UpToDate

The value of this product is the corpus, and the company frames it that way itself. UpToDate Expert AI is described as the generative AI powered evolution of the reference product, built solely on the existing curated content: more than thirty years of topics maintained by over 7,600 physician authors and editors. Remove the generative layer and what remains is the clinical reference product that was sold without it until September 2025.

The work around the model is substantial: an assumptions and rationale architecture, a published validation framework, and, per the company's own release of 3 June 2026, OpenAI models deployed inside a model agnostic enablement platform it calls FAB. None of that changes how little the product depends on the model. A customer who switched the generative layer off would still hold almost all of what they paid for.

Adoption of the layer is moving quickly. The same release states that as of 30 April 2026 more than half of United States Enterprise Edition customers, approximately 2,000 hospitals, had signed up for the Expert AI version, with the company expecting 70 percent by mid year. That is uptake of a feature, and it matters because a feature most customers switch on can become the product over time.

One part of an unusually detailed oversight design is missing: the threshold at which the system stops.

Two mechanisms carry it. The assumptions block opens every response by naming what the system took for granted about the patient, in the company's own illustration that the patient is not pregnant, not a child and has no major comorbidities, so a clinician can invalidate an answer in one glance rather than reading to the end to find it answered a different question. The nudges are inline prompts asking the clinician to supply context the system lacks, such as comorbidities or intensive care status. Rather than abstaining or emitting a confidence score, the system names the gap and asks the user to close it, which is a real answer to what the system does when it does not know. Inline citations tied to specific statements and a rationale panel sit alongside both.

What is not published is the other side of the boundary: the conditions under which the system declines to answer, escalates, or routes a case away from itself, and a statement of what it may not be used for. A nudge asks for more context; nothing describes what happens when the context never arrives.

Trade coverage of the launch reported the company as saying that closing the system to the broader web eliminates the chance of hallucination. That wording is not the company's published language, which speaks of grounding and traceability. If it is said in a sales setting it is an overclaim, because grounding constrains the source material without removing synthesis error.

The supplier and the platform are named and the version discipline is not. The company's release of 3 June 2026, published as a PDF on the Euronext company news service, states that UpToDate Expert AI was built working closely with OpenAI, that moving to the latest OpenAI models improved cost, responsiveness and speed and significantly increased clinical relevance, and that OpenAI technologies are deployed within Wolters Kluwer's own secure, model agnostic generative AI enablement platform, which it calls FAB. The same release describes the product as supported by a multi layer expert in the loop framework.

Provenance and reasoning are exposed in the product itself. 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 has also published a validation framework for evaluating clinical AI at the point of care.

No specific model or version is identified, latest models is not a version, no update practice or change notice is published, and the model-agnostic platform means the model can change underneath the product without any published record of when. The significant increase in clinical relevance is stated with no figure, method or denominator.

There is no word on which model answers a given account today, how a model change is tested before release, or how customers are told when it happens.

The model provider is named for this product, and nothing else in the chain is. The company's release of 3 June 2026 states that UpToDate Expert AI runs on OpenAI models, deployed within Wolters Kluwer's own model-agnostic enablement platform, FAB, while staying within the company's existing enterprise infrastructure and adhering to customer governance requirements.

That is the model layer and the platform named, and it is where the disclosure stops. No hosting provider, no subprocessor list, and no statement of what OpenAI receives, whether it retains anything, or whether any terms bar training on what it receives. The phrase staying within existing enterprise infrastructure is the company's description of its own architecture, not a published data flow that can be checked.

The interface makes the missing terms matter. Expert AI invites clinicians to describe cases in natural language and prompts them for comorbidities and intensive care status, and it is reachable from inside the record system where patient detail is already on screen.

Because the same product is sold both as a personal subscription bought with a card and as an enterprise deployment under a health system agreement, a clinician who bought their own license and types a case into it is contracting personally while handling their employer's patient data.

What the model provider receives and retains on each route is not published, and neither is what an institution can require of individually purchased licenses.

The published evidence is for the reference product, it is fourteen years old, and nothing independent evaluates the generative layer. All three halves of that sentence matter.

The peer reviewed study is Isaac, Zheng and Jha, Use of UpToDate and outcomes in US hospitals, Journal of Hospital Medicine, February 2012 (doi 10.1002/jhm.944), from Beth Israel Deaconess Medical Center. It is a retrospective study across a national sample of US inpatient hospitals using fee for service Medicare beneficiaries, comparing risk adjusted length of stay, mortality and quality performance between hospitals with and without UpToDate. Patients in UpToDate hospitals had shorter stays, 5.6 days against 5.7 overall and 0.1 to 0.3 days shorter across six prespecified conditions, lower risk adjusted mortality for three of the six, by 0.1 to 0.6 percentage points, and better quality performance on every Hospital Quality Alliance condition. The benefit concentrated in smaller and non teaching hospitals and was small or nonexistent in larger teaching institutions. The authors call it a very small but consistent association. It is independent and multi site, and it is retrospective and associational where the outcome claim it supports would need a prospective design.

For UpToDate Expert AI there is no peer reviewed evaluation of accuracy, no error rate and no published output of the validation framework the company designed. What exists is vendor produced: whitepapers, the framework itself, an analyst report the company promotes as independent validation, physician surveys, and named customers, St. Luke's University Health Network and the University of Texas Medical Branch. The release of 3 June 2026 reports that about 2,000 hospitals had signed up for the Expert AI version by 30 April 2026, which is adoption rather than benefit.

PubMed lists no evaluation of UpToDate Expert AI.

The position published is a scope statement rather than an answer to the questions a generative product raises. The terms of use state that UpToDate and its licensors are not intended to receive, transmit, store or process protected health information under HIPAA, which places the product outside the patient data relationship by contract. That is a clear position on protected health information and it says nothing about what happens to the free text a clinician types.

The product specific question is live. Expert AI is a conversational interface that invites clinicians to describe cases in natural language and prompts them to add comorbidities and intensive care status, so clinical detail about identifiable patients will predictably be typed into it, and the product is reachable from inside the electronic health record where that detail is already on screen. The company's release of 3 June 2026 confirms that OpenAI models run the generative layer inside Wolters Kluwer's own enablement platform and existing enterprise infrastructure. Nothing published states whether queries are retained, for how long, whether they are used to improve the service or the underlying models, or what reaches the model provider.

The privacy and cookies policy was out of reach and may say more about retention.

The same product is sold as a personal subscription bought with a card and as an enterprise deployment under a health system agreement, and only the second puts an institution in a position to negotiate terms.

Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

UpToDate takes an explicitly published no protected health information position. 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. The position is stated plainly in a document that governs the service, so a customer knows not to enter patient identifiers and not to expect business associate obligations.

A disclaimer allocates risk rather than preventing the problem, and the product design makes the problem foreseeable. UpToDate is embedded in electronic health records inside the clinical workflow, and Expert AI invites clinicians to ask questions in free text and prompts them for comorbidities and intensive care status. A clinician working a live case is 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 logs persist, or who can read them. The company's release of 3 June 2026 confirms that OpenAI models now answer those questions inside Wolters Kluwer's own infrastructure, which makes the question of what the model layer receives a live one.

The privacy and cookies policy, which was unreachable, may cover query logs.

Whether a business associate agreement can be obtained for integrated or enterprise deployments, despite the standard terms disclaiming the relationship, is not said.

Nothing published attests to the security of UpToDate itself. The product's policies and legal section publishes terms of use, a privacy and cookie notice and jurisdiction specific privacy addenda, and nothing describing security controls, independent audits or certifications for the service.

The same parent publishes assurance elsewhere, and where its boundary falls matters. Wolters Kluwer maintains dedicated privacy and security certification pages for products in its other divisions, TeamMate and CCH Integrator among them, naming ISO 27001, ISO 27701 and SOC 2 Type 2, and one of those pages describes a group level security program aligned to ISO 27001 and the NIST Cybersecurity Framework under executive oversight. Those certifications carry their own scope and it is not this product. A program described as aligned to a framework is an assertion rather than an outside check. An attestation whose boundary excludes the service being bought says nothing about that service.

None of trust.wolterskluwer.com, trust.uptodate.com or security.uptodate.com resolves, and no UpToDate SOC 2 or ISO 27001 report or certificate is published; material on those standards covers only the other divisions. Assurance for a product of this size usually exists somewhere, and the Wolters Kluwer pages that stayed closed may hold it.

Expert AI and the FAB platform it runs on may also sit inside a different boundary from the reference content.

No clearance, none claimed, and the product is scoped as a clinical reference. What is missing is a stated regulatory rationale, not a clearance.

This product illustrates the clinical decision support exclusion test well. The 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 about the patient, and follow the reasoning, which is the inverse of an embedded predictive score whose basis is disclosed to nobody.

The direction of travel is the open question. The company's release of 3 June 2026 describes a pipeline of features equipped with domain grounded generative and agentic AI built with OpenAI. An agent that acts is a different regulatory object from a reference that answers, and nothing published says how the company's position is maintained as the product moves that way.

Named process stands behind the governance material. The company publishes a corporate AI principles statement, a validation framework for evaluating clinical AI at the point of care released as a public document, and 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. The release of 3 June 2026 describes UpToDate Expert AI as supported by a multi layer expert in the loop framework and governed under the company's Responsible AI Principles. The company also publishes survey research on unsanctioned AI use inside healthcare organizations, which is self interested and useful.

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 is published.

A structural exposure sits underneath and is 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 generated answer. Grounding to a curated corpus inherits the corpus's blind spots as faithfully as it inherits its rigor, and the validation framework could measure that but reports no such measurement.

The strength here lies in the product's mechanisms, and nothing stands behind the output. 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 naming what the system took for granted about the patient. Publishing the premises an answer rests on is useful, because a clinician can see at once when the system has assumed something that does not hold for the patient in front of them.

The company has also authored the test and not published its score. It publishes a validation framework for evaluating clinical AI at the point of care and an argument that evaluation must go beyond benchmark wins, and it publishes no result of applying that framework to its own product: no evaluation output, no error or hallucination rate, no benchmark performance. The release of 3 June 2026 states that moving to newer OpenAI models significantly increased clinical relevance, with no figure, method or denominator behind the statement.

The company publishes no warranty, indemnity or remediation commitment attached to the output, though the terms of use, inaccessible for now, may hold one.

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. The company's release of 3 June 2026 states that more than half of United States Enterprise Edition customers, approximately 2,000 hospitals, had signed up for the Expert AI version by 30 April 2026, so the generative layer is reaching the chart through the same route.

No FHIR or SMART on FHIR detail, no record system beyond Epic, and no marketplace listing or Epic Toolbox designation is published, and nothing describes whether the integration reads patient context, writes back, or both.

There is also no technical description or credential for the integration, and no word on whether Expert AI receives patient context from the chart when launched from inside it.

Delivery options are published and where the system runs is not. The product is delivered on the web, through dedicated mobile applications for the reference product and the generative layer, as personal subscriptions bought directly, and to health systems through UpToDate Enterprise Edition embedded in the record. Use is reported at more than three million clinicians with localized sites in more than thirty countries. Named reference: St. Luke's University Health Network.

Residency is the weak point. No hosting provider is named, no region is stated, and nothing describes where queries are processed or where data rests. The company's release of 3 June 2026 adds one fact and it points the other way from a commitment: the OpenAI models behind Expert AI run inside Wolters Kluwer's own model-agnostic enablement platform and existing enterprise infrastructure, which says the processing stays within the company's estate and does not say where that estate is. For a company headquartered in the Netherlands and selling across more than thirty countries, location is implied by the corporate structure rather than committed.

Availability of the generative product is bounded and stated: Expert AI is included in United States personal subscriptions, in Pro Plus and trainee subscriptions in Canada, and in Enterprise Edition accounts. Reach of the reference product is not reach of the AI.

Whether processing differs by country, and which party hosts the enablement platform, is not stated.

Part of this product can be bought without a conversation, and the pricing basis and inclusion scope are published. A public webstore operates at store.uptodate.com with a see pricing route, personal and trainee subscriptions are sold directly, and trainee rates are described as reduced against proof of status. What is included is stated plainly: Expert AI comes with all United States personal subscriptions and with the Pro Plus and trainee subscriptions in Canada, and reaches health systems through UpToDate Enterprise Edition.

The figures themselves go unstated. The store is a Salesforce commerce application that renders its catalog in the browser, and the prices may sit there. The openness is established by structure rather than by a rate.

Enterprise pricing, which is what a health system buys and the tier that reaches most clinicians, is gated behind a sales conversation. The transparent tier is the one that matters least to the institutional buyer.

Nothing published says whether Expert AI is included in an Enterprise Edition license or priced as a module, whether pricing is per named clinician, per concurrent user or per institution, whether Lexidrug is bundled, or whether an institutional agreement changes the data handling terms that apply to questions typed into Expert AI compared with a personal subscription bought with a card.

Coverage is 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 localized sites in more than thirty countries. The trainee and medical university segments are substantial, and continuing medical education accrues inside the AI product, so the product serves workforce and training as well.

Two negative scope statements are published. The generative product is available to individual subscribers only in the United States and Canada, and continuing education credits earned inside it are not yet recognized by certain accrediting authorities. Publishing what such a credit will not count for stands out.

Validation reaches part of the coverage and not the generative product. A peer reviewed retrospective study of US hospitals (Journal of Hospital Medicine, 2012, doi 10.1002/jhm.944) associated use of the reference product with slightly shorter stays and lower mortality for some conditions, concentrated in smaller and non teaching hospitals and small or absent in large teaching institutions. That is inpatient validation for the reference content, fourteen years old. No performance evidence is published for Expert AI in any setting or specialty.