InpharmD

The model drives the core service, and the service would survive without it. The terms of use (effective May 2026) describe an AI-powered platform delivering drug information, literature summaries and clinical insights, and the home page describes a daily crawl of PubMed and other sources into a clinical library that the InpharmD AI draws on. Published answers carry a named author in the form of a drug information pharmacist plus InpharmD AI, with a background summary, a literature review, a stated level of evidence and references.

Remove the model and a complete, saleable product remains: a drug information service staffed by board certified pharmacists, which is what health systems used to employ in house, and the company runs a one year fellowship to train the pharmacists who staff it. The model is what changes the unit of work, from a pharmacist answering each question from scratch to a pharmacist reviewing and completing a draft.

The oversight structure is described and one part is missing. The service is built around a drug information pharmacist: the clinician submits a question with the time by which it is needed, the published answers name a pharmacist as author with the AI, and a 2023 company post describes the full pipeline, in which a pharmacist formulates the researchable question, the engine generates sub questions and a draft summary, and a board certified drug information pharmacist reviews it, discards it if it does not fit, and composes the response that is delivered. The service is asynchronous by design, which is the cost of putting a person on every answer.

What is missing is a current statement of what the AI may not do on its own. The terms of use (effective May 2026) describe an AI-powered platform delivering literature summaries and disclaim the reliability of AI generated outputs without saying that each output passes a pharmacist, the home page describes an AI library built from a daily crawl and says about a third of questions have already been asked, and a free trial account is offered alongside the pilot. Whether any AI output reaches a user without pharmacist review, for example from the library or on a trial account, is not stated, and the detailed pipeline description is three years old.

The approach and one supplier are named, without the model or any version discipline. A 2023 company post describes the pipeline step by step: inquiry parsing, a pharmacist's formulation of a researchable question, sub question generation with prompt templates and large language models, semantic retrieval against a vector database, summary generation, pharmacist review, and reinforcement from prior pharmacist judgments. The vector database is Pinecone, named by the company and in Pinecone's own customer story, whose fourfold query response time figure is the supplier's claim about infrastructure rather than a measurement of the answers. The current home page adds that the InpharmD AI draws on a library built by crawling PubMed and many other sites every day, and each answer shows a level of evidence.

No model class or provider is named for the generation step, no release notes say when the model or the pipeline changed, and the only detailed description is three years old, while the current terms describe an AI-powered platform without restating it. The home page shows a side by side comparison with an answer attributed to OpenAI GPT; that is a comparator, not a statement of what produces the company's own answers.

The architecture is described and no model provider is named. Large language models generate the sub questions and draft summaries, so a clinician's question reaches a model, and neither the model nor its provider is identified on the home page, FAQ, terms or privacy policy. Pinecone is named as the vector database in a 2023 company post and in Pinecone's customer story.

The privacy policy (effective May 2026) names recipients by category: service providers for hosting, analytics, customer support and payment processing, which process data only on the company's behalf under confidentiality and security obligations and may not use it for their own purposes; health authorities, regulators or manufacturers for adverse event and pharmacovigilance reporting; and parties to a corporate transaction. The cookie settings name Microsoft Clarity for analytics and Zendesk for support chat. No subprocessor list is published, and no retention or training terms govern data once it reaches a model provider.

The chain carries a drug information question rather than a patient record, since users are told not to submit protected health information.

Customers are named only in aggregate, and the figures are the vendor's own, with no method, denominator or reference standard. The home page and FAQ state that more than 30 health systems retain an InpharmD virtual drug information pharmacist, that the service is trusted by more than 20,000 clinical pharmacists, 161,050 clinical pharmacist hours saved, a return on investment of more than four times, and a 100 percent customer satisfaction rate. None carries a period, a method or a denominator, and the hours figure is not derived anywhere published.

The FAQ also states that, based on the largest study conducted, hospitals with this service saved $5.2 million and had 45 fewer deaths than hospitals without it. No study is cited, and the sentence does not make clear whether the study measured this product or hospital drug information services in general; read either way, it cannot be checked.

No peer reviewed evaluation, accuracy measurement or error rate for the service or its model is published. The scale figures show use, not benefit.

General assurances that do not answer the questions AI raises. The privacy policy (effective May 2026) tells users not to submit protected health information, financial information or other sensitive personal information about themselves or anyone else, and says that if protected health information is submitted inadvertently the company assumes no responsibility for its handling under HIPAA. The FAQ says the company does not take protected health information. The policy states that personal information is kept only as long as necessary and then deleted, anonymized or aggregated under retention schedules that are not published, and that the company does not sell personal information. Security measures listed include encryption in transit and at rest.

What is not stated: how long a clinician's question and the generated answer are kept, whether questions or answers are used to train or improve any model (the policy allows use to operate, maintain and improve the services), and whether anything filters patient detail typed into a free text question. The published example questions are close to a patient, for instance whether a particular population is at higher risk of skin staining with an iron infusion, so the instruction not to include patient detail carries real weight and sits entirely with the user.

No safety engineering is published beyond the pharmacist review of answers: no guardrails, no hallucination mitigation and no process for handling a safety event.

The company takes a plain position that the service does not take protected health information, explains why, and has a service privacy document behind it. The FAQ says the company does not take protected health information and that customers need not integrate, and the privacy policy (effective May 2026), which covers the services and not only the website, requires users not to submit protected health information and says the company assumes no responsibility under HIPAA for any that is submitted inadvertently. The service takes a clinician's drug information question and, on the formulary side, the health system's product and medication lists, neither of which needs a patient record.

The policy also says the company is committed to maintaining safeguards designed to protect protected health information in accordance with applicable HIPAA requirements, which sits oddly beside the instruction not to submit any; the earlier policy's plain statement that the company is not a covered entity or a business associate platform is no longer in it. No business associate agreement is offered, and with no protected health information taken, none would be.

The exposure is the free text question: the risk of a clinician including patient detail is placed on the clinician and the institution.

The company names a recognized certification without producing the report, and its own documents describe it two ways. The privacy policy (effective May 2026) states that the company maintains a SOC 2 Type II compliant environment and undergoes periodic independent audits to validate its security controls, and lists encryption in transit and at rest, access controls and monitoring, regular security reviews and vulnerability testing, and staff training. The terms of use of the same date describe SOC 2 Type II aligned controls, which is weaker language than an attestation.

The company gives no report date, auditor, trust services scope or renewal cadence, and there is no trust center, so nothing settles which of the two descriptions is accurate.

No device claim is made and none is apparent for a literature search and drug information service. The terms of use (effective May 2026) state that content does not constitute medical advice, diagnosis or treatment, that no clinician patient relationship is created, and that the services are for informational and professional use and not a substitute for a licensed clinician's independent judgment. That is a responsibility clause rather than a statement of intended purpose or a regulatory position.

The design sits close to the clinical decision support exclusion in the 21st Century Cures Act: the output is a cited literature synthesis with a stated level of evidence, and a licensed pharmacist is named as its author, so the basis is reviewable by the clinician who receives it. That is an observation about the design; the company publishes no regulatory rationale.

A related question sits outside device classification: a licensed pharmacist composing a written answer about a clinical situation is practicing pharmacy, and nothing published says which state licenses the reviewing pharmacists hold.

Governance runs through professional review rather than published AI policy, and nothing about the model is measured. The published answers name a drug information pharmacist as coauthor with the AI, each answer grades its evidence (for example level D, case reports or unreliable data), the company trains its reviewers through a one year drug information fellowship, and a 2023 company post describes reviewer judgments feeding back into retrieval. That is a competence standard applied to a named person at the point of output.

No AI principles, responsible use framework, external advisory body, bias evaluation, subgroup analysis or error taxonomy is published. The terms disclaim any guarantee of the accuracy, completeness or reliability of AI generated outputs.

The one figure the service holds by construction and does not publish is the rate at which its pharmacists reject or materially rewrite the generated summary. It would evidence the review step and measure the model's error rate at the same time.

Mechanisms exist that let a clinician challenge an answer, and nothing stands behind the output. Each published answer names the drug information pharmacist who authored it with the AI, cites its references and states a level of evidence, so a clinician can check the literature and knows which licensed professional signed the answer. The FAQ offers a request form.

The terms of use (effective May 2026) provide the services as is and state that the company does not guarantee the accuracy or completeness of content, including AI generated outputs, or their reliability. Liability is capped at the greater of the fees paid in the prior twelve months or $100, and users indemnify the company for their use of the services.

The pilot page promises that the company will guarantee outcomes or be the first to say there is no value, and a business case with an ROI guarantee at the point of purchase. That is a guarantee about value to the health system, not about whether an answer is right, and its terms (what is paid or refunded if the return misses) are not published.

No accuracy figure, error rate or correction commitment is published.

Integration is claimed with no system named and nothing to verify. The FAQ says that because the company takes no protected health information, a health system does not have to integrate, though the company can integrate with the customer's EHR and wholesaler. No record system, wholesaler, marketplace listing, application programming interface or health data exchange standard is named on the home page, FAQ, terms or privacy policy.

The product is used alongside the record: a clinician submits a question through the web or mobile app and receives an answer. That fits an asynchronous service. The formulary work depends on a health system's own medication and product data, and the pilot page says the company compiles the customer's data into a business case, so some data transfer happens; how, and in what format, is not published.

A single hosted option with location implied rather than committed. The service is a web application with iOS and Android apps, and the FAQ describes it as running in the cloud with no data or integration needed from the health system's IT team. The privacy policy (effective May 2026) says the company operates in the United States and that information from users outside it may be transferred to, stored and processed in the United States or other countries, with standard contractual clauses where required.

No hosting provider, region, tenant isolation model or on premises option is published. A 2023 company post and the vector database vendor's own customer story name Pinecone as the retrieval layer, which is a processing location for the library and says nothing about where a question rests.

The formulary side of the business uses a health system's own medication and product data, which is commercially sensitive even when it identifies no patient, and nothing says where that data is held.

No price is published and the sales posture is discoverable in unusual detail. The pilot page sets out the route: a meeting to establish fit, then a free pilot with a 30 day calibration period aligning the service to the health system's templates, workflows, formulary and style guide, while the company compiles the customer's data into an internal business case. The company says it will guarantee outcomes or be the first to say there is no value, and that a customer moving forward gets a business case and an ROI guarantee on the spot. A free trial account with self serve sign up is offered as an alternative, and the home page offers to get started free.

What is not published is anything a buyer can size a cost from: no seat or subscription price, no unit of charge, no tier, no minimum commitment and no implementation fee, and no terms for the ROI guarantee. A process a buyer can run before committing money is useful, and it is a process rather than a price.

Coverage is stated clearly and nothing validates it. The scope is drug information and formulary work for health systems: clinical pharmacists are the primary users, the terms extend use to qualified healthcare professionals, and the FAQ says the service is not yet available to patients. The published example questions show the depth intended within that scope, including iron sucrose administration in pediatric patients, flat dosing of a prothrombin complex concentrate for anticoagulation reversal, fibrinogen dosing in postpartum hemorrhage, and the effect of medication reconciliation on readmissions.

No evaluation by question type, specialty or setting is published, no ambulatory or community pharmacy offering is described, and the terms make no representation that the service is appropriate in any particular jurisdiction or clinical context.