How AI Is Actually Reshaping Investment Research: Evidence from 4.2 Million API Queries

A friend’s data on 4.2 million research queries reveals something the headlines keep missing.

Most of what gets written about AI in finance assumes the same plot. Analyst sits at a desk. AI eats their lunch. Headcount drops. Tidy story. Also wrong in a specific, interesting way.

A friend of mine (who writes as MokumKiwi over on Substack) just published **a piece drawing on 4.2 million real API queries** hitting their company’s research stack.

The number that surprised me most: about 1 in 5 queries comes from agents scanning hundreds of stock tickers (240–350+) every single week.

No analyst has ever done that. Not because they didn’t want to but because they couldn’t afford to.

The work that didn’t exist

What makes this strange is that it isn’t a faster version of old work. It’s a new category. 78% of those queries come from buy-side firms. Funds that, historically, would deep-dive maybe 40 to 60 names at a time, because each name cost analyst-hours. The economics simply forbade casting a wider net.

Now an agent sweeps 300 tickers every Monday morning for the cost of a moderately fancy lunch.

This isn’t automation in the layoff-the-junior-analyst sense. It’s the thing economists keep promising and rarely delivering: a genuine expansion of what’s possible to do at all. You don’t replace headcount with this. You change what the headcount points at.

The split that explains the industry

The other detail that’s been rattling around in my head: sell-side and buy-side are using the same technology for almost opposite ends.

• Sell-side (Goldman, JPM, Morgan Stanley). 74% of their queries cluster around investment theses, filings, and transcripts. They’re going deeper on names they already cover.
• Buy-side. They’re going wider, surveilling universes they previously couldn’t afford to look at.

This isn’t a quirk. It’s the business models rendered in query data. Sell-side gets paid to publish; depth is the product. Buy-side gets paid to find mispricings; breadth raises the odds of finding one. Same tool. Opposite use.

If you want a single mental model for how AI lands in any industry, this is it. The tool doesn’t dictate the use. The economics of the user do!

The cost line nobody mentions

One number in the piece I keep coming back to: A complete investment thesis run that used to consume 14 to 18 analyst hours now resolves in 75 minutes at roughly $50 per run.

The headline framing (“80% faster!”) obscures the more important shift. The unit economics of analytical thinking just changed. A $50 thesis run means a junior PM can speculatively explore ten investment hypotheses in an afternoon for the cost of dinner. That’s a different job than the one that existed in 2023.

The 7% of queries the piece labels “Consensus Synthesis” makes this concrete. Investment committees are now running AI-generated devil’s advocate passes against 7 to 12 broker reports before they vote. The conviction call is still human. The substrate the conviction sits on is not.

What actually becomes valuable

The piece ends on a point worth pulling out: as the infrastructure layer collapses into commodity APIs (filings parsing, financial data ingestion, transcript retrieval), the moat moves upstream. Prompt design. Agent logic. Discernment about which question to ask.

That last one is the only part I’d push back on. Discernment about which question to ask has always been the moat in research. AI didn’t create it. It just stripped away the busywork that disguised who actually had it.

The analysts who survive (and thrive) won’t be the ones who type fastest. They’ll be the ones whose taste in what to investigate was always the load-bearing skill, even when 14 hours of grunt work was hiding it.

If you work in markets, the original piece is worth your ten minutes. It’s one of the better pieces of primary data on AI-in-finance I’ve read, mostly because it isn’t theorizing, it’s looking at what people actually do when you hand them an API.