CaryHealth Clair

The artificial intelligence is the product. Clair's underlying sources are public: FDA drug package inserts, the NIH clinical trials database, the National Library of Medicine drug interaction database and published clinical guidelines, plus whatever a customer uploads. The company does not own a curated corpus built over decades and does not claim to. What it sells is the retrieval, routing and generation layer over that material, organized as six tools (open search, clinical summary, drug interactions, antimicrobial stewardship, ICD-10 code definitions and medication review), with answers linked to their references.

The terms of use (last updated 10 November 2025) describe Clair as a conversational query tool that generates drug information and clinical summaries using third party large language models, including OpenAI's. Remove the model and what remains is a set of public databases anyone can already reach.

Clair stands apart from CaryHealth's pharmacy, manufacturer access, field medical and care gap products, which are separate lines.

The company publishes an absolute claim about the accuracy of every answer, which grades here whatever review sits behind it. The Clair product page says a team of clinical experts continuously monitors search results to ensure all responses are accurate and credible, that real time processing ensures responses without outdated guidance, and that open search gives an instant, ultra accurate answer to any complex clinical question; the company's App Store listing says answers are generated through a semantic search algorithm that guarantees zero hallucinations. A buyer who believes all responses are accurate and current has no reason to build the review that would catch the one that is not.

The company's own terms of use contradict the promise: AI features may not generate accurate information, information provided may be out of date, and users should independently verify responses.

The structure that does exist would otherwise sit higher: every response links to its references, the product is split into six narrow tools, and the stewardship tool bounds its output to treatment, dose and duration for a stated infection. Against that, the product page says the system continuously learns from user interactions and feedback to refine its models, and nothing published describes a validation gate on that loop, a versioning scheme, who approves a change before it reaches a clinician, or what the system does when sources conflict or say nothing.

A model supplier and the technique are named, and the company's own descriptions of the model disagree. The terms of use (last updated 10 November 2025) say Clair is powered by third party large language models and developers, including but not limited to OpenAI. The product page names retrieval augmented generation, embeddings, complex pipelines and self learning, and describes intelligent sourcing that routes a question to package inserts, the trials registry or the interaction database by its type.

The product page also says the models are trained on the company's database of 200 million or more clinical studies, 147,000 or more FDA inserts, 2 million or more drug interactions and 100 or more guidelines, and elsewhere that they are large language models trained on diverse medical texts; the App Store listing calls it a proprietary AI model trained on FDA and NIH data. Training on a corpus, retrieving from it and calling a third party model are different architectures with different failure modes, and a buyer cannot tell which combination ships. The 200 million clinical studies figure is not defined and exceeds common counts of the published biomedical literature.

No model version is named and no release notes say when the model or the pipeline changed, which matters more because the company says the system continuously learns from user feedback.

The model provider is named, the list is left open, and the terms that follow the data are the company's own use rather than a protection. The terms of use (last updated 10 November 2025) say Clair runs on third party large language models and developers including but not limited to OpenAI, and that questions and prompts are provided to third party LLMs to generate responses. The privacy policy (last updated 14 May 2025) lists disclosure of personal information to a third party LLM vendor such as OpenAI for AI features, and to contractors and service providers bound to confidentiality and to use it only for the purposes disclosed.

What is not published: the full list of model providers the phrase including but not limited to leaves open, any subprocessor list or hosting provider, and the retention and training terms that govern prompts once they reach OpenAI or any other provider. The company's own stated use runs the other way: the privacy policy says it uses AI feature data to improve the LLMs it uses, and the terms grant it a perpetual license to what users submit. A customer's uploaded documents follow the same chain.

One independent evaluation includes Clair, and on what the abstract reports it does not support the company's accuracy claims. Ipema, Munir, Sarna, Wasynczuk and Saiyad, University of Illinois Chicago Retzky College of Pharmacy, Journal of the American College of Clinical Pharmacy 2026;9(6):e70231 (doi 10.1002/jac5.70231): 30 drug information questions previously answered by drug information pharmacy faculty were put to four clinically focused chatbots (OpenEvidence, Clair, Glass Health and DougallGPT) and to ChatGPT, and three pharmacy faculty scored each response independently on the 25 point CLEAR framework (completeness, lack of false information, supporting evidence, appropriateness, relevance). OpenEvidence scored highest at 17.82 and ChatGPT second at 15.72, both in the average band, so Clair scored lower than the general purpose chatbot; every chatbot scored significantly below the faculty answers, lack of false information scored highest across chatbots and supporting evidence lowest. No company author is listed. Clair's own score and component results sit in the full text, which is behind a browser challenge (28 September 2026), so the product's figure cannot be stated here.

The company's own figures carry no method or source: the App Store listing says Clair can reduce search time by up to 90 percent, and the product page says it is faster and more concise than legacy tools and safer and more accurate than generic AI solutions. No named customer deployment or outcome is published. The company reports industry awards and growth rankings, which measure the business rather than the answers.

The practices are stated, and what they state is use of what users type rather than protection of it. CaryHealth's privacy policy (last updated 14 May 2025) covers Clair and lists chat and prompt logs, output logs, content and usage data among what it collects; among its stated uses are to develop, improve and test the service and, for AI features, to improve on the large language models the company uses; and it discloses personal information to a third party LLM vendor such as OpenAI. The terms of use (last updated 10 November 2025) say questions and prompts are provided to third party LLMs to generate responses, that content is used to improve the AI features, and grant the company a worldwide, royalty free, transferable, perpetual license to content users submit, including content entered into AI features.

The same policy covers the whole company, including a pharmacy that handles protected health information, so the policy lists protected health information among what the company may collect, and the stewardship tool accepts optional patient demographics. No retention period for prompts, no de identification step before prompts reach the model provider, and no segregation of a customer's uploaded documents between tenants is stated.

No safety engineering is published: no guardrails of the company's own (enterprise guardrails are defined by the customer), no hallucination mitigation beyond reference linking, and no process for handling a safety event.

A substantive privacy document reaches the service, and no business associate status or scope position is stated for Clair. CaryHealth's privacy policy (last updated 14 May 2025) covers Clair and its AI features by name, including the chat, prompt and output logs they collect, and the terms of use incorporate it. The company operates a licensed pharmacy that is a HIPAA covered entity, and its notice of privacy practices (effective 8 July 2025) says it applies only to the pharmacy services, not to CaryHealth's digital platforms.

Nothing says whether Clair is meant to receive protected health information, and the company does not tell users to keep it out: the privacy policy lists protected health information among what the company may collect across its services, the stewardship tool accepts optional patient demographics, and enterprise customers upload their own documents. No business associate agreement, tier or execution path is published for Clair, and the enterprise offering's claim to meet the compliance and security needs of enterprise partners names no requirement.

Nothing independent stands behind the security claims. The product page says the API architecture meets the unique compliance and security needs of enterprise partners and offers customer defined guardrails for compliance and security, without naming a control, a framework or an auditor. The privacy policy tells users that no security measures are perfect or impenetrable. No SOC 2 report, ISO 27001 certificate, HITRUST certification, trust center, penetration testing statement or vulnerability disclosure program is published on the product page, the site footer's legal pages (terms, privacy policy, notice of privacy practices, attributions) or the developer documentation link.

Public API documentation is a technical surface, not an outside check on how the company operates security.

No device claim is made, and the company's responsibility clause sits in tension with what two of its tools are sold to do. The terms of use (last updated 10 November 2025) carry a Clair AI medical disclaimer: Clair is not a healthcare provider, provides AI generated drug information for educational purposes only, and does not provide medical advice, diagnosis or treatment recommendations. The product page sells a stewardship tool that returns treatment, dose and duration information for an infection, adapted to optional patient demographics, and a medication review tool described as providing evidence based recommendations for patient care. Both are addressed to clinicians, who remain responsible for the decision, so this is tension between a clause and the marketing rather than a contradiction of a stated regulatory position.

No clearance, submission or regulatory rationale is published. Answers link to their sources, which is the reviewable basis the clinical decision support exclusion in the 21st Century Cures Act turns on; dosing adapted to patient demographics is the output where that basis is tested hardest, and the company's statement that its models continuously learn from user feedback bears on it too.

Governance is asserted as a process, with no names, method or output behind it. The product page states that a team of Clair clinical experts reviews all data sources in a vetting process and continuously monitors search results, and that users can rate and give feedback on results so the company can refine them. No AI principles, responsible use framework, named clinical lead, evaluation method, bias assessment, subgroup analysis or error taxonomy is published.

The enterprise offering adds tailored AI guardrails, described as letting a customer define custom AI governance frameworks to its own standards, alongside upload of the customer's own documents and criteria for the model to answer from, white labeled branding, and a consulting engagement to shape response style. The company's industry navigation lists pharmaceutical manufacturers first. Together those describe a clinical reference a manufacturer could deploy under its own brand, answering from its own documents, under guardrails it defines. Nothing published shows the configuration used that way, and nothing published says what the company requires of a customer's guardrails or uploaded content.

An independent evaluation that included Clair (Journal of the American College of Clinical Pharmacy, 2026) is not cited on the company's pages; it is recorded under clinical and operational evidence.

Mechanisms exist that let a clinician challenge an answer, and the terms disclaim everything behind the output. Every response links to its references, and users can rate and give feedback on results.

The terms of use (last updated 10 November 2025) state that Clair provides AI generated drug information for educational purposes only and does not provide medical advice, diagnosis or treatment recommendations; that AI features may not generate accurate information and information provided may be out of date; that users should independently verify responses; that the company gives no warranty or guarantee as to the accuracy of information provided by AI features; that AI features may contain errors or design flaws; and that the company is not responsible or liable for the results of decisions made on health related content.

The marketing makes promises the terms withdraw: the App Store listing says answers come from a semantic search algorithm that guarantees zero hallucinations, and the product page says clinical experts monitor results to ensure all responses are accurate. None of those promises carries a remedy, an error rate or a correction route that binds the company.

Integration is claimed with no record system named. The product page states that Clair integrates seamlessly into EHR systems and CaryHealth platforms and describes a flexible API based architecture for integrating into an organization's infrastructure. Public API documentation is published at documentation.cary.health, which is a real integration surface for developers.

No record system is named, and no marketplace listing, health data exchange standard, named live customer integration or implementation effort is published on the product page, terms, privacy policy or footer. The integration a buyer can verify from published material is the API, not a connection to a record.

A single hosted option with location implied rather than committed. Clair is delivered as a hosted web application at askclair.ai, iOS and Android apps, and an API for embedding, from Cary Rx Incorporated in Washington, DC. The terms of use offer the service only to users who reside in the United States or its territories. Enterprise options include custom document ingestion, white labeled branding, a consulting engagement to shape response style, and customer defined guardrails.

No hosting provider, cloud region, data residency statement or tenant isolation model is published, and prompts go to third party language model providers including OpenAI, whose processing location is not stated. For a product that ingests a customer's own documents, nothing says where those documents rest or whether they are kept apart from other customers' material. No on premises or customer environment deployment is offered.

A price is published without full tiers. The company's App Store listing, a field the developer controls, shows the app as free to download with in app purchases, says a subscription is required after the trial period ends, and names one: Clair Professional at $9.99. The web product at askclair.ai and the Android app are linked from the product page for individual clinicians to start without a sales conversation.

What is not published: the subscription's term, the length of the trial, and any enterprise pricing, which is where custom document ingestion, white labeling, customer defined guardrails, the consulting engagement and the API sit; every enterprise route on the site ends at a demo request or meeting. The product page also claims lower cost and greater value, saying Clair is competitively priced because every healthcare setting should have access to innovative clinical tools, a comparison the published figure only partly lets a buyer check.

Coverage is claimed broadly, and the sources behind it are narrower. The product page presents Clair as having a tool for any clinical question, across healthcare professionals, health plans, pharmaceutical manufacturers and life sciences, and the terms limit use to adults in the United States. The six tools cover disease and drug summaries, drug interactions, antimicrobial stewardship, ICD-10 code definitions and medication review, and the named sources are drug package inserts, a clinical trials registry, a drug interaction database and about 100 clinical guidelines, so four of the six tools and most of the corpus are pharmacologic. That is a drug and therapeutics reference presented as a general one, coherent with a parent built around pharmacy and manufacturer programs.

No specialty or care setting is enumerated, and the one independent evaluation (Journal of the American College of Clinical Pharmacy, 2026) tested drug information questions only.