Doximity GPT

The product is a language model assistant and nothing survives its removal. There is no standalone reference library a clinician would license, no workflow platform and no device. What is sold, or in this case given away, is the ability to ask a question in natural language and receive a drafted or synthesised answer.

One qualification belongs on the record without changing the grade. Part of the asset is not the model: a structured peer reviewed drug dataset of more than 3,200 monographs, and the verified clinician identity layer the parent network has built since 2010. The dataset is content and the identity layer is a network effect. But the drug reference is one component inside an assistant, not the product a clinician opens, and the identity layer governs who may use the tool rather than what it does.

One design decision here is genuinely distinctive and no other vendor in this category states it: the company draws an explicit line inside its own product between where it retrieves and where it generates. Drug answers are stated not to be produced on the fly but returned from a structured peer reviewed dataset, and the material contrasts that with what it calls guesswork.

That matters because drug information is the highest consequence output in this category, and the vendor has chosen to make it the part of the product that does not generate. A vendor publishing which of its own answers are synthesised and which are looked up is giving a clinician something real to calibrate against. Ask every vendor here whether it can say the same.

Supporting credits: responses to clinical questions are referenced, literature is directly reachable, and a clinician can select preferred journals and sources which are then prioritised in later answers, which is user control over the corpus rather than over the interface.

Held at B on three absences. Nothing states what happens outside the structured dataset, where the system does generate. No abstention behaviour or confidence signal is described. And a claim that outputs carry physician reviewer validation appears on third party review sites rather than on the company's own material, so it is recorded here as unverified and is not graded on.

More architectural disclosure than most of this category, and it is the useful kind rather than the decorative kind.

The retrieval and generation boundary is stated explicitly. The drug corpus is quantified at more than 3,200 monographs and its provenance is described as peer reviewed and structured. Answer construction is described at the level of output design, with conclusions first and tables for comparison, which tells a clinician what shape of answer to expect. Data handling is stated with unusual precision for this axis: encryption in transit and at rest, per clinician session isolation, and prompt data never used to train the models.

What is absent is the model layer itself. No model class, no provider, no version and no retrieval architecture for the generative half were located. No evaluation harness, benchmark or accuracy figure of any kind is published.

One housekeeping point worth a buyer's attention: the product is named inconsistently across the company's own material, appearing as Doximity GPT, DoxGPT and Doximity Ask, with a support article carrying one name at a web address carrying another. That reads as a rename in progress rather than a defect, and it makes verifying claims against the right product harder than it should be.

No published evaluation of the assistant was located in two retrieval passes: no accuracy study, no error rate, no clinician review panel, no benchmark result and no peer reviewed assessment.

What exists is adjacent rather than evidential. The reach of the parent network across United States physicians is real and is verifiable in the parent's public filings, but it measures the network rather than this product, and a clinician using a free tool inside a network they already belong to is a weak proxy for the tool working. Individual clinician accounts of the product published on medical education sites are testimony, not measurement, and one such account describes uses the vendor does not itself claim, which is a reason to treat that genre carefully rather than as corroboration.

The drug reference sits differently and is credited: describing the underlying dataset as structured and peer reviewed is an evidentiary claim about content that a buyer can at least interrogate, which is more than a bare accuracy assertion.

Graded C rather than lower on that basis. The move that would change this axis is the same one available to every vendor in this category and taken by almost none: submit the assistant to an evaluation the company does not run.

The published commitments are specific rather than decorative, which is what this axis rewards. Data is stated to be encrypted in transit and at rest. Sessions are stated to be private to each individual clinician. Prompt data is stated never to be shared and never to be used to train the company's models. Access is gated to verified clinicians, so the population entering data is credentialed.

Those commitments carry more weight here than they would elsewhere, because the product invites protected health information rather than warning it away. Its entire positioning against general purpose assistants is that a clinician may put patient detail into it, and attachments for uploading laboratory results and notes are a named feature. Several vendors in this category resolve the stewardship question by refusing patient data. This one accepts it, so the commitments have to do real work.

Held at B on what is not stated: no retention period, no data residency position, no subprocessor detail, and no description of what happens to an uploaded document after a session ends.

One structural question is unaddressed on the product material and a buyer should resolve it from the parent's public filings rather than from the product page: nothing explains how a free product is funded, and the parent is a listed company whose network revenue model is disclosed there.

The claim is prominent and repeated. The product is described as fully compliant with the federal health privacy rule and as compliant by design, and that claim is the product's central differentiator against general purpose assistants.

No instrument was located. No business associate agreement, no terms of one, no execution path and no statement of who counts as a covered entity in the arrangement were found on the company's own material in two passes. Third party review sites state that business associate coverage for protected health information is the main reason to choose this product over a general assistant, which is exactly the claim that ought to rest on a document rather than on secondhand description.

The gap is more consequential here than for a reference product, and that is the point of the grade. Almost every other vendor in this category asserting compliance also arranges not to receive protected health information. This one actively invites it, positions on that basis, and does not publish the agreement that would govern it. A clinician uploading a patient document on the strength of a marketing line is relying on terms they have not seen.

Ask for the agreement, and establish whether it runs to the individual clinician or requires the employing organisation to execute it, since the two are not the same and the distinction has already caught a competitor in this category.

Two concrete controls are published and are credited: encryption in transit and at rest, and session isolation between individual clinicians. That is more specific than the general assurances several vendors in this category offer.

No attestation was located. No SOC 2 report of either type, no ISO 27001, no HITRUST, no trust centre, no penetration testing cadence and no vulnerability disclosure programme were found against this product in two passes.

Recorded as a retrieval outcome rather than as a finding that controls are absent, with one contextual note. The parent is a publicly listed company and therefore subject to disclosure obligations covering cybersecurity risk management and material incidents, so a buyer has a route to relevant information that does not exist for the private vendors in this category. That route is the parent's regulatory filings rather than the product page, and it is worth using.

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

The product's own described surface sits reasonably outside device territory. Referenced answers to clinical questions, retrieval from a structured drug dataset, drafted documentation and correspondence, and translation are all either reference or administrative acts, and the basis of a referenced answer is independently reviewable by design.

Two things keep this from being straightforward. Drafted clinical documentation goes into a patient record under a clinician's signature, which raises accountability questions even where it raises no device question. And a published clinician account of using the product describes generating predictions of likely in hospital complications, which is a materially different act from anything the vendor claims. That description is a user's, not the company's, and is recorded here as such rather than treated as a product claim, but it illustrates why a vendor benefits from stating its own intended use rather than leaving users to define it.

Graded C because no such statement exists. A single published intended use and limitation, of the kind a professional society in this category publishes, would close the question.

No governance artefact was located: no published artificial intelligence principles, no responsible use framework, no clinical advisory board, no external validation membership, no bias evaluation, no subgroup analysis and no error taxonomy.

Two design commitments do governance work and are credited even though they are not framed as governance. The commitment never to train models on clinician prompt data is a substantive constraint on the company's own use of the most valuable data it holds. And the published boundary between retrieved and generated answers is a governance decision about where the product is permitted to synthesise, applied at the highest consequence part of the surface.

What is missing is measurement, in common with almost every vendor graded in this category.

One governance question is specific to this product and untouched. A clinician may select preferred journals and sources, and those choices then shape later answers. Letting a user tune the corpus is genuinely useful and it also means two clinicians can receive different answers to the same clinical question, with no statement of whether the system discloses that a preference is in effect, or whether any floor prevents a clinician narrowing the corpus to the point where the answer becomes unrepresentative.

No integration with an electronic health record was located on the company's own material for this product. No named record system, no marketplace listing, no application programming interface, no conformance statement for a health data exchange standard and no write back path.

What exists instead is integration with the company's own surrounding products: secure messaging and fax for sending material, and attachments for bringing laboratory results and notes in. Those are real workflow connections and they are credited, but they connect the assistant to its own family rather than to the record.

That matters because a substantial part of the product's value is drafted documentation, and a chart note or discharge summary that must be copied into the record by hand has moved the clinician's work rather than removed it. Third party descriptions state that outputs can be copied into the record, which is an accurate description of copying rather than of integration and is not graded on.

The companion ambient documentation product from the same company is graded separately and a buyer weighing documentation workflow should look at both.

Delivered inside an existing professional network, reached through web and mobile, with access gated to verified clinicians. Deployment for the individual clinician is therefore close to frictionless, since the account already exists, and that distribution advantage is real and hard for competitors to replicate.

Geographic scope is stated clearly and negatively, which this index credits: access is restricted to verified United States clinicians. A buyer outside that market is not a customer at any price, and the company says so rather than leaving it to be discovered.

Held at C on everything else. No hosting provider, no cloud region, no data residency position, no retention statement and no deployment option for an organisation requiring processing inside its own environment. There is also no described institutional deployment at all: the unit of adoption is an individual clinician, so a health system cannot deploy, configure or govern the tool centrally even though its clinicians may already be using it.

The price is published, unambiguous and zero. The product is free to verified United States clinicians, including nurse practitioners, physician assistants and students, with no paid tier, upsell or usage limit described. A clinician can establish the full cost of what they will use without speaking to anyone, which remains true of very little in this category.

The eligibility condition is equally clear: verification as a United States clinician is required, and no route exists for anyone else.

Held at B rather than A for two reasons. No institutional or enterprise offering is described, so an organisation wanting to deploy, configure or govern the tool across its staff has no published option even though its clinicians can already use it individually. And nothing published explains how a free product of this scale is funded. The parent is a listed company and its revenue model is disclosed in its filings rather than on the product page, so the answer exists; a buyer weighing what a free clinical assistant costs in some other currency should read it there.

Coverage is broad across roles and tasks and is enumerated rather than implied. The stated audience runs across physicians, nurse practitioners, physician assistants and students. Specialty specific prompting is offered. The task surface spans clinical questions, literature access, a drug reference quantified at more than 3,200 monographs, chart notes and discharge summaries, prior authorisation and appeal letters, patient education, coding assistance and multilingual translation.

Translation deserves specific credit on this axis rather than being treated as a feature. Generating patient education material in a patient's own language addresses an access gap that most vendors in this category do not attempt, and it reaches a population that is systematically underserved by English language reference tools.

Held at B on two limits. Availability is confined to one country by design, so a product otherwise well suited to broad use is unavailable to most of the world's clinicians. And no evidence of usefulness is published for any specialty, role or task, so breadth of coverage is established while depth in any of it is asserted.