CaryHealth Clair

The clearest centrality case among the reference vendors graded so far, and it comes from the opposite direction to the incumbents. The underlying sources are largely public: drug package inserts, a public clinical trials registry, a public drug interaction database, published guidelines. The company does not own a curated corpus built over decades and does not claim to. What it sells is the retrieval, routing and synthesis layer over that material, plus whatever the customer uploads.

Strip the model out and there is no product left, only public databases the buyer could already reach. That is the definition this axis is testing for, and it holds regardless of how the other axes grade.

It also explains the shape of the rest of this record. Where UpToDate, DynaMed and ClinicalKey AI grade low on centrality and high on editorial process because the content is the asset, this vendor inverts both.

Real credits first. Every response is stated to carry links to its original references, source routing is explained with worked examples, and the product is split into six narrow tools rather than one open ended assistant, which bounds what a given query is being asked to do.

What holds it at C is a claim in the company's own explanation of how the product works. It states that the system continuously learns from user interactions and feedback to refine its models, names self learning among its techniques, and describes a user community that rates results, giving the company the opportunity to continuously refine those results. Nothing published describes a clinical validation gate on that loop, a versioning scheme, or who approves a change before it reaches a clinician.

That matters more here than it would elsewhere. Every other vendor in this category has moved deliberately towards a fixed, curated, editorially governed corpus precisely so that answers do not drift. A reference tool whose outputs are shaped by aggregated user feedback is making the opposite bet, and the buyer is not told how it is controlled.

No abstention behaviour is described, and nothing states what the product does when sources conflict. Ask specifically what the self learning loop can change, who signs off, and how a clinician would know the answer to a repeated question had moved.

More technique is named than several better graded peers manage: retrieval augmented generation, embeddings, pipelines and self learning are listed explicitly. That is worth crediting, because most vendors in this category describe an outcome and leave the mechanism entirely unstated.

Two problems hold it at C. The descriptions do not reconcile. One answer says the models are trained on the company's database of clinical studies, package inserts, interactions and guidelines; another describes retrieval augmented generation over those sources; a third says the models are trained on diverse medical texts. Training on a corpus and retrieving from it are different architectures with different failure modes, and a buyer cannot tell which is being described.

The second problem is a number. The corpus is repeatedly given as more than 200 million clinical studies. That figure is not reconcilable with the size of the published biomedical literature as the rest of this category describes it: the two nearest competitors cite roughly 38 million and 60 million peer reviewed articles for corpora they present as comprehensive. The company does not define what it is counting as a clinical study, and the figure sits on the homepage, the product page and the frequently asked questions unqualified. Ask for the definition and the denominator before treating it as a comparison point.

No model class or provider is named, and no evaluation harness is described.

Nothing was located in two retrieval passes that measures whether the product is correct.

There is no peer reviewed publication, no accuracy evaluation, no benchmark result, no error rate, no clinician review panel, no named health system outcome and no third party assessment. The only quantified claim is that clinicians search around 90 percent faster than before, which appears on the vendor's own material and in a case study published by the design agency that built the mobile application, a source with a commercial interest in the result and no clinical measurement in it.

Recognition that does exist is not evidence of performance. The company reports an award from an industry awards programme and repeated placement on a fastest growing private company list. Those measure commercial growth and a submission process, not clinical accuracy.

This is the lowest evidence position in the category so far, and the gap is not close. The three incumbents publish either an evaluation with results, a structured customer satisfaction ranking, or a large content literature; two of the newer entrants publish at least vendor run benchmark figures. Here there is no number about correctness at all.

The base product carries no patient record data, since the named sources are public drug, trial and guideline databases, which is the favourable architecture this category generally enjoys.

Two features complicate that and neither is addressed in published material. Customers can upload their own documents and criteria for the model to answer against, so customer data of unknown sensitivity is ingested and retained by design. And the stewardship tool accepts optional patient demographics as an input, which is patient data even at a coarse grain.

Nothing published on the product material states an encryption practice, a retention period, a data residency position, a de identification approach, or whether uploaded customer documents are segregated between tenants. The company does publish a privacy policy, terms and a notice of privacy practices, none of which were opened this pass; that is recorded explicitly rather than left implicit, and it is where a fuller answer would sit.

The specific question to ask, because it follows from the product design rather than from a general worry: what happens to an uploaded document set, is it used only for that customer's answers, and is it ever used to refine the shared model given that self learning is named as a technique.

A dedicated notice of privacy practices page is published and linked from the site footer, which is more infrastructure than several peers in this category offer and is consistent with the parent company operating a licensed pharmacy and therefore holding covered entity obligations of its own.

Beyond its existence there is nothing a buyer can act on from the product material. No business associate agreement was located, no terms of the agreement, no statement of the tier at which one becomes available, no covered entity or business associate language on the product page, and no execution path. The enterprise offering is described as meeting the unique compliance and security needs of enterprise partners without naming a single requirement it meets.

Graded on what was found rather than as a finding that nothing exists. The notice of privacy practices, the privacy policy and the terms were all identified but not opened this pass, and any of the three could carry a business associate commitment. Flagged for refresh, and the notice of privacy practices is the first place to look.

No certification of any kind was located in two retrieval passes: no SOC 2 report of either type, no ISO 27001, no HITRUST, no trust centre, no penetration testing statement, no vulnerability disclosure programme and no named encryption practice on the product material.

What exists is an assertion without content. The enterprise architecture is described as meeting the unique compliance and security needs of enterprise partners, and customers are offered the ability to define their own guardrails. Neither names a control, a framework or an auditor.

A genuine credit belongs on the record even so: public application programming interface documentation is published at its own address, which is a real technical surface and an unusual thing for a vendor of this size to maintain openly.

Recorded as a retrieval outcome and not as a finding that controls are absent, with the legal pages noted above still unopened. But the comparison inside the category is unavoidable and cuts against the vendor: two startups of comparable size publish attestations with the type specified and run public trust centres.

No clearance, no submission and no published regulatory positioning statement were located.

The reference tools have a reasonable claim on the clinical decision support exclusion, since answers are linked to their sources and a clinician can review the basis. Two of the six tools sit closer to the line than that framing suggests and the record should say so plainly rather than treat the product as uniformly reference. The stewardship tool returns treatment, dose and duration for an infection and accepts patient demographics. The medication review tool is described by the vendor as providing evidence based recommendations for patient care.

Dosing guidance is the highest consequence output anywhere in this category, because an error carries directly into an order rather than into a clinician's background understanding. A product that generates it, adapts it to demographics, and separately describes its models as self learning is exactly the configuration a regulator would want an analysis of.

Graded C because no such analysis is published. Ask for the regulatory position on the stewardship and medication review tools specifically, not on the platform as a whole.

No governance artefact of any kind was located. There are no published artificial intelligence principles, no responsible use framework, no external advisory council, no named clinical leadership for the product, no membership of any external validation body, no bias evaluation, no subgroup analysis and no error taxonomy. What exists is a statement that a team of clinical experts vets data sources and monitors search results, with no names, no method and no output.

One published feature inverts the usual arrangement and is the reason this sits at D rather than C. The enterprise offering includes tailored guardrails, described as letting a customer define custom artificial intelligence governance frameworks tailored to that organisation's standards. Governance is therefore a configurable product option rather than a property the vendor guarantees, and the vendor supplies the mechanism while the customer supplies the standard.

Read that alongside two other published facts and it becomes a concrete exposure rather than an abstract one. Customers can upload their own documents for the model to answer from, and the branding can be white labelled. The industry list on the vendor's own navigation names pharmaceutical manufacturers first, ahead of health plans and healthcare professionals. Those three capabilities together describe a clinical reference tool that a manufacturer can deploy under its own brand, loaded with its own documents, under guardrails it defines.

State it as structure rather than as accusation, because that is what it is. Nothing here shows the configuration has been used that way. But the index has now recorded three distinct commercial arrangements in this category where industry money sits close to the clinical answer, and this is the one where the mechanism is most fully built out and least governed.

A public application programming interface documentation site is published at its own address, which is a real integration surface and is credited. The architecture is described as flexible and interface based, and the product is stated to integrate with the vendor's own platforms, including the care gap closure product used by health plans.

The claim about electronic health records does not go further than a sentence. The product page states that Clair integrates seamlessly into record systems and names none of them. Two retrieval passes located no marketplace listing on any record vendor's storefront, no conformance statement for a health data exchange standard, no named live customer integration and no published implementation effort.

This axis exists to separate an integration a buyer can verify from one asserted in a sentence, and this is the second case. Compare DynaMed in the same category, which holds a marketplace listing issued by the record vendor and names its standard explicitly.

Delivery is a hosted web application, native applications for both mobile platforms, and an interface based architecture for embedding. Enterprise deployment options are more developed than most vendors of this size offer: custom document ingestion, white labelled branding, a consulting engagement to shape response style for a customer's use cases, and customer defined guardrails.

Against that, the residency and hosting picture is blank. No hosting provider, no cloud region, no data residency statement, no retention position and no statement of which countries the product is available in were located. For a product that ingests customer uploaded documents, the absence of a residency and tenancy statement is a more pointed gap than it would be for a pure public corpus tool, because the buyer is being asked to place its own material somewhere unspecified.

No alternative deployment model exists for an institution that requires data to remain inside its own environment, and none is offered.

A distinctive form of non disclosure sits on this record and it is worth naming precisely. The product page carries a pricing claim as one of three headline benefits: lower cost and greater value, competitively priced because every healthcare setting should have access to innovative clinical tools. No figure, no tier, no basis and no rate card appears anywhere.

Asserting a comparative price position while publishing no price is weaker than saying nothing, because it invites a comparison the buyer cannot make and cannot check. This index has recorded vendors that gate pricing entirely and vendors that publish it openly. This is a third posture and the note records it as such.

Every route on the site ends at a demo request or a meeting booking. Mobile applications are listed on both major application stores, so an individual clinician may be able to reach the product directly, but the terms of that access were not established this pass and no self serve subscription tier is published. Enterprise pricing, which is where custom ingestion, white labelling and the interface all sit, is entirely gated.

Six named tools cover disease and drug summaries, interactions, antimicrobial stewardship, diagnosis code definitions and medication review, and the stated audience spans healthcare professionals, health plans, pharmaceutical manufacturers and life sciences.

The honest reading of the source list is that this is narrower than the general clinical reference framing implies. The named corpora are drug package inserts, a clinical trials registry, a drug interaction database and guidelines, and four of the six tools are pharmacologic. It is a strong drug and therapeutics reference presented as a general one, which is coherent with a parent company built around pharmacy, manufacturer programmes and a pharmacist chief executive, and it is a genuine strength in that lane rather than a criticism.

Held at C because no specialty coverage is enumerated, no setting is named, no geographic availability is stated, and no evidence of usefulness is published for any specialty or care setting. A buyer whose questions are diagnostic or disease management rather than pharmacologic should test that directly before assuming parity with the disease reference incumbents.