DynaMed

The moat is the dataset, and this is that precedent at close to full strength. The asset is a curated evidence corpus built over decades under a stated seven step methodology, maintained by a named clinical editorial team with daily literature surveillance. Remove Dyna AI and a complete, saleable and award winning product remains, which is exactly what it was for years before July 2024.

The grade describes the mechanism, not the quality. Dyna AI is a real shipped generative product rather than a label: retrieval augmented generation over the company's own content sets, extended in February 2026 into a dedicated interface mode rather than a search box overlay. It is nonetheless sold as an add on to a content subscription, and the company's own positioning argues that the content, the editorial process and the human review are what distinguish it, with the model as the delivery mechanism.

That framing is honest and it is the reason for a C rather than a criticism of the product. Compare UpToDate, graded identically for the same structural reason, and contrast the retrieval first startups in this category where the model and the retrieval stack are the entire company.

Several real oversight affordances, and one claim that needs resolving.

On the credit side: generative output is stated to be clearly labelled as such, which is a published principle rather than a feature note; answers link to the source they came from; the corpus is bounded to the company's own curated content with an explicit statement that the tool never pulls from external unvetted sources; and Dyna AI Mode is designed so a clinician chooses between AI synthesis and classic topic search and can move between them, rather than having AI imposed on every query. The company's own framing criticises AI bolted onto existing clinical tools.

One mechanism is genuinely new to this category and worth naming as a benchmark. Dyna AI Mode captures the question, the response, the references and a timestamp in copy ready form, and lets a clinician re run an earlier query. That preserves what the tool said and when, in a form that can go into a record. Every other vendor in this segment leaves the clinician to reconstruct that after the fact. New standing ask: can a clinician preserve the evidence as it stood at the moment of care.

Held at B on an ambiguity in the vendor's own marketing. Promotional material states that every response is validated by a clinician and that there is always an expert clinician in the loop. For a real time retrieval augmented system that is implausible read literally, and is far more likely to describe review at the content curation stage than validation of each generated answer. Where the human actually sits is the single most consequential fact about an oversight model, and it should not require interpretation. Ask it directly. Nothing published states what the system does when the evidence is thin or in conflict.

The architecture is named rather than implied. The company states a retrieval augmented generation framework, bounds the corpus explicitly to its own DynaMed, DynaMedex and Dynamic Health content sets, states that generative use for user facing applications is limited to information found in that curated content, indicates the origin of each answer with direct source links, and labels AI driven output. It also publishes the composition of the team behind the system, naming physicians, advanced practice providers, pharmacists, nurses and search engineers on a clinical prompt engineering team.

Below that it stops. No model class or provider is named, no evaluation harness is described, and the seven step evidence methodology that governs the underlying content is not mapped onto how the generative layer applies it.

One contrast is worth recording because it runs opposite to the rest of this category. This vendor publishes no benchmark figures at all, where two of the newer entrants publish headline percentages against named competitors with no methodology attached. Publishing nothing is not transparency, but it is a different failure from publishing a number nobody can reproduce, and a buyer comparing across the category should not read the silence as the weaker position without asking for the internal results.

Two products answer this question differently and separating them is the right analysis. The generative layer reads a curated content corpus rather than a patient record, so it carries no patient data at all, which is the strongest of the structural answers this index tracks: there is nothing for a chain to leak because nothing patient identifying enters it.

A separate decision support module does take patient data, through a standards based record integration described as compliant, to pre populate decision aids and calculators, and that module is where every stewardship question actually lands. A published principle addresses security and patient privacy directly and states that systems are designed and monitored according to established safety principles, with governance including continuous monitoring by clinical experts.

Held at C on what was not retrieved for the module that matters. No retention period, data residency statement, de identification practice or sub processor detail was located, no model or provider is named for the generative layer, and the corporate privacy policy and terms of use were not opened this pass, which is recorded explicitly rather than left implicit. Ask which of the two products your deployment includes, and for the integrating module ask what patient fields are drawn, what is retained, where processing occurs, and for a sub processor list.

Revised upward on 2 August after a peer reviewed study surfaced during the Micromedex build. Recording the correction rather than quietly amending it, because the first pass under found and the reason is instructive: the study is indexed under the joint product's earlier name, not under DynaMed.

A peer reviewed usability evaluation of the combined drug and disease product exists, published in 2023, with authors from Brigham and Women's Hospital, Mass General Brigham, Harvard Medical School and a college of pharmacy alongside authors from the partner vendor. It observed clinicians using the tool, collected structured feedback, and published what needed improvement, including navigation difficulty and gaps in comprehensiveness. A peer reviewed study of an actual product in this category is rare, and one that publishes its own product's shortcomings is rarer.

Alongside it, recognition in the Best in KLAS report is extensive: DynaMed named best in category for point of care disease reference in 2021, 2022, 2024 and 2025, and the combined product named top performing in 2026. Be precise about what that measures. It is a structured survey of customer experience across many provider organisations, real and hard to buy, and it is not a measurement of whether the generative layer produces correct answers.

Held at B rather than A for three stated reasons. The peer reviewed study measured usability rather than correctness. It predates Dyna AI by roughly a year and evaluated the earlier assistant, so it does not speak to the current generative layer. And the vendor co authored it.

No evaluation of Dyna AI's own accuracy, error rate or failure modes was located. The company describes rigorous qualitative and quantitative research and extensive user testing during development without publishing any of it, which remains the gap. The vendors in this category carrying the strongest evidence positions got there by submitting to evaluations they did not run.

The architecture is favourable and it is worth separating the two products, because they answer the stewardship question differently. Dyna AI reads a curated content corpus rather than a patient record, so the generative layer carries no patient data at all, which is the strongest of the three architectural answers this index tracks. DynaMed Decisions is the separate module that does take patient data, through SMART on FHIR integration the company describes as HIPAA compliant, to pre populate decision aids and calculators.

A published principle addresses security and patient privacy directly, stating that data are protected using best practices in data security in accordance with HIPAA standards, and that systems are designed and monitored according to established safety principles in AI. Governance of the AI is stated to include continuous monitoring for quality and usability by clinical experts.

Held at B on what was not retrieved. No retention period, no data residency statement, no de identification practice and no subprocessor detail were located, and the corporate privacy policy and terms of use were not opened this pass. Scope recorded explicitly rather than left implicit.

An explicit assertion exists and it is repeated consistently across product announcements, which is more than several peers in this category manage. DynaMed Decisions is described as a HIPAA compliant clinical decision support solution, its EHR integration is described as HIPAA compliant, and a published principle commits to protecting data in accordance with HIPAA standards.

The assertion is where it stops. No business associate agreement was located, no terms were published, no covered entity or business associate language appears, no execution path is described and no tier at which a BAA becomes available is stated. The corporate privacy policy and terms of use were not opened this pass, and both are places such a statement could sit, so this is graded on what a buyer can find rather than as a finding that nothing exists.

The contrast within the segment is stark and useful. Vera Health, a company a fraction of this size, publishes the full text of its business associate agreement at a public address. Establish before contracting whether a BAA is available at the subscription tier being bought, or only for the enterprise EHR integrated module.

Two retrieval passes located no SOC 2 report of either type, no ISO 27001, no HITRUST certification, no trust centre, no penetration testing cadence and no vulnerability disclosure programme for EBSCO Clinical Decisions or for DynaMed specifically. What exists is the security limb of the published responsible AI principles, which commits in general terms to best practices in data security in accordance with HIPAA standards.

This is recorded as a retrieval outcome rather than as a finding that controls are absent. A division of a large private information services company selling into hospitals and health systems almost certainly holds attestations, and the absence here is about publication rather than about practice. Flagged for refresh, and the specific thing to ask for is the SOC 2 report and its type.

The comparison that makes the gap conspicuous is within this category rather than outside it. C8 Health, a startup, publishes SOC 2 Type II with the type specified and runs a public trust centre. Vera Health publishes four compliance marks and a trust centre on its homepage. Scale is not the reason for the difference.

No clearance, no submission and no device positioning statement were located, which is unsurprising for a reference product and is not by itself a mark against it.

The position on the clinical decision support exclusion is stronger here than for most of this category, and the vendor has built towards it deliberately even if it does not say so in regulatory language. The exclusion turns on whether a clinician can independently review the basis of what the software presents. A bounded corpus the vendor names, output labelled as AI generated, and direct links from each answer to its source together make the basis reviewable in a way an unbounded generative assistant does not. The February 2026 addition of timestamped, copy ready capture of question, answer and references strengthens that further.

Against it, DynaMed Decisions takes patient specific data through the Smart on Fhir standard and returns personalised guidance, which is a materially different act from displaying what the evidence says, and the two products are sold under one brand. Graded C because the vendor publishes no analysis of where either product sits. A claim that the product aligns with the federal algorithm transparency requirements for certified health information technology appears on a third party review directory and could not be confirmed on the vendor's own material, so it is recorded here as unverified and is not graded on.

The most developed governance apparatus in this category to date, and still no measurement.

Five principles for the responsible use of AI are published with substance rather than as headings: quality, limiting generative use for user facing applications to curated content; security and patient privacy, referencing HIPAA standards and established AI safety principles; transparency, requiring generative uses to be clearly labelled and evidence sources shown; governance, placing clinical experts over development, validation and continuous monitoring; and equity. The company also commits publicly to ongoing review for bias, quality, safety, ethics, regulatory considerations and scientific rigour, and states plainly that applying generative AI to clinical diagnosis and treatment is experimental in nature, which is an unusually careful thing for a vendor to write down.

Two structures back the principles. An external expert advisory council selected for unbiased representation includes not only clinicians but legal counsel and data security and privacy experts, which is a broader composition than the clinician only councils elsewhere in this segment. And the company joined the Coalition for Health AI in September 2024, the first third party AI governance affiliation recorded in this category.

Held at B because none of it produces a published result. No bias evaluation, no subgroup analysis, no error taxonomy, no accuracy floor and no external audit output. Equity is named as a principle and nothing is published against it. The apparatus is real; the measurement is absent, which is the same ceiling that holds UpToDate at B for a differently shaped version of the same posture.

The constraint on what this system can say is architectural and stated, which is the substantive disclosure here. The framework is named as retrieval augmented generation, the corpus is bounded explicitly to the company's own named content sets, and generative use for user facing applications is stated to be limited to information found in that curated content.

A bounded corpus is a real limitation on the failure mode: a system that can only surface what is in a curated library cannot invent a guideline that does not exist, though it can still misapply one that does. Each answer indicates its origin with direct source links and generated output is labelled as such, so a clinician knows which text was produced rather than quoted. The team composition behind the system is published, naming the clinical and engineering roles involved.

Held at C because nothing is measured and one mapping is missing: the multi step evidence methodology governing the underlying content is not mapped onto how the generative layer applies it, so a buyer cannot tell whether a generated answer inherits the evidence grading of its source or flattens it. One contrast belongs on the record because it runs against the rest of this category.

This vendor publishes no benchmark figures at all, where two newer entrants publish headline percentages against named competitors with no methodology attached. Publishing nothing is not transparency, and it is a different failure from publishing a number nobody can reproduce. Ask for the internal results.

The deepest integration position in this category, and it clears the marketplace credential test that this index treats as the hard evidence rather than the claim.

DynaMed Decisions was listed in the Epic App Orchard marketplace in June 2021, which is a credential Epic issues rather than a statement the vendor makes about itself. The company names SMART on FHIR and HL7 standards explicitly as the mechanism, describes the integration connecting to a patient's record to gather the data a decision aid needs, and describes that integration as HIPAA compliant. Named integration methods across the product line include embedded links, search box widgets, the HL7 Infobutton context retrieval API and a web services API, alongside single sign on and a mobile application.

Infobutton deserves specific credit. It is the interoperability standard built for exactly this job, surfacing reference content in context from inside the record, and a vendor supporting it is meeting the EHR on the standard rather than asking for a bespoke build.

Two honest limits. The App Orchard listing dates from 2021 and Epic has since renamed that marketplace, so confirm the current listing rather than relying on the announcement. And broader lists of named EHRs appear on third party directories rather than on the vendor's own material, so they are recorded as context and not graded on.

Delivery is cloud and web with a mobile application, single sign on, institutional affiliation across multiple organisations for a single clinician, and EHR embedded access. Implementation for institutional subscriptions runs through a named professional services team.

What earns the grade is that the geographic boundaries of the AI are stated plainly and negatively, which this index credits because vendors so rarely publish what they do not cover. Dyna AI is stated to be currently unavailable in the European Union, with a separate page maintained for European audiences, and DynaMed with Dyna AI for individual subscribers is stated to be available only in the United States.

Those two statements are worth reading together with the rest of the category, because a pattern is now visible. One vendor withdrew from the European Union and United Kingdom entirely rather than meet emerging AI requirements, one bounds its generative layer to two North American markets, and this one geofences its generative layer out of the European Union while continuing to sell the underlying reference product there. The generative layer is being treated as the regulated component and the content beneath it is not.

Held at B on the absence of any hosting, region or residency statement, which for a product sold internationally is the obvious question to close.

A public webstore exists at the product's own domain with a subscribe route and a free trial, so an individual clinician can begin a purchase without speaking to anyone. Subscription tiers for individuals and for trainees are sold directly, and the store carries a plain statement of what is not included, namely that DynaMed with Dyna AI for individual subscribers is currently available only in the United States. Publishing a limitation at the point of sale costs a vendor something and is credited.

Two reasons for B rather than A. Exact figures were not captured this pass because the store renders through client side script, so the price points on the record are unverified and should be pulled on refresh; figures circulating on third party directories are not graded on. And institutional pricing, which is the tier a health system actually buys and which governs Dyna AI as an add on, is entirely gated behind a sales representative with no mechanism, rate card or implementation fee published.

That is the same asymmetry recorded against UpToDate, and it is structural across this category: the transparent tier is the one that matters least to the institutional buyer.

Coverage across the product family is broad and the segments are distinct rather than overlapping: DynaMed for disease and condition reference, Dynamic Health aimed at nursing and allied health, and DynaMedex adding Micromedex drug content for clinical care teams and pharmacy. Continuing medical education and maintenance of certification credit accrue from usage and are tracked for the clinician, and a single user can affiliate with multiple organisations for access across sites.

On continuing education this is the third vendor in this category to attach credit to ordinary point of care use, after UpToDate and Vera Health, which makes it a category feature rather than any one vendor's differentiator.

Held at B for a reason specific to how this record is scoped. The reference product's coverage and the AI's coverage are not the same, and the gap is published: Dyna AI is unavailable in the European Union and, for individual subscribers, available only in the United States. A buyer outside those markets is licensing the content product, not the graded AI. No performance or usefulness evidence is published for any specialty.