UpToDate

This is the strongest instance in the index of a vendor whose moat is the dataset rather than the model, and the company frames it that way itself, calling the generative product the next evolution of the reference product and stating that it is built solely on the existing curated corpus. The asset is more than thirty years of content maintained by over 7,600 physician authors and editors and used by a reported three million clinicians. Remove the generative layer and what remains is the market leading clinical decision support product, which is what it was until September 2025.

The grade describes the relationship between the mechanism and the moat, not the sophistication of the engineering, and that distinction needs stating plainly here. The AI work is more advanced than at several vendors graded higher on this axis, with a stated multi layer validation framework, an assumptions and rationale architecture, and a named enterprise model partnership.

It is graded C because a buyer removing the model still holds almost all of the value. That is the opposite of the position at a vendor whose product ceases to exist without its model, and this axis measures that dependency rather than quality.

The best oversight architecture for a generative clinical product in this index, and it earns the grade on two mechanisms nothing else here has.

The first is the assumptions block. Every response opens by naming what the system took for granted about the patient, in the vendor's own illustration that the patient is not pregnant, not a child and has no major comorbidities. A model that states its own premises before answering lets a clinician invalidate the answer in one glance rather than reading to the end to discover it was answering a different question.

The second, and more important, is the nudges: inline prompts asking the clinician to supply context the system lacks, such as comorbidities or intensive care status. This index asks nearly every vendor what its system does when it does not know, and almost never gets an answer. This is a real answer, and a different shape from the one expected. Rather than abstaining or emitting a confidence score, the system names the gap and asks the user to close it. Supporting this are inline citations tied to specific statements, and positioning that customers themselves endorse, one quoted preferring supportive decisions over automated ones.

One counterweight is recorded and precisely attributed. Trade coverage of the launch reported the company as saying that closing the system to the broader web eliminates the chance of hallucination. That claim does not appear on the product page, which uses the more careful language of grounding and traceability. But if it is made in a sales setting it is an overclaim, because grounding constrains the source material without eliminating synthesis error, and telling clinicians a tool cannot hallucinate is precisely the message that lowers the scrutiny this architecture is otherwise designed to support.

Strong on provenance and reasoning, silent on the model itself.

What is published is unusual and creditable. Every answer carries inline citations tied to specific statements with single click access to the source topic, a side panel exposing rationale and supporting material, and a stated assumptions block. The company named its enterprise model partner publicly in June 2026, which direct competitors in this segment both decline to do, and which lets a buyer reason about the underlying technology at all. It has also published a validation framework for evaluating clinical AI at the point of care, a whitepaper on the approach, and an argument that clinical AI evaluation must go beyond benchmark wins.

What is absent is every number. No model class or version, no evaluation results, no hallucination or error rate, no benchmark performance, and no published output of the validation framework the company itself designed.

Publishing a framework and withholding its results is a distinct posture worth naming, and it is what separates this from the top band. A buyer should ask for the measured output of that framework specifically, since the company has already conceded it is the right way to assess the product.

The model partner is named, which the direct competitors in this segment both decline to do, and that alone lets a buyer reason about the underlying technology rather than treating it as a black box. What is unaddressed is what the interface receives.

The generative layer is conversational and invites clinicians to describe cases in natural language, prompting them to add comorbidities and intensive care status, so free text clinical detail about identifiable patients will predictably be typed into it, and the product is reachable from inside the record system where that detail is already on screen.

Nothing located states whether queries are retained, for how long, whether they are used to improve the service or the underlying models, or what the model partner receives and holds. One distinction matters more here than at most vendors and a buyer should establish it first.

The same product is sold as a personal subscription bought with a credit card and as an enterprise deployment under a health system agreement, and only the second route puts an institution in a position to negotiate terms. A clinician who bought their own licence and types a case description into it is contracting personally on consumer terms while handling their employer's patient data. Ask for query retention and training use per route, what the model partner receives, and what the institution can require of individually purchased licences.

Graded on what was retrieved, and the retrieval gap is stated explicitly because it matters more here than on most records.

What was located in this pass is vendor produced or vendor promoted: whitepapers on evaluating clinical AI, a published validation framework, an analyst report the company promotes as independent validation, company surveys of physicians, and named customer testimony from St. Luke's University Health Network and the University of Texas Medical Branch. No published evaluation of the generative product's accuracy, no error rate, and no peer reviewed study of it was found.

An important caveat applies to any refresh. The company maintains a research studies page, and the underlying content product has an older peer reviewed outcomes literature associating its use with hospital level outcomes. None of that was retrieved in this pass. It was not assessed and is not reflected in this grade.

Pull it specifically on refresh, because this axis could move materially. The absence recorded here is a retrieval limit rather than an established evidentiary gap, and the distinction matters for a product of this age and scale.

A privacy and cookies policy and terms of use are published and linked, and neither was opened in this pass. That scope limit is recorded explicitly rather than left implicit, and it is the first thing to close on a refresh.

The product specific question is live and unanswered by anything read. The generative layer is a conversational interface that invites clinicians to describe cases in natural language and prompts them to add comorbidities and intensive care status, so free text clinical detail about identifiable patients will predictably be typed into it. The product is also reachable from inside the electronic health record.

Nothing located states whether queries are retained, for how long, whether they are used to improve the service or the underlying models, or how the answer differs between a personal subscription bought with a credit card and an enterprise deployment under a health system agreement. That last distinction matters more here than at most vendors, because the same product is sold both ways and only one of those routes puts an institution in a position to negotiate terms.

UpToDate takes an unusual and explicitly published position. Rather than describing how it protects patient information, its terms of use state that UpToDate and its licensors are not intended to receive, transmit, store or process protected health information under United States HIPAA regulations. The product is contracted as a clinical reference and decision support resource rather than as a processor of patient data, and no business associate agreement is offered because the service is scoped to sit outside that relationship.

This is clearer than most vendors manage and it earns credit for being a stated position rather than a silence. A buyer knows exactly where they stand: do not enter patient identifiers, and do not expect the vendor to accept business associate obligations.

Held at B because a disclaimer allocates risk rather than preventing the problem, and the product design makes the problem foreseeable. UpToDate is embedded into electronic health records within the clinical workflow, and the company now ships a generative artificial intelligence component, UpToDate Expert AI, which invites clinicians to ask questions in free text. A clinician working a live case is precisely the person most likely to type patient specifics into that box, and nothing published states what happens when they do, whether such queries are logged, how long any logs persist, or who can read them.

Buyers should ask what the vendor does with free text query content, whether queries are retained and for how long, and whether a business associate agreement can be obtained for integrated or enterprise deployments despite the standard terms disclaiming the relationship.

No security certification page, trust centre or attestation specific to UpToDate was located. The product's policies and legal section publishes terms of use, a privacy and cookie notice and jurisdiction specific privacy addenda, but nothing describing security controls, independent audits or certifications for the service itself. That is a statement of what was found rather than a claim that none exists, and it is the first thing to test on a refresh.

The absence is notable because of what the same parent publishes elsewhere. Wolters Kluwer maintains dedicated privacy and security certification pages for products in its other divisions, naming ISO 27001, ISO 27701 and SOC 2 Type 2, and one of those pages describes a group level security programme aligned to ISO 27001 and the NIST Cybersecurity Framework under executive oversight. Those certifications belong to those products and those divisions. This record does not import them, because a buyer cannot rely on an attestation whose scope excludes the service being purchased, and a parent company certification says nothing about which systems were actually in the assessment boundary.

Buyers should ask which certifications cover UpToDate specifically, request the scope section of any report offered so the covered systems can be checked against the service being bought, and then ask the same questions separately about UpToDate Expert AI, since a newer generative component may sit inside a different assessment boundary from the reference content it draws on.

No FDA clearance, none claimed, and none apparently required. The grade records the absence; the note records that the regulatory position is strong and that this product illustrates the underlying test better than most.

The clinical decision support exclusion turns on whether a clinician can independently review the basis for a recommendation. Here the basis is an inline citation to a specific expert authored topic, alongside a stated set of assumptions and a rationale panel. The clinician can read the source, see what was assumed, and follow the reasoning. That is the same structural position held by the strongest pathway products in this category, and the direct inverse of an embedded predictive score whose basis is disclosed to nobody.

The company never states a regulatory rationale explicitly, so ask for it directly. Ask specifically how the position is maintained as the product moves toward the agentic capabilities named in its enterprise model partnership, because an agent that acts is a different regulatory object from a reference that answers.

The grade rests on published artefacts rather than assertions. The company publishes a corporate AI principles statement, has released a validation framework for evaluating clinical AI at the point of care as a public document, and has published a whitepaper describing how the generative product is assessed for clinical usefulness, knowledge integrity and potential risks, alongside an argument that clinical AI evaluation must go beyond benchmark wins. It also publishes survey research on unsanctioned AI use inside healthcare organisations, which is self interested but genuinely useful and an unusual thing for a vendor to fund.

The gap is that the framework is published and its results are not. No subgroup analysis, no performance by specialty or question type, no error taxonomy and no measured output of the company's own evaluation method appear anywhere.

A second exposure is structural and unaddressed. The corpus is authored by 7,600 experts through an editorial process, so whatever is under represented in that editorial tradition is under represented in every answer the generative layer produces. Grounding to a curated corpus inherits the corpus's blind spots as faithfully as it inherits its rigour, and that is a question the validation framework is well placed to answer and does not.

This company authored the test and has not published its score, which is a distinct posture worth naming rather than folding into an ordinary disclosure gap. It has published a validation framework for evaluating clinical artificial intelligence at the point of care, a whitepaper on the approach, and an argument that evaluation must go beyond benchmark wins. That is a real contribution and it advances the field.

What is absent is every number, including the measured output of the framework the company itself designed: no model class or version, no evaluation results, no hallucination or error rate, no benchmark performance. A vendor that defines the right way to assess a product and then does not report how its own product scores has conceded the standard while withholding the result, and a buyer should ask for that specific output rather than a generic accuracy figure, because the company has already agreed it is the right question.

The mechanisms in the product are genuine and carry the grade: every answer has inline citations tied to specific statements with single click access to the source topic, a side panel exposing rationale and supporting material, and a stated assumptions block. Publishing the assumptions an answer rests on is rare and useful, since a clinician can see when the system has assumed a patient characteristic that does not hold. No warranty, indemnity or remediation commitment was located. Ask for the framework results.

Epic integration is named, demonstrated at Epic expert group meetings, and covers both clinical decision support and patient education leaflets delivered at the point of care, with a dedicated EHR integration page and alignment across the UpToDate and Lexidrug content sets. Patient education material is offered in up to nineteen languages, which is a real interoperability property in a multilingual health system and one almost nothing else in this index provides.

Held at B rather than A on specifics: no FHIR or SMART on FHIR detail was retrieved, no EHR beyond Epic is named in the material read, and no marketplace or Epic Toolbox credential was located, the latter mattering more since the Connexall record established Toolbox designation as the stronger integration credential.

The largest installed base of anything in this segment by a wide margin, with a reported three million clinician users across thousands of hospitals and localised sites in more than thirty countries. Delivery spans web, dedicated mobile applications for both the core product and the generative layer, personal subscriptions bought directly, and health system deployment through the enterprise edition. Named reference: St. Luke's University Health Network, quoted through its associate chief medical information officer.

Two things hold it at B, and both are credited for being stated rather than hidden.

Availability of the generative product is geographically bounded. It is included in all United States personal subscriptions, but only in the higher tier and trainee subscriptions in Canada, and it reaches health systems only through selected enterprise accounts. Global reach of the reference product is therefore not global reach of the AI, and a buyer outside those markets should confirm what they are actually licensing.

And no data residency statement of any kind was retrieved despite operations across more than thirty countries, which for a European headquartered company selling into jurisdictions with strict transfer rules is a gap a buyer should close directly.

The most open commercial posture in this segment and one of the most open in the index, because part of this product can simply be bought. A public webstore operates with a pricing route, personal and trainee subscriptions are purchasable directly, and trainee rates are described as specially reduced against proof of status. A clinician can establish the cost of the AI enabled product without speaking to anyone, which is not true of a single other vendor graded in this segment.

The scope of what is included is also stated plainly rather than left to a sales conversation: the generative layer is included in all United States personal subscriptions, and in the higher tier and trainee subscriptions in Canada.

Two reasons it is B and not A. The actual figures were not retrieved in this pass, so the openness is established structurally rather than by a captured rate. And enterprise pricing, which is what a health system actually buys and the only tier that reaches most clinicians, is entirely gated behind a sales conversation. The transparent tier is therefore the one that matters least to the institutional buyer this index serves.

Breadth is exceptional and explicitly enumerated rather than implied: physicians, physician assistants, nurse practitioners and advanced practice nurses, pharmacists, dentists, residents, fellows and students, and medical librarians, across health plans and payers, providers, hospital pharmacies, life sciences, virtual care technology companies and retail pharmacy, with localised sites in more than thirty countries. The trainee and medical university segments are substantial enough, alongside continuing medical education accrual inside the AI product, to justify the workforce and training cross listing.

Two negative scope statements are credited, because this index asks for them and rarely gets them. The company states that the generative product is available to individual subscribers only in the United States and Canada, and that continuing education credits earned inside it are not yet recognised by certain accrediting authorities. Publishing what is not covered, and specifically what an earned credit will not yet count for, is the kind of disclosure that costs a vendor something.

Held at B because coverage is asserted across every setting while no performance or accuracy evidence is published for any of them. Breadth of availability is established; breadth of demonstrated usefulness is not.