The physician panels

The competitor most likely to take this market is not another data company. It is a physician survey panel that has not yet noticed it is in the market.

Sermo has 1.3M+ healthcare professionals across 150 countries, over a million of them physicians, 600+ staff, and offices in New York, Puebla, London, Barcelona, Vilnius, Tokyo, Dalian and Cape Town (Sermo). The number that decides everything: Sermo paid $20 million to members in one year, with members able to earn over $15,000 (White Coat Investor).

Read that as infrastructure rather than as a marketing claim. Sermo already has verified identity, engaged attention, tax and compliance handling, and funded payment rails to a million physicians. It moves eight figures a year through them today. If Sermo decided tomorrow to sell expert judgement to AI labs, it would not need to build supply — it would need to learn what to ask.

M3 Global Research claims over 2 million verified physicians and HCPs (M3), the largest panel in the world, with a Tokyo-listed parent (2413). Panel-level revenue is consolidated and not broken out, so it could not be established. Doximity counts roughly 80% of US physicians as members and monetises through pharma marketing and hiring, never by the hour; FY2026 revenue of $644.9M appears only at headline level [WEAK] (StockTitan).

And Hippocratic AI, at $3.5B, has already built every component internally: clinician recruitment at 7,700 licensed clinicians, credential verification, a standing network of 6,000+ nurses and 300 physicians, a Nurse Advisory Council and a clinical-safety taxonomy (hippocraticai.com/safety, MobiHealthNews). It does not sell any of it. It has the recruiting pipeline, the payment rails and the safety taxonomy that a clinical-judgement business needs, and it uses them entirely for its own product — see Health-AI companies as buyers.

#The wage floor that governs everything

Whatever a new entrant pays, it is a price-taker. The physician market-research industry set the rate years ago and pharma keeps paying it.

Two readings fall out of that ladder.

AI training work has priced itself into the physician market-research band and not above it. $110–250/hr sits squarely inside $60–300/hr, which is exactly where it must sit to attract a practising physician's marginal hour, because that is what pharma already pays for the same hour. There is roughly a 10–25× spread between the crowd-labelling floor and the expert-witness ceiling for what is nominally the same act — a doctor looking at a case and saying what they think. Everything in the middle is a negotiation about credentialing and task difficulty, not about scarcity.

A new entrant cannot pay meaningfully less than $100/hour for practising-physician judgement and expect supply. It can pay far less for trainees, students and non-US clinicians — which is precisely the arbitrage the crowd model exploits, and precisely why that model produces labels rather than judgement. So the margin cannot come from buying the hour cheaply. It has to come from task design: what you ask in the hour, and how much of the output is sellable. This is the same conclusion GMV is not revenue reaches from the accounting side, and the one What actually gets sold builds the product on.

#The recruiting funnel benchmark

The one clean public number for what it costs to convert physicians into raters comes from a 2026 study. Researchers emailed 12,000 physicians drawn from IQVIA's OneKey provider database and converted 1.2% — 149 completers across 36 states, who produced 1,156 pairwise ratings over 620 questions at a median 127 seconds per rating, 7.8 assignments each. The physicians "were compensated"; the amount is not disclosed (arXiv).

1.2% from a clean, licensed, commercially maintained provider database. That is the number to plan a supply model against — not the response rates implied by a marketplace's signup page. It also puts OpenAI's HealthBench funnel in perspective: 1,021 inbound interest forms converted to 262 working physicians, a 26% yield on people who had already raised their hands. The expensive step is getting the hand raised at all.

#Why the panels have not moved

They sell surveys, and surveys are a mature, high-margin, low-variance product with a known buyer. AI data is a lumpy, technically demanding, unfamiliar sale to a buyer who churns fast and negotiates hard. The panels are not blocked; they are uninterested — for now.

Two things could change that. A frontier lab could approach one directly, at which point the panel learns the price. Or a startup could visibly demonstrate the margin, which teaches every panel the price at once. The second is the standard fate of intermediaries who prove a market for a supplier that already owns the supply — the Getting cut out problem in its sharpest form.