Deep dossier · 31 pages · 52,000 words · 30 August 2026

Health, examined properly

The niche screen liked clinical medicine's evidence and killed it on cost — physicians are the most expensive supply in the shortlist. A proper look finds that the cost objection applies to one specialty out of thirteen, that the regulator has written the product specification, and that the hardest problem is not law or price but whether the labs will buy from anyone at all.

Read the verdict first

01

The read

What the evidence says, and the one shape of this business the regulators are effectively demanding.

The health read

Build, in four of thirteen specialties, and not the ones the screen liked. The cost objection that killed clinical medicine on the shallow screen turns out to apply to exactly one specialty out of thirteen. The finding that should worry you is that the largest buyer ran its own physician recruitment funnel, published it in full, and named no vendor anywhere in it.

Characterised disagreement

Models now beat physicians on HealthBench Professional, 59.0 against 43.7. That is good news, because it moves the product from producing the right answer to producing a measured map of where the right answer is contested — which is the one artefact that appreciates as models improve, and the one FDA asks for by name.

02

Thirteen specialties

Healthcare is not one market. Each specialty has its own buyer, its own wage floor and its own legal position.

Clinical reasoning

The only healthcare sub-niche where frontier labs have repeatedly, publicly and by their own admission paid physicians — and consequently the one with four incumbents and no moat.

Dermatology and ophthalmology

The best risk-adjusted specialist bet: the only specialty with a reimbursed autonomous AI service, the most permissive image corpora, a near-census membership list, and a reasoning-quality gap that is wide open while answer accuracy is nearly saturated.

The best intellectual thesis in the dossier and the worst substrate: ground truth genuinely is a distribution over expert opinion, which makes a blinded adjudicated panel the only honest product — but only about 10% of US labs are digitised.

The biggest headline market in medical AI and the worst unit economics in this dossier: 88.2% of image-based FDA authorisations, a real capability gap, and an arbitrage that runs backwards.

Genomics and variant curation

The cleanest artefact and the cleanest legal position in the dossier, a shortage quantified to the individual variant, a frontier failing below 50% — and no buyer at all.

The largest and most internationally reachable clinician pool of the six, a severe documented workforce shortage — and every road to the valuable artefact runs through a closed door.

The best labour arbitrage in the dossier and probably the best niche of all thirteen: a 3x wage spread, the only pool with first-party date-stamped credential counts, no licensure, no malpractice, no PHI, and the best model on earth scoring 63.57%.

Three and a half million verifiable RNs, no PHI requirement, a wage multiple that works — and the emptiest benchmark space in the dossier. Held back by a hostile union and by the largest buyer having already built its own supply.

Behavioural and mental health

The only niche where frontier labs buy directly and say so in writing, and the highest willingness-to-pay multiple in the dossier — occupied by a funded vertical specialist with a psychologist CEO, a published benchmark and foundation-model customers.

Veterinary medicine

No HIPAA is a structural cost advantage nothing else in healthcare matches, and no expert-data vendor has a veterinary vertical — but three expert-curated benchmarks landed in 2025–26 and the sharpest play inside the niche is not about pets at all.

Clean artefacts, free verification, no PHI and real benchmark headroom — undone by a 1.1–1.8x wage multiple and the total absence of any named commercial buyer anywhere in the record.

Ambient scribing

The loudest, best-funded buyers in health AI and the only niche with two published studies showing external clinician review being purchased — attached to a benchmark saturating at 91% and raters priced at physician rates.

Clinical trials and regulatory writing

Where the dossier's own buyer thesis breaks: every disclosed pharma-AI deal from 2020 to 2026 is compute, cloud or platform, none is expert labelling — and freelance regulatory writers already earn $131.65/hr, so there is no arbitrage to capture.

03

Who would pay

Four labs that matter, a field of funded application companies, and pharma — which turns out not to buy this.

The labs as health buyers

OpenAI published its physician recruitment funnel in full and no vendor appears anywhere in it — 1,021 applicants down to 262 paid physicians, run first-party. The lab TAM is four, realistically two and a half, and the whole programme has plausibly cost $12M.

Health-AI companies as buyers

Probably the real first customer, and the evidence is discouraging: the only unambiguous large-scale buyer hired 7,700 clinicians itself, and across eight readable job boards there is not one US clinical-reviewer posting.

Pharma, payers and providers

Every disclosed pharma-AI deal from 2024 to 2026 bought model access, deployment or trial infrastructure — not one bought expert-labelled data. But $11–12B of clinician judgement is already priced next door, in pharmacovigilance and medical affairs.

04

Who would sell

What a clinician hour costs, how many there are, and the registries that verify them.

What a clinician hour costs

The wage ladder runs from a $22/hr coder to a $475/hr expert witness for what is nominally the same act — a clinician looking at a case and saying what they think. Healthcare's structural advantage over every other domain in this atlas is that the state maintains the register, and the recruiting funnel benchmark is 1.2%.

05

Who is already there

One incumbent a layer below, one direct competitor nobody has noticed, and a million-physician panel that could enter tomorrow.

Centaur.ai, in full

The clearest incumbent in medical annotation sits a layer below the business under consideration — volume labels from a semi-credentialed gamified crowd at an implied $10–25/hr, against Mercor's $110–250/hr for physician judgement.

BioStack and Sepal

You are not first. A seven-person YC company founded in October 2025 is already selling clinical RL environments and evaluations to top AI labs under a six-figure contract — and it barely appears in any market map.

The physician panels

The pivot threat is Sermo, not Centaur. It has 1.3M+ HCPs and paid $20M to members in one year — funded payment rails to a million physicians already in place — and it sets the wage floor that governs the whole dossier.

The real-world data brokers

A $1B-revenue records industry sitting next door — and not one of them is moving into expert labelling or third-party evaluation. Records and judgement remain separate goods with separate buyers.

06

What you would sell

The artefact, the benchmarks, and why characterised disagreement outlives the expert being right.

What actually gets sold

Mercor spent over $500,000 on 200 benchmark tasks — about $2,500 each, the best cost-per-unit anchor in public. Everything else in this market is priced behind an NDA, because no vendor publishes a rate card.

The measurement landscape

Models now beat physicians on HealthBench Professional, 59.0 against 43.7, and by four times on SDBench — while humans still beat GPT-4V on pathology. When the model exceeds the expert, the buyer stops wanting ground truth and starts wanting harder cases.

07

What could stop it

PHI, the FDA, Europe, and the clinician's own exposure — several of which turn out to be the reason the business exists.

Build it without ever touching a patient record

MIMIC is closed to a commercial vendor on four independent grounds, and 44% of the ISIC archive is silently non-commercial — but the de-novo case route clears HIPAA, the Common Rule and institutional data ownership at once, and a JAMA trial in 2000 found written vignettes scored 71.0% against chart abstraction's 65.6% as a measure of real practice.

The regulator wrote your product spec

FDA's January 2025 draft guidance asks device sponsors for annotator expertise, blinding, adjudication method and an inter-clinician variability assessment — by name, as submission content. One of those four cannot be reconstructed after the fact, and that is the whole moat.

Europe prefers the design you were going to build

GDPR Article 89(1) says purposes that can be met without identifying data subjects 'shall be fulfilled in that manner' — and AI Act Article 10(5), preserved and extended by Regulation (EU) 2026/1744, permits real special-category data only where the job 'cannot be effectively fulfilled by processing other data, including synthetic or anonymised data'.

What the doctor on the other end is risking

The largest practical risk in this business is not patient privacy but the hospital employment contract — and two questions that determine whether clinicians say yes have no answer anywhere in the record: whether this is the practice of medicine, and what organised medicine thinks of doctors selling judgement for AI training.

08

What to do next

The first ninety days, and what is still unknown.

Ninety days in health

Ask the one question nobody has asked, seat twenty coders against a 1.2% funnel, and build a single 500-chart blinded panel — because the inter-rater statistic is the only deliverable that cannot be added afterwards. Every step below is ordered so that a no arrives early and cheap.

What the health dossier could not establish

Ranked by how much the answer moves the decision. The top item is that nothing in the entire record shows any lab, device sponsor or health-AI company buying clinical expert data from a vendor — not one contract, not one named supplier, not one rate.