InpharmD

A genuine hybrid, and the services test is the right lens rather than the usual one.

The artificial intelligence is load bearing rather than decorative. It parses the inquiry, generates sub questions from a researchable question using prompt templates and large language models, retrieves against a vector database, and drafts the summary. The company's own framing puts the model first and the human second: the information is described as artificial intelligence generated and clinical pharmacist reviewed, in that order.

What holds it short of A is that a complete saleable product survives the removal of the model. What remains is a drug information service staffed by board certified pharmacists, which is exactly what health systems used to employ in house, and the company runs a fellowship to train the pharmacists who staff it. The workforce is part of the asset, not just a quality gate on it.

Graded B rather than penalised further because the honest reading is that the model is what makes the economics work. Without it a pharmacist answers one question at a time from scratch; with it the pharmacist reviews and corrects a draft. That is a real change in the unit of work, not a wrapper.

The clearest human in the loop architecture in this category, and one of the clearest anywhere in this index, because the human is load bearing rather than advisory.

A board certified drug information pharmacist sits between the model and the clinician on every inquiry. That pharmacist formulates the researchable question the engine works from, reviews the generated summary, discards it outright if it does not match the inquiry, and composes the response that is actually delivered. Nothing generated reaches a clinician unreviewed. Prior pharmacist judgements feed back into later retrieval, so the human contribution accumulates rather than being spent once.

What makes this credible rather than a slogan is the architectural cost the company accepted to get it. The service is asynchronous by design, and the user states how quickly they need an answer. Every competitor in this category has optimised for an instant response, which is precisely what makes review by a named professional impossible. This vendor traded latency for oversight and says so.

Set it directly against the two nearest comparisons the index now holds. One competitor states that every response is validated by a clinician in language that cannot plausibly mean per response review, and the ambiguity is unresolvable from outside. Another describes models that continuously learn from user interactions with no validation gate described at all. This is what an in the loop claim looks like when it is real.

The question that follows from the architecture, and the vendor has the data by construction: what proportion of generated summaries do reviewing pharmacists reject or materially rewrite. Publishing that number would be the strongest single disclosure available to any vendor in this category.

The pipeline is published step by step in more detail than almost anything else in this category: inquiry parsing, pharmacist formulation of a researchable question, sub question generation using prompt templates and large language models, semantic retrieval against a vector database, summary generation, pharmacist review, and reinforcement from prior pharmacist judgements. The assistant is named, and the vector database provider is named openly, which is a subprocessor disclosure most vendors here do not make.

Two limits hold it at B. No model class or provider is named for the generation step. And the detailed description sits in a company blog post from 2023, so a buyer cannot tell how much of it still describes the current system three years on. That currency gap matters more for a pipeline description than for a static claim.

One quantified figure exists and its provenance should be stated: a fourfold improvement in query response time appears in a case study published by the vector database vendor, a supplier with a commercial interest in the result. Treat it as infrastructure marketing rather than as a measurement of the product.

No peer reviewed evaluation, no accuracy measurement and no published error rate were located in two retrieval passes.

What exists falls into three categories and each needs handling differently. Deployment figures are real but are not evidence of correctness: more than 30 health systems retaining the service, and an earlier figure of more than 10,000 clinician users. A named commercial mechanism with a number does exist, a health system saving several million dollars annually by standardising a drug class after a formulary review, and the mechanism is credible because formulary standardisation savings are straightforward to compute; but the account appears in a sponsored publication whose author also discloses being an investor, so it is a self interested source twice over and is recorded rather than relied on.

The third category is the one to be careful with. The company's blog carries outcome percentages including a 30 percent reduction in medication errors, a 20 to 30 percent decrease in adverse events and an 18 percent cost reduction, with no study cited and no cohort described. Read in context these appear to be general claims about what evidence based practice achieves rather than measurements of this product, but they sit inside a post about this product and a reader will not make that distinction. Ask which of those figures, if any, were measured on this service.

The honest summary is that the strongest evidence for this vendor is commercial: more than 30 health systems pay for it, and a health system that buys a virtual pharmacist has done its own arithmetic.

The stated position is unambiguous and architecturally coherent. The privacy policy states that the company is not a covered entity or a business associate platform for the storage or processing of protected health information, instructs users not to submit protected health information, financial information or other sensitive personal information about themselves or anyone else, and states that only the information necessary to provide and improve the service is collected.

Coherence is worth something and this index credits it. The service answers questions about drugs and the literature, so it does not need a patient record to function, and saying so plainly is better than the contradictory positions recorded elsewhere in this segment.

What holds it at C is the gap between that policy and how the product is actually used. The example inquiries the company publishes are patient specific in substance: whether a particular patient with a severe infusion reaction to one agent can transition to another, or the risk profile for a named induction regimen in transplant patients. A clinician asking a question that specific will frequently include clinical detail, and the policy places the entire consequence of that on the user. No encryption practice, retention period, data residency position or de identification approach was located.

The practical question for a health system: what happens operationally when a clinician does paste patient detail into a free text box, and is there any filtering or warning at the point of entry rather than only a clause in a policy.

Unusual in this category for being completely unambiguous, and weak on the merits, which are two different things and the record should carry both.

The company states directly in its privacy policy that it is not a covered entity or a business associate platform for protected health information, that users agree not to submit it, that if it is submitted inadvertently the company assumes no responsibility for its handling under the relevant rules, and that the user is solely responsible for ensuring their use complies. Terms of use are governed by Delaware law and place the obligation to comply with applicable healthcare and privacy law on the user.

Credit where it is due. This index has recorded a competitor whose security page invites covered entities to transmit protected health information under a standard agreement while its about page states the opposite, and several vendors that simply never mention the subject. A clear, consistent, published refusal is more useful to a buyer than either.

Graded C nonetheless, because the axis measures posture and this posture offers a health system nothing to sign. No business associate agreement is available at any tier, and the risk of inadvertent disclosure through a free text clinical question is allocated entirely to the clinician and their institution. For a product sold into health systems and used by their pharmacists on named patient problems, that allocation deserves to be surfaced during procurement rather than discovered in a policy afterwards.

No certification 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 and no vulnerability disclosure programme.

One real disclosure sits on the credit side and it is unusual for a company this size. The vector database provider underpinning retrieval is named publicly, which tells a buyer where a material part of the processing happens. Most vendors in this category publish no subprocessor information at all, and one publishes a processor list only on request.

Recorded as a retrieval outcome rather than as a finding that controls are absent. A vendor selling to more than 30 health systems will have completed security reviews for each of them, so documentation exists; the question is why none of it is published. Ask for the security assessment package that those health systems already received, which is the fastest route to an answer here.

No clearance, no submission and no published regulatory positioning statement were located, and none would be expected for a literature search service.

This record sits further from the regulatory line than anything else in the category and the architecture is why. The clinical decision support exclusion turns on whether a clinician can independently review the basis of the output. Here the output is a cited literature synthesis, and a licensed pharmacist has already reviewed it before the clinician sees it, so there are two independent reviewable layers rather than one. A human professional taking responsibility for the response also changes who is accountable in a way software alone does not.

Graded C only because nothing is published stating that position. The related question worth asking is not about device classification but about professional practice: a licensed pharmacist composing a written response about a specific patient's therapy is performing a professional act, and a health system should establish which state licences the reviewing pharmacists hold and how that is handled across state lines.

No published artificial intelligence principles, no responsible use framework, no external advisory council, no membership of any external validation body, no bias evaluation, no subgroup analysis and no error taxonomy were located.

What exists instead is a control that most vendors in this category do not have, and it should be credited on this axis rather than only on oversight. Governance here runs through professional licensure: every response passes a board certified drug information pharmacist, the company trains those reviewers through its own year long fellowship, and reviewer judgements are fed back as a signal into retrieval. That is a documented competence standard applied to a named human at the point of output, which is a stronger control than a published principle with nothing behind it.

Compare the two nearest cases in this category. One vendor sells governance as a configurable customer option with the standard supplied by the customer. Another publishes five substantive principles and no measurement. This vendor publishes no principles and has the most concrete control, which is a useful reminder that the two halves of this axis measure different things.

Held at C because nothing is measured or published. The single most valuable disclosure available to this company, and one it holds by construction, is the rate at which its own pharmacists reject or rewrite the generated summary. That number would evidence the oversight claim, quantify the model's error rate, and be almost impossible for a competitor to match.

No integration claim of any kind was located. Two retrieval passes found no named electronic health record, no marketplace listing, no conformance statement for a health data exchange standard, no application programming interface and no published integration path.

Delivery is a web application and native mobile applications on both platforms, with the workflow being submit a question and receive a response, so the product is used alongside the record rather than inside it. That is a coherent design for an asynchronous service, and integration matters less for a product a clinician consults deliberately than for one meant to surface in context.

It is still a real limitation for the pharmacy and therapeutics side of the business, where formulary work depends on a health system's own medication data. The company describes taking a customer's product list to run standardisation analysis, which implies a data exchange of some kind, and nothing published describes how that transfer happens or under what terms. Ask about that path specifically rather than about record integration generally.

Hosted service reached through a web application and native mobile applications on both platforms, deployed at more than 30 health systems with seats bought for clinicians and the pharmacy and therapeutics committee as the usual entry point.

One partial infrastructure disclosure is on the record and is credited: the vector database provider is named. Beyond that, no hosting provider, no cloud region, no data residency statement, no retention position and no geographic availability statement were located. There is no alternative deployment model for an institution that requires processing to stay inside its own environment.

The residency question is less acute here than for peers, because the company states it does not accept protected health information, so in principle what is transferred is a clinical question and a literature response. It is not moot, though, since the formulary arm involves a health system's own medication and utilisation data, which is commercially sensitive even when it is not patient identifiable.

No price is published, and this still earns a B, which needs explaining.

What the company publishes is a commercial process detailed enough for a buyer to run before committing money. A free pilot is offered explicitly so the vendor can use the customer's own data to establish whether value exists. A 30 day calibration period is stated, during which the service is aligned to the customer's templates, workflows, formulary and style guide. At the end the buyer is given a business case built from their own data and a stated outcome guarantee, with the company committing in writing to be the first to say if there is no value. A free trial account with a self serve signup flow exists alongside the enterprise route.

This index credits published mechanism, and a process a buyer can actually execute is materially more useful than a comparative price claim with no figure behind it, which is what a peer in this category offers. It is also unusually candid for a vendor to publish that its own pilot might conclude the product is not worth buying.

Held at B rather than A because it is a process and not a price. No seat rate, no tier structure, no subscription basis and no implementation fee schedule were located, so a buyer can establish whether to proceed but not what proceeding will cost until they are inside a sales conversation.

Deliberately narrow, and the narrowness is a design choice rather than a shortfall, which the note records so the grade is not misread.

Scope is drug information and formulary work in the health system setting. The primary users are clinical pharmacists and pharmacy and therapeutics committees, with physicians and nurse practitioners consuming responses. Coverage within that scope is deep and the published example inquiries show it: transitions between monoclonal antibodies after an infusion reaction, cumulative dosing questions in transplant induction, dose dependent cardiac effects of common antiemetics. These are specialist questions that a general reference tool answers poorly or not at all, which is the actual argument for the product.

Graded C because the axis measures coverage. No specialty breadth is claimed, no care setting outside the health system is addressed, no ambulatory or community pharmacy positioning exists, and no geographic availability is stated. A buyer looking for a general clinical reference is looking at the wrong product, and a buyer with hard pharmacotherapy questions is looking at one of the few products in this category built for them.