Atropos Health

Three delivery speeds sit on one engine, and which one you want changes the conversation materially.

The Green Button informatics-consult service is the original: you pose a clinical question, an Atropos analyst-mediated process returns a bespoke observational study (a "Prognostogram") in under 48 hours. ChatRWD is the generative-AI front end that compresses the same output to minutes — Atropos describes it as "the first generative AI application to deliver full observational studies on healthcare data in minutes." The Atropos Evidence Agent is the newest layer: the same capability surfaced inside a clinician's existing workflow rather than in a separate tool.

The underlying platform is called GENEVA OS. The point worth internalising is that this is a study-generation system, not a retrieval system. When it works, it answers questions that have no published literature behind them — which is precisely the case the vendor's own study found it strongest in, and precisely where a retrieval tool like OpenEvidence cannot help you.

This is the part worth reading closely, in both directions.

In June 2025 a study appeared in SAGE's DIGITAL HEALTH (DOI 10.1177/20552076251348850, PMID 40510193) submitting 50 clinical questions to five systems: ChatRWD, OpenEvidence, ChatGPT-4, Claude 3 Opus and Gemini 1.5 Pro. Nine independent physicians graded the answers. Verbatim from the abstract: general-purpose LLMs "rarely produced relevant, evidence-based answers (2–10% of questions)," against 24% for OpenEvidence and 58% for ChatRWD. On actionability the picture splits usefully: OpenEvidence produced actionable results for 48% of questions that had existing evidence versus 37% for ChatRWD — but for questions that lacked existing literature, ChatRWD reached 52% against under 10% for everything else. The paper's own conclusion is that the two approaches are complementary, not that one wins.

That design — blind physician graders, a named competitor, published numbers that are not flattering in every column — is genuinely more than most vendors in this space submit to, and it should count in Atropos's favour.

The qualifier is stated in the paper itself, and we quote it rather than characterise it: "ChatRWD, the LLM system evaluated in this study, is developed by Atropos Health where many of the authors are employed. NHS is not an Atropos Health employee but sits on its board." A companion medRxiv preprint introducing the "Answered with Evidence" evaluation framework carries the same structure — "JB, CD, AM, NS, and SG are employees of Atropos Health, which provided funding for this study." So: peer-reviewed, independently graded, vendor-funded and vendor-authored. We found no evaluation of Atropos conducted by a party with no commercial relationship to it.

For category context rather than a measurement of this product: an independent March 2026 benchmark, RWE-bench (arXiv 2603.22767), had LLM agents reproduce 162 peer-reviewed observational studies against MIMIC-IV and found the best agent completed 39.9% of tasks end-to-end. Atropos is not among the systems evaluated there, so this says nothing about ChatRWD's accuracy — it is simply the best available independent read on how hard the underlying problem is, and a reason to run your own pilot questions rather than accept any vendor's numbers.

In April 2026 Atropos announced Alexandria, which it calls "the largest source of novel medical content, featuring more than 33 million artifacts of evidence generated through a high-throughput, standardized observational evidence-creation workflow," and stated it is "expected to scale to two billion studies by the end of 2026, exceeding all known medical evidence by 100x in volume."

Three things a buyer should hold in mind. First, these are machine-generated observational analyses, not peer-reviewed literature — the 100x comparison is a volume comparison, not a quality one, and Atropos does not claim otherwise. Second, the figures vary by source: the Atropos Evidence Agent listing on Microsoft Marketplace describes searches across "over 10M novel studies in Alexandria" while the product page and April release say 33M, so establish which number applies to the corpus you would actually be querying. Third, the two-billion figure is a projection, not a delivered state; as of this writing we found no source confirming it has been reached.

Atropos did address the obvious question of quality-at-scale. A separate April 2026 announcement describes a multi-step review combining the "Answered with Evidence" evaluation, an AI review, on-demand expert clinical review by Atropos's own teams, and "an independent audit process modeled on peer review by international publisher Becaris Publishing Limited." The Becaris involvement is the only externally-run component; how large a sample it audits is not established publicly, and it is a reasonable thing to ask for in diligence.

The Atropos Evidence Network is described by the vendor as covering "300M+ de-identified records" across 40+ clinical specialties, with data spanning open and closed claims, EMR, labs, vitals, registry data and RxNorm/CPT coding, sold in packages by therapeutic area rather than as one undifferentiated data purchase.

On deployment, the most concrete public data point is Stanford Health Care, where — per the announcement text carried by HIT Consultant — Atropos is "already installed inside Stanford Health Care's firewall" and generates evidence to help clinicians finalise encounter notes, including notes produced by Microsoft's ambient AI. That is a meaningful architectural signal for any health system whose security review will object to PHI leaving the building. We would not generalise it to every deployment without asking: whether that arrangement is the standard offering or a Stanford-specific build is not established from public sources.

On compliance, we could not find a public trust or security page. Atropos's published privacy policy governs the website — it states the site does not collect protected health information "unless you or your affiliated institution has executed an agreement with us specifically providing for our collection of such information" — and does not describe platform-level PHI handling. We found no public evidence of SOC 2 or HITRUST attestation, and no public statement of the de-identification standard used (HIPAA Safe Harbor versus Expert Determination). None of that means the controls are absent; enterprise health-tech vendors routinely keep this behind an NDA. It does mean you cannot pre-screen it, and it should be an early question rather than a late one.

Atropos has moved decisively toward being consumed by other systems rather than through its own UI, which is the most relevant recent development if you are assembling an agent stack.

In June 2026 it launched the Atropos Evidence Agent MCP on Databricks Marketplace, extending a Databricks partnership first announced in June 2025 — the pitch is asking clinical questions in natural language and getting evidence-backed answers without leaving your Databricks environment. In March 2026 it launched a collaboration with Microsoft Dragon Copilot at Stanford Medicine, and the Atropos Evidence Agent is listed on Microsoft Marketplace as a Teams app (offer WA200009979). The marketplace listing describes the agent as interpreting question context, orchestrating LLM searches across published literature and Alexandria, and returning a result "badged with an Answered with Evidence rating to help requestors understand the quality of answer for that specific question." August 2026 added a partnership with Health Universe to surface the same capability inside its governed agent environment.

The per-answer quality rating is a real design choice worth crediting: a system that tells you how well-supported each individual answer is, rather than presenting all outputs with uniform confidence, is doing something most clinical AI tools do not.

The practical caveat is that none of these are self-serve. The Databricks route requires discovering the MCP, then completing onboarding with Atropos to activate access before configuring it in your own workspace. Treat marketplace availability as a shorter integration path, not as a way to skip the sales conversation.

There is no pricing page anywhere on atroposhealth.com — we checked the full page sitemap — and no self-serve tier. Every entry point is a demo request or a contact form; the older ChatRWD access page is still a beta-access signup. Network access is packaged by therapeutic area (the site lists cardio-metabolic, gastro/renal/urology, immunology/derm, neuro/psych and rheumatology among others), which suggests scope-based pricing rather than seats, but no figures are published and we are not going to guess at them.

The named buyers in public material are academic medical centres, health systems, and life-sciences and pharma teams — the Series B explicitly earmarked funds for pharma partnerships, and the investor list (Cencora Ventures, McKesson Ventures, Merck Global Health Innovation Fund) reflects that. Atropos also announced a Merck collaboration and, in January 2026, a commercial analytics product aimed at life-sciences companies. If you are an individual clinician or a small practice, this is not currently sold to you.

Atropos spun out of Stanford Medicine, where the underlying "Green Button" informatics-consult service ran inside Stanford Health Care for roughly two years before the company launched with a seed round in autumn 2020, backed by investors including the Boston Millennia Founders Fund. Co-founders include Brigham Hyde (CEO) and Saurabh Gombar (CMO), who retains a Stanford Pathology affiliation, alongside Stanford's Nigam Shah. Note that secondary sources disagree on the founding year — some say 2019 — and we could not resolve it from a primary source, so we are describing the seed round rather than asserting a date.

The last publicly announced raise is a $33M Series B in May 2024, led by Valtruis with Cencora Ventures, McKesson Ventures and Merck Global Health Innovation Fund joining existing investors Breyer Capital, Emerson Collective and Presidio Ventures. We found no publicly announced round since, which at a bit over two years is worth noting neutrally: it is unremarkable for a company with revenue and strategic investors, and it is also the sort of thing to ask about directly if you are signing a multi-year deal. Total-funding figures from aggregators disagree ($50.3M and $53.8M both appear) and we would not repeat either as fact. Headcount is not established — the figure Tracxn serves is not credible and we did not find a reliable source.

Signs of ongoing investment rather than drift: the company hired its first CTO and a CFO in February 2026, and Fierce Healthcare named it to its "Fierce 15" list of private healthcare companies in March 2026.

One independently verifiable strength: searching PubMed today for the affiliation "Atropos Health" returns 41 indexed papers. The homepage's claim of "over 50 publications" presumably includes non-PubMed venues, and we have not verified the fuller figure — but a genuine, checkable research output at that scale is uncommon for a company this size and supports the methodological posture the marketing claims.

Stated as trade-offs, because that is what they are.

If your questions already have good published answers, a retrieval tool may serve you better and cost less — the vendor's own study has OpenEvidence ahead on actionability for questions with existing evidence (48% vs 37%). Atropos's edge is specifically in the gap where literature is thin.

If you want to evaluate before you talk to sales, you cannot. No trial, no pricing, no public sandbox; even the MCP requires onboarding.

If you need independent validation for a governance committee, it does not exist yet. The best evidence is a well-designed vendor-funded study. Committees that require third-party evaluation will need to run their own.

If observational evidence is not decision-grade for your use case, no amount of speed fixes that. Confounding and unmeasured-variable bias are properties of non-randomised data, not of the tool, and generating a study in minutes does not resolve them. Atropos's answer is its review layers and the per-answer quality rating — reasonable, and something you should stress-test on your own questions during a pilot.

If you are a small team or an individual clinician, this is not sold to you today.