a deal is passed, and has to be redone when it comes back. Each output is handed on by email and spreadsheet, if it is handed on at all. Even with AI in the building, most teams choose to run these stages separately.
What problems do mining finance teams raise most often?
Between August and November 2026, Pulse audited 138 calls with clients and prospects. Of the 38 organisations that described their problems on those calls, these are the shares that raised each one:
| Problem | Share of organisations |
|---|---|
| Monitoring means reading everything | 37% |
| Origination is a one-off hunt each time a mandate lands | 37% |
| Data you cannot trust: inaccurate, stale or unverified | 29% |
| AI on raw data: cost, and output that sits outside your tools | 24% |
| Diligence evidence buried inside reports | 21% |
| Comps built by hand, and not like-for-like | 21% |
| Hand-offs between teams and systems | 21% |
Source: Pulse Client Problem Inventory, 138 client and prospect calls, 12 August to 12 November 2026.
One royalty company put the trust problem precisely: the validation is the killer, because the difficulty is not finding information but the number of versions of it. A metric reported three or four different ways across a company's filings, or a project spelled two ways by two divisions, is enough to send a model in the wrong direction. At royalty companies, which run lean by design, it is common to find people at VP and SVP level spending 200 to 300 hours a year validating technical metrics for onward reporting.
What are Goldilocks conditions?
Goldilocks conditions are the deal written down before the search starts. The traditional sequence is to find a possible deal and then spend weeks working out whether it should be a deal at all. The alternative is to define the deal first, in quantitative and qualitative terms (commodity, jurisdiction, stage, scale, cost, cheque size, permitting status, ownership) and let the system find every asset that fits.
The list is not a one-off output. It stays current: projects join when a study publishes or a permit is granted, and leave when a condition stops being true. Origination becomes something pushed to the team rather than something the team goes looking for.
What is the difference between sequential and vertical automation?
Two words carried the rest of the session.
Sequential, or horizontal, automation hands each stage's output to the next. The target list becomes the comp set, the comp set and the target list together feed the diligence instructions, and all three become the watch list and the investment memo. Each stage can carry its own Goldilocks conditions.
Vertical automation is depth inside a single stage, as far down as the question needs. A criterion as narrow as more than 3 g/t over more than 5 m across at least ten holes. A processing cost the feasibility study never prints, rebuilt from three of its own cost tables. Every drill hole behind a resource estimate. That depth is only reachable because 83M+ drill intercepts and 100M+ pages have already been extracted, validated and normalised.
Case study: how does a royalty company screen for a US$10-20M royalty?
A royalty company wanted royalties that would pay US$10-20M a year at the book's own 1-2% NSR, in precious metals, on projects at development stage.
| Stage | Result |
|---|---|
| Define the deal | US$10-20M a year at a 1-2% NSR; precious metals; development stage |
| Origination | 1,309 development-stage projects, of which 505 carry an economic study |
| Valuation | 116 projects where the modelled royalty prices inside the cheque |
| Diligence | 49 of the 116 carry no existing royalty; 60 already have one |
| Monitoring | A weekly origination brief as studies publish and projects enter or leave the band |
The valuation is a simple financial model applied project by project: annual metal production from each project's own study, spot prices, mine life capped at 20 years, discounted at 5%. Without an economic study there is no production profile to value, which is why the study filter comes first. The existing-royalty check runs against Pulse's royalty and stream records, public and private, so the team knows before the first call whether someone is already in the neighbourhood.
116 of 1,309 projects fit the cheque, each sized from its own study.
Case study: how does a mining private equity fund build a target chain?
A mining private equity fund set four rules: gold as the primary metal, tier-1 jurisdictions, 100,000 ounces a year or more, and all-in sustaining cost at or below US$1,500 an ounce.
| Stage | Result |
|---|---|
| Origination | 46 assets pass: 10 producers and 36 developers, with 30 more on a watchlist |
| Comps | 10 operating peers on a subject project's processing route, with cost and recovery from each report |
| Diligence | 262 of 274 holes behind a target's resource pre-date its current owner; 5 discrepancies found |
| Monitoring | Past diligence targets watched for permits and high-grade drill results |
The same criteria were put to a general AI model searching the open web. It returned 13 developers against the 36 that pass in Pulse. What separated the two results was the material underneath: a validated system of record in one case, whatever the model could assemble from the open web in the other.
36 developers pass in Pulse; a general AI model on the open web found 13.
Case study: how can a credit fund see a financing need before the mandate?
A credit fund did not need to write its conditions down. Its own ten deals since 2019 set the ticket size, the capex band and the entry point, and those were applied to the market.
| Stage | Result |
|---|---|
| Origination | 57 projects in the band, across every commodity |
| Qualification | 29 companies with an open funding need: 11 live, 12 in 2027, 6 royalty-first |
| Diligence | Median capex moved +18% from pre-feasibility to feasibility |
| Monitoring | 29 names watched for mandates, term sheets, cost, cash and permits only |
Across the band, 11 of 12 recent financings followed a study that was already on file, a median of 58 months earlier. The signal was public long before the mandate. The watch now reports movement on 29 names and nothing else, which is the part that matters: knowing what to ignore.
11 of 12 financings followed a study already on file, a median of 58 months earlier.
How is a financial model built from mine data?
This was the first question from the room. In the live part of the session, the royalty screen above was run through the Pulse MCP in Claude, restricted to Pulse data, and returned as a ranked Excel workbook: the top 25 projects, years to first production, potential revenue, and the theoretical value of a 1% and a 2% royalty, every row carrying its source link.
The method is a plumbing exercise. The model's structure comes either from the client's own model, whether it has ten fields or 250, or from what the AI infers the question needs. Each field is then filled from Pulse's 850+ structured metrics, drawn from technical reports, financial reports and company disclosure. Where a metric is not standard, the AI derives it from the base metrics, says that it has done so, and shows the formula.
How do you know the number is right?
Every asset name, company name and metric that enters Pulse is validated against a canonical value: one true reference for each thing in mining, whether it is a project that two divisions spell differently or a grade reported in three units. Five successive feasibility studies on one project sit in one place, so the evolution of each number is visible. Pulse's AI now spends more of its effort on validation than on ingestion.
For the analyst, that means a system of record. Ask a general model the same question twice and the answer moves, because there is nothing fixed underneath it. Ask it against a validated record and it reads the same information every time, and every figure opens its source page so it can be checked in one click.
Where does client data sit?
Pulse runs two kinds of work. The public layer shown in the webinar is a web platform, and the Pulse MCP connects a client's own AI to it in one direction only: the AI asks Pulse for information and Pulse draws nothing out of the client's environment. Work on a client's proprietary data, from digitising geological archives to automating internal workflows, is single-tenant and end-to-end encrypted, hosted as standard on EU-based AWS, on Pulse infrastructure, or entirely on the client's own.
How is Pulse priced?
Pulse is priced by organisation and use case, not per seat. The fee depends on the workflows built and the data provided, so it is scoped in conversation rather than read off a rate card.
Does automation replace the analyst?
No, and it is not designed to. The aim is to compress the time that is not needed for thinking as close to zero as possible: the searching, re-keying, reconciling and re-reading that sit between an analyst and a decision. The judgement stays with the person paid to make it, who can open any step, any figure and any source document whenever they want to.
Three takeaways
- Define the deal before you search for it. A written set of Goldilocks conditions turns origination from a hunt into a maintained list.
- Let each stage hand its output to the next. The target list should become the comp set, the diligence pack and the watch list without being rebuilt.
- Judge the system by what you no longer read. The most valuable output of monitoring is the noise it removes.
The Pulse webinar series
This was the third session in the series. The first, in July, asked how to get decision-grade answers from AI in mining finance and corporate development. The second, in September, covered why AI fails on geological data and what it takes to digitise and normalise a century of exploration records.
About Pulse Intelligence
Pulse Intelligence is an AI enablement company for the mining industry. We digitise, validate and structure technical reports, exchange disclosure, government archives and historical geological records (100M+ pages, 32,000+ mining assets, 83M+ drill intercepts, 850+ metrics and 500 to 700 news releases a day) and deliver them through the platform, a documented API, the Pulse MCP for Claude, ChatGPT or your own model, and Excel. Every figure opens the original page of the original document.
To scope a workflow on your own criteria, contact hello@pulseintelligence.com.
Less searching. More strategising.™
Less Searching. More Strategising.™
See the platform running on your own deal criteria.