Miju LabsFinance has the best demand evidence in this research and the worst competitive position, and both facts trace to the same cause: it is the first domain anyone thinks of.
OpenAI has a dedicated Subject Matter Expert, Investment Banking role in San Francisco, hybrid three days, at $185,000–$205,000 plus equity, asking for 2+ years of IB experience including live transaction execution. The duties read as a product specification: "designing evaluation tasks that reflect realistic banking workflows", "creating reference materials like financial models, valuation analyses, and pitch books that meet professional banking standards", "developing assessment rubrics for financial correctness, analytical judgment, and practical usefulness" (OpenAI).
A search of OpenAI's full 754-role board for "Subject Matter Expert" returns exactly one result, and this is it (OpenAI careers search). Read that twice, in both directions. It proves finance is the one domain OpenAI staffs internally. It also proves OpenAI is willing to insource rather than buy — which is the single most important sentence on this page.
What the data actually is
Reconstructed deal work, and the industry has already agreed on how.
Live deal data is material non-public information and cannot leave the bank. Full stop, no workaround, no consent flow. But an ex-banker building an LBO from a public 10-K is clean, and the whole market operates on that basis — see the line Mercor puts on its own engineering listings: "No access to confidential employer information required" (Mercor). The blocker that looks fatal from outside is soft from inside.
The artefacts:
| Line | Input | Why it is scarce |
|---|---|---|
| Model construction traces | Public filings, built forward into a DCF or LBO | The order of the build and which assumption gets stress-tested first |
| Precedents and comps judgement | Public transaction data | Which comparable is defensible and which one a client will laugh at |
| Accretion/dilution reasoning | Two public companies, a hypothetical structure | The structure choice, not the arithmetic |
| Pitch-book critique | Constructed decks | What a managing director strikes through, and why |
The benchmark is open at exactly the sub-skill you would sell. Vals AI's Finance Agent Benchmark v2 tests nine analytical categories including "DCF/NPV, LBO, and M&A accretion/dilution models" and credits the financial experts who built it. The best model manages 60.60% on partial credit overall — but the two hardest categories are Financial Modeling at 34.52% and Precedents at 36.37% (Vals AI).
Proof scores 3, not higher. The number to beat is public and low, which is good. But Vals AI owns the benchmark layer here, publishes into model cards, and is not going to vacate it — so the standard entry weapon, publishing the benchmark nobody has built, is unavailable. That is the same problem Offensive security has, arriving from a different direction.
Is anyone buying
Yes, harder than anywhere else in this set of eight. Budget scores 5.
- OpenAI's SME role above — a named, permanent, in-house headcount with a six-figure band.
- xAI ran a family of contracts — Investment Banking Expert (M&A), (ECM), (DCM), Finance Expert (Quant) — at $45–$100/hr, remote or Palo Alto, US-only, tasked with "data annotations and labels for sell-side finance AI training using proprietary software" (xAI posting). Press put the headline at "$100 an hour to teach Grok finance"
[WEAK — secondary](Entrepreneur). - When xAI enumerated the four domains it was surging specialist tutors into, finance was one of them.
- Anthropic's product motion corroborates the spend — Claude for Financial Services and a bank-and-insurer agent suite
[WEAK — vendor marketing, not a data contract](Anthropic).
No system card or disclosed contract in this domain names purchased finance expert data as a line item. What exists is one permanent lab headcount, one family of hourly contracts, and one benchmark vendor's methodology page. That is the strongest demand evidence in this set of eight — and it is still a rate card, not an invoice. Every "the lab is buying" claim across these pages rests on the same class of evidence.
Getting the experts
Reach scores 4. There are 443,100 US financial analysts at a median $103,570/yr, $49.79/hr (BLS Occupational Outlook Handbook; no URL was captured in the research, so treat the citation as second-hand), and the channels are good but unverified: Wall Street Oasis and Mergers & Inquisitions, where the population self-selects by career obsession; CFA Institute; banking alumni networks; university IB clubs.
Why 4 and not 5: there is no register. The CFA charter is a credential but most bankers do not hold it, and "two years at a bulge bracket" is verifiable only by reference. Compare Accounting, audit and tax, where NASBA hands you 650,667 licensed CPAs at one URL, or Architecture, engineering and construction, where the state PE registers are public and searchable. Here you are back to LinkedIn and trust — the same vetting problem Sales and GTM has, in a population with more incentive to inflate.
What it costs to run
Cost scores 5, the highest in this set alongside nothing else. No rig, no licence, no building. Excel and public filings — EDGAR is free, the 10-K is free, transaction data has cheap tiers.
Against a $49.79/hr floor and an observed $45–100/hr AI-data rate, the arbitrage is real but not extraordinary: roughly 1x to 2x. The bind is subtler and it is worth naming, because it is the same trap Sales and GTM falls into. A second-year analyst will take $100/hr happily. A vice-president whose deal judgement is what a lab actually wants has an opportunity cost measured against a bonus, not a base, and will not. You are structurally recruiting from the junior end of a profession whose value is concentrated at the senior end.
The workaround is per-artefact pricing — a flat fee for one hard structuring opinion that takes twenty minutes — the same move Clinical medicine needs for procedural specialists and Offensive security gets for free.
Who is already there
Room scores 2, and this is the number that reclassifies the page from build to crowded.
Halluminate occupies this niche by name. It builds "RL environments that train frontier AI models on real financial services work", explicitly targeting investment banking, private equity and hedge funds, with a Data Partnerships page implying it licenses real workflows from firms (Halluminate). Its Strategic Project Lead req — $150–225K base, 0.15–0.30% equity, SF in person — wants "finance analysts or associates with deep interest or expertise in AI/ML" to create "modeling exercises, pitch decks", recruit contractor staff, and build "end-to-end systems and processes for data and environment delivery" (Ashby). That is this business, already staffed, already YC-funded, already in San Francisco.
One qualifier worth holding onto: the RL directory lists Halluminate's disclosed funding at only $160K [WEAK — directory figure, almost certainly stale for a YC S25 company with a closed seed and $225K salaries] (rl-list). Whether this niche is occupied or barely touched turns on a number nobody has.
Also present: Vals AI on the benchmark layer, raising $40M at a $400M valuation led by a16z, with results appearing in model cards from five labs. AfterQuery builds finance environments [WEAK — directory] (alignlist). And Mercor's APEX benchmark used Goldman Sachs and JPMorgan graders.
Defense scores 2. No licence gates entry, 443,100 people are reachable by anyone, the artefacts are constructible by any competent ex-banker, and the buyer has already demonstrated it will hire the expert directly rather than buy the data. Nothing here stays scarce.
What would kill it
OpenAI already did the killing move. A lab that hires its own subject-matter expert at $185–205K has decided the judgement is worth owning rather than procuring. If that pattern spreads — and finance is where it started — the addressable spend is capped at the labs that have not yet hired.
Halluminate raises properly. A seed-stage YC company with the position, the pitch and the SF network converts into a real incumbent on one round.
Vals extends downward. The benchmark-to-data path runs in both directions, and Vals is standing on the step you would enter through, with $40M and five labs' citations.
MNPI, if you ever get sloppy. Reconstructed work is clean. One expert pasting a live deal model into your platform is a regulatory event, not a data-quality event.
The first ninety days here
The honest answer is that this niche fails the test the rest of this atlas is built on: The specialist wedge argues you want an open benchmark or a proven budget, and prefers the first. Finance has a proven budget with the benchmark layer captured and the vertical occupied. That is the losing side of the trade.
If you enter anyway, enter where Halluminate is not: not banking, but the un-glamorous adjacent judgement — credit, restructuring, FP&A, insurance underwriting — which shares the modelling substrate, has no named competitor, and whose practitioners are cheaper and less mobile than bankers. Recruit 40 through Wall Street Oasis and alumni networks, pay per artefact, and publish a modelling-error taxonomy rather than a leaderboard. FAB v2 already tells the buyer the score is 34.52%; nobody has told them which modelling errors produce it. See The first ninety days.
Where the record is thin
Halluminate's revenue, customers and true funding are unknown, and the $160K directory figure is almost certainly stale. Whether this domain is occupied or open turns entirely on that gap.
The xAI band comes from a job aggregator; the OpenAI SME salary is a posting, not a contract. Nobody publishes what a finance environment or eval set sells for — the only price point anywhere in this research is SemiAnalysis's roughly $20,000 per UI gym, which is a different product. And the BLS analyst figures arrive here without a captured URL.