Clinical Reference & Evidence

AI that answers a clinical question from the published evidence rather than from the patient in front of the clinician: literature search and synthesis, guideline and drug reference, and retrieval over an institution's own approved knowledge base, delivered at the point of care. The category separates from clinical decision support on a single testable property. A product here answers what the evidence says and takes little or no patient context to do it, which is why several vendors in this category process no protected health information at all. A decision support product answers what to do for this patient, takes the chart as input, and inherits the evidentiary and regulatory obligations that a patient specific output carries. Products that genuinely do both are graded on their reference half and cross listed. Two failure modes define the category, and the accuracy percentages vendors publish address neither of them well. The first is the fabricated or misattributed citation, which is conspicuous once someone checks it and which the serious vendors here have largely engineered out. The second is the one to ask about: an interpretive error sitting underneath an accurate citation. Verifiable provenance creates an appearance of checkability that the synthesis layer has not earned, and a clinician who spot checks the reference rather than the reasoning will not catch it. A third exposure is structural rather than technical. Grounding an answer in a curated corpus inherits that corpus's blind spots as faithfully as it inherits its rigour, and a superseded entry is wrong for every clinician it reaches until someone notices, with no drift monitoring concept in this segment to catch it. Buyers should establish what the corpus actually contains and who maintains it, what the review cycle is and what detects a stale entry, whether the system states its own assumptions and tells the clinician what it does not know, and how the product is funded, since advertising and industry funded clinical content are both live commercial models here and only some vendors publish a policy governing them.