Healthcare AI data · · 7 min read
Healthcare AI data vendors compared
Acumen Editorial · Updated November 21, 2026
Clinical AI changed shape this month. On November 16, 2026, OpenAI added computer-use agents to ChatGPT for Healthcare, software that acts across existing hospital systems with access controls and audit logs. Google Cloud opened Gemini for healthcare in preview on November 10, an agentic platform for providers and payers with launch partners that include Epic, HCA Healthcare and Northwell Health, according to Fierce Healthcare.
Agents that file referrals and pull records raise the bar on the data behind them. They need to be trained and tested on work that looks like real care, graded by people who do that work. That has made the vendor decision harder, because the market for expert data is large, loud and growing fast.
Who sells data for healthcare AI?
The money is real. Mercor raised a $750 million Series D at a $20 billion valuation on November 10, 2026, and says its network spans more than 5 million experts, per TechJack Solutions' summary of the round. Handshake's AI training business reached nearly $1 billion in gross annualized revenue by April 2026, with contracted workers that include medical doctors, according to Dealroom. TechCrunch reported in August that Micro1 hit a $500 million gross run rate. Frontier labs work with several of these providers at once.
Not every vendor sells the same thing. We see five types.
| Vendor type | What you buy | Strongest at | Watch for | Best fit |
|---|---|---|---|---|
| Horizontal expert marketplace | Access to vetted contractors across every domain | Speed and scale for broad programs | Specialty depth, clinical QA design, health documentation | Frontier labs running many domains at once |
| Managed labeling firm | Annotation workforce plus tooling | High-volume, well-defined tasks | Clinical judgment on ambiguous cases | Imaging labels, coding, extraction |
| Clinical data licensor | De-identified records, notes, claims or images | Real-world patterns at volume | No reference answers or graded reasoning | Pretraining and fine-tuning on real data |
| Synthetic data provider | Generated patients, notes or conversations | Privacy, rare cases, fast iteration | Validation against real care, drift | Augmentation and stress tests |
| Vertical expert data provider | Specialty-matched clinicians, rubrics, adjudicated sets | Evaluation, post-training, regulated documentation | Narrower scope outside health | Clinical deployment and regulated buyers |
Most mature programs use more than one. The mistake is buying one type for a job that belongs to another.
Which vendor type fits which job?
Start from the job, not the vendor list.
Pretraining and fine-tuning on real patterns. Licensed records are the right tool. They show how care is actually documented. They do not tell a model what the right answer was.
Reasoning, preference and reward data. This is where expert data earns its price. Physicians write cases with reference answers, build rubrics and compare model outputs. OpenAI's HealthBench used rubrics written by 262 physicians, and its November ChatGPT for Healthcare update cited a blind evaluation of more than 1,300 physician reviews across multiple specialties. That is the standard buyers now hear in every pitch.
Evaluation before deployment. Here specialty match and adjudication matter most. An agent that drafts oncology referrals should be graded by oncologists, with disagreements resolved and agreement reported.
Evidence for regulators and health-system review. FDA's January 2025 draft guidance on AI-enabled devices asks sponsors to describe how data were collected, cleaned and processed, and how test data stay independent. In August 2026, FDA's device center floated a competency-based evaluation for generative AI devices: benchmarking against prespecified thresholds, then clinical confirmation, with independent clinician adjudicators proposed for postmarket review, according to Cooley. Whoever builds your test data is now part of your evidence.
How do you compare healthcare AI data vendors?
Score every shortlisted vendor on the same eight criteria. We weight the first four highest for clinical work.
- Specialty match. Can they staff the exact specialties and care settings in your intended use, not "physicians" in general?
- Credential verification. Do they verify license, board status and years in practice, and can they show it per contributor?
- Adjudication. How many clinicians review each item, how are disagreements resolved, and do they report inter-rater agreement?
- Hard-case coverage. Do sets include omissions, rare presentations, demographic variation and cases where the right answer is to escalate?
- Provenance and documentation. Can they deliver a data description that fits an FDA submission or a health-system security review?
- Privacy handling. How is PHI kept out, and how is de-identification checked?
- Test-set independence. Can they guarantee evaluation items never enter your training data?
- Turnaround and pricing model. How fast to first sample, and are you paying per hour, per item or per accepted item?
Ask for a sample dataset against your own use case before you sign. A vendor's sample tells you more than its deck.
Marketplace or vertical provider: which should you pick?
Pick a horizontal marketplace when you need many domains at once and your team will design the clinical tasks, rubrics and QA itself. The large marketplaces are built for that, and their scale is real.
Pick a vertical provider when the hard part is clinical judgment: choosing the right specialists, designing rubrics that catch omissions, adjudicating disagreement and writing documentation regulators and procurement accept. The vendor's value is in the design, not only the headcount.
Many teams do both. They use a marketplace for breadth and a vertical provider for the evaluation sets that decide whether a model ships.
What should you ask a healthcare AI data vendor?
1. Which specialties have you staffed in the last six months, and how many reviewers per specialty?
2. Show us a rubric you built and the agreement rate on it.
3. How do you handle cases where reviewers disagree?
4. Who owns the data, and can we audit contributor credentials?
5. What does your data description look like for a regulatory file?
6. How do you keep evaluation items out of training sets?
7. Can we get a sample dataset on our use case within two weeks?
If the answers are vague, the data will be too.
How Acumen helps
We are a vertical expert data provider for healthcare, life sciences and pharma. Our clinical AI training data service supplies physician-written cases with reference answers, rubrics, graded reasoning and preference data, matched by specialty and adjudicated by clinicians per item. Every delivery ships with provenance documentation built for procurement and FDA-style data descriptions.
Use us alongside a marketplace, or for the evaluation sets that decide whether your agent goes live. For the sourcing steps in more detail, see our guide to sourcing expert-verified healthcare AI training data.
FAQ
What are healthcare AI data vendors?
Healthcare AI data vendors supply the data used to train and evaluate clinical AI. They include horizontal expert marketplaces, managed labeling firms, clinical data licensors that sell de-identified records, synthetic data providers, and vertical expert data providers that supply specialty-matched clinicians, rubrics and adjudicated evaluation sets. Each type fits a different job.
What is the difference between an expert marketplace and a vertical provider?
An expert marketplace gives you access to vetted contractors across many domains, and your team usually designs the tasks and quality checks. A vertical provider focuses on one industry and designs the clinical work itself: specialty matching, rubrics, adjudication and documentation. Marketplaces win on breadth. Vertical providers win on clinical depth.
Is licensed EHR data enough to train clinical AI?
No, not on its own. Licensed records show how care is documented at scale, which helps pretraining and fine-tuning. They do not include reference answers, graded reasoning or judgments about which response is safer. Models used in care still need clinician-written cases, rubrics and evaluation sets to learn and prove good decisions.
How do you evaluate a healthcare AI data vendor?
Score each vendor on specialty match, credential verification, adjudication and agreement reporting, hard-case coverage, provenance documentation, privacy handling, test-set independence, and turnaround. Then ask for a sample dataset built on your own use case. A sample shows rubric quality and reviewer depth far better than a sales deck or a contributor count.
Why does FDA care who builds AI training and test data?
FDA's January 2025 draft guidance asks device sponsors to describe how data were collected, cleaned and processed, and to keep test data independent. Its August 2026 discussion paper on generative AI devices proposes benchmarking against prespecified thresholds plus clinical confirmation. The people and process behind your data become part of the evidence.
Can you use more than one healthcare AI data vendor?
Yes, and most mature teams do. Frontier labs work with several providers at once. A common pattern is a marketplace or labeling firm for broad, high-volume tasks, licensed records for real-world patterns, and a vertical provider for the specialty evaluation sets that decide whether a model or agent is ready for clinical use.