ASCO Guidelines Assistant

The moat is the dataset in its purest form anywhere in this category. The asset is a guideline library built over decades through systematic review and expert panel process, and it exists, is published, and is widely used independently of any artificial intelligence. Strip the assistant away and the guidelines remain exactly as authoritative as before.

The generative layer is a chat interface over that library, and it is built on a third party platform using a third party model family rather than on anything the society developed. That is disclosed openly, which is credited elsewhere on this record, but it is also the reason this axis sits where it does.

Graded C on mechanism rather than on quality, consistent with UpToDate, DynaMed and ClinicalKey AI, all of which are content organisations that added a model. The distinction worth holding onto is that for those three the content is a commercial product; here the content is a professional standard that the society publishes as part of its mission, which changes what the artificial intelligence is being asked to do but not how central it is.

The published disclaimer is the most precise scope statement in this category and it deserves to be quoted in substance. The assistant is described as a voluntary educational resource that does not account for variation among patients and does not substitute for the independent judgment of a licensed healthcare professional, with a dedicated page carrying the full text.

That middle clause is the valuable one. This index has spent the category asking whether generative clinical products state what they assumed about the patient. Here the answer is that the tool does not individualise at all and says so at the point of use. Guidelines describe populations; this tool retrieves guidelines; a clinician is told plainly that the gap between the two is theirs to close. Very few vendors state their own limitation that exactly.

Supporting it: every response carries citations and sources, the corpus is bounded to one named body of guidance, and the update cadence is tied to guideline publication rather than left vague.

Held at B on the question a bounded single society corpus makes unavoidable. Oncology practice regularly runs ahead of or outside published guidance, and nothing states what the assistant does when the guidelines do not cover the question asked, or when two guidelines address the same situation differently. Silence on the out of scope case is the gap here, and it is more consequential for a narrow corpus than for a broad one.

This record does something no other vendor graded in this category does: it names the model family and the platform provider. The assistant is stated to be built on Google Cloud's Vertex platform using Gemini models, said plainly in the launch announcement, on the product page and in the society's own publications.

That is worth stating as a category level observation rather than a footnote. Across this category the index has recorded vendor after vendor describing retrieval architectures, evidence grading logic and curated corpora while declining to say which model generates the text. A professional society disclosed it immediately, and the platform provider was quoted by name doing so.

The corpus is also unusually well specified, because it is a public artefact: the tool draws solely from published guidelines that anyone can read, so a clinician can verify not just a citation but the whole scope of what the system could possibly have drawn on.

Held at B because the disclosure stops at the name. No model version, no retrieval or ranking method, no evaluation harness, no accuracy figure and no description of how conflicting or superseded guidance is handled were located.

No evaluation of the assistant was located in two retrieval passes: no accuracy measurement, no error rate, no clinician review study, no benchmark result and no published pilot. The tool was demonstrated and tested during the society's 2025 annual meeting, which is exposure rather than measurement.

The distinction that matters here is the same one recorded against DynaMed and it is worth restating because it will recur throughout this category. The underlying content carries an exceptionally strong evidence process: systematic reviews, expert panels, formal guideline methodology. None of that is evidence about the retrieval and synthesis layer built on top of it. A guideline being rigorous does not establish that a model answering questions about it returns the right passage, or summarises it without distortion.

That gap is more pointed here than elsewhere, because the corpus is the one place where evaluation should be cheapest. The guidelines are public, finite and structured, the correct answer to a guideline question is usually determinate, and the society has the panel members who wrote them. A vendor better placed to run and publish an evaluation is hard to imagine.

The architecture requires no patient data. Users ask questions about guidelines, the system answers from published guidance, and nothing in the described workflow needs a record. Access runs through a society account login, so users are identified but patients are not involved.

One mechanism is disclosed and it cuts two ways. User feedback is collected and stated to refine both the tool and the development of guidelines over time. As a quality loop that is legitimate and unusually candid. It also means clinician queries and feedback are retained and used, and nothing published states for how long, in what form, or whether query content is separable from the feedback attached to it.

No retention period, data residency statement, encryption practice or query handling policy specific to the assistant was located, and the society's general privacy policy was identified but not opened this pass. Recorded explicitly rather than left implicit.

The practical question for a clinician: a question typed in a hurry about a real case may contain clinical detail even when the tool does not ask for it, and nothing at the point of entry addresses that.

No statement of any kind was located. No business associate agreement, no covered entity or business associate language, no reference to the federal health privacy rules in the assistant's own material, and no execution path.

The architectural case for that being acceptable is stronger here than for most of this category. The product is positioned as a voluntary educational resource, access is by individual professional membership rather than institutional licence, and the system answers from published guidance rather than from a record, so a business associate relationship would not ordinarily arise.

Graded C rather than lower for that reason, and recorded as a retrieval outcome rather than a finding. The society's general privacy policy and terms of use were identified but not opened this pass and are where any relevant statement would sit.

The question worth asking is institutional rather than individual: if a practice or cancer centre wants to make the assistant part of a documented workflow rather than leaving it to individual member access, there is currently no published instrument to support that, and the disclaimer positioning suggests the society may not intend one.

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 specific to the assistant.

One real disclosure sits on the credit side. The platform the tool runs on is named, which tells a buyer where processing happens and which provider's security posture is inherited. Most vendors in this category name no infrastructure at all, and one publishes its processor list only on request. Naming a hyperscale platform does not transfer that provider's certifications to this product, and the record should not imply it does, but it is more than a blank.

This is graded as a retrieval outcome. A professional society running member systems at this scale holds security documentation; none of it was published against this product. Access being membership gated also means the population with access is credentialed, which reduces some exposure without addressing any of the underlying controls.

The strongest regulatory positioning in this category, and it is achieved through a disclaimer rather than a filing.

No clearance exists and none is sought. What exists instead is an explicit, published statement of what the product is not: a voluntary educational resource that does not account for variation among patients and does not substitute for the independent judgment of a licensed healthcare professional, carried at the point of use with a dedicated page behind it. Every other vendor graded in this category publishes no regulatory positioning of any kind and was graded C for exactly that absence. This one publishes a position.

The substance also fits the exclusion criteria rather than merely gesturing at them. The output is retrieved published guidance with citations to documents any clinician can read in full, so the basis is independently reviewable, and the stated refusal to individualise to a patient goes to the heart of what separates a reference tool from patient specific decision support.

Held at B rather than A because no formal analysis is published, the disclaimer page itself was not opened this pass, and the product is described in the society's own launch material as delivering treatment recommendations, which is a looser characterisation than the disclaimer supports. Those two framings should be reconciled.

No artificial intelligence governance artefact was located: no published principles, no responsible use framework, no evaluation methodology, no bias assessment, no subgroup analysis and no error taxonomy.

Two adjacent structures are real and are recorded without being credited as artificial intelligence governance, because they govern the content rather than the model. The society operates a formal conflict of interest management function covering the panels that author its guidelines, which is meaningful assurance about the corpus. And the feedback loop is disclosed rather than hidden.

That feedback loop deserves closer attention than its one line description invites, because of where it terminates. Feedback given to the assistant is stated to inform not only the tool but the development of guidelines over time. That creates a path from what clinicians ask a chat interface back into the standard of care itself. It may well be a good path. Nothing published describes how that input is weighted, who reviews it, or how it is kept separate from the formal evidence review process that guideline development otherwise rests on.

Graded C because the artificial intelligence layer has no published governance of its own. The bias question specific to this product is untouched: guideline recommendations are shaped by the evidence base available to the panel, and a retrieval tool inherits every population gap in that evidence without surfacing it.

No integration exists and none is claimed. Two retrieval passes located no named electronic health record, no marketplace listing, no application programming interface, no conformance statement for a health data exchange standard and no context launch mechanism.

Delivery is a web application reached through the society's site and a member mobile application on both platforms. That is a coherent design for a membership benefit and an honest one, since nothing on the product page implies otherwise.

It is nonetheless the axis where the product's distribution model shows its cost. The stated purpose is helping community oncologists who carry large volumes across many cancer types make decisions quickly. Those clinicians work inside an oncology record system all day, and a tool they must leave that system to reach competes with the record's own content rather than complementing it. Two competitors in this category reach the clinician inside the record; this one does not.

Hosted service reached through the society's website at a dedicated address and through a member application on both mobile platforms, with access gated behind a society account.

The platform provider is named, which is a partial infrastructure disclosure and is credited, since it tells a buyer whose cloud the processing runs on. Beyond that nothing is stated: no region, no data residency position, no retention statement and no description of the deployment for an institution rather than an individual.

Geographic availability is the specific gap worth closing. The society has substantial international membership and its guidelines are used well beyond the United States, but nothing published states whether the assistant is available to members everywhere or is bounded to particular markets. Three other vendors in this category publish explicit geographic limits on their generative layers, in two cases excluding the European Union entirely, so the question is live rather than theoretical and the absence of a statement here should not be read as global availability.

The access condition is stated unambiguously and there is no separate charge. The assistant is available to society members, requires a society account to log in, and appears as a benefit inside the member application. No purchase route, upsell, paid tier or add on price exists or is implied.

That is a genuinely transparent commercial position and it compares well within this category. A clinician can establish exactly what they must do to obtain the tool without contacting anyone, which is not true of most vendors graded here, and two of the three reference incumbents gate the tier that matters behind a sales conversation.

Two honest limits hold it at B. The membership dues schedule was not retrieved this pass, so the actual cost of access is on the record as a condition rather than a figure; pull it on refresh. And there is no route at any price for a non member, so an oncologist outside the society or an institution wanting to license the tool for its staff has no published option at all. Access is clear, and it is also closed.

Single specialty by design, and the design is coherent. Coverage is the society's full guideline library across medical oncology, which is broad within that discipline and covers exactly the ground a community oncologist handling many cancer types needs. The stated target is community and rural practice, where volume is highest and subspecialty depth thinnest, and that is a well chosen problem.

Graded C because the axis measures coverage and this is one specialty in one body of guidance, with nothing outside oncology and no evidence published for performance within any tumour type.

The scoping limit worth naming is inside the specialty rather than outside it. Oncology practice draws on more than one society's guidance, on compendia, and on trial protocols, and treatment decisions frequently sit at the boundary between them. A tool bounded to a single society's library will answer confidently where that library speaks and is silent on how it behaves where the library does not, which is the same gap recorded on this vendor's oversight axis and matters more in oncology than it would in most fields.