Development of an exploratory taxonomy of veterinary AI query types and stage-specific query preferences: an expert panel study

Background/objectives The integration of large language model (LLM)-based AI tools into veterinary clinical practice is rapidly increasing; however, veterinary-specific frameworks for organizing AI query types and examining stage-specific preferences remain limited. This study aimed to develop an exploratory taxonomy of veterinary AI query types and to characterize veterinarians’ stage-specific query preferences through a structured expert-panel survey. Methods Candidate query types were drafted with the assistance of a locally deployed LLM from a preliminary review of de-identified veterinary AI chatbot query logs and refined through literature review. In a structured online survey, 43 veterinarians then selected the Clinical Support query types (A–H) they would prefer to direct to AI at each of five clinical stages and indicated preferred response formats; the survey measured self-reported query preferences, not observed use. Results A taxonomy of 3 categories and 21 subtypes was established: Clinical Support Queries (Types A–H), Evidence-Based Research Queries (Types I–L), and Terminology and Drug Reference Queries (Types M–U). Across stages, Type B (Differential Reasoning) was most often ranked first (64/213 first-choice responses), and Type D (Clinical Decision Support) was most often ranked first for the post-consultation stage (21/42, 50.0%). In analyses accounting for repeated measures using generalized estimating equations, respondents with ≥10 years of experience more often ranked Type G (Evidence Search) first than those with <10 years (OR 3.09, 95% CI 1.11–8.59; unadjusted p = 0.031), but this contrast was not significant after Benjamini–Hochberg adjustment across four contrasts (adjusted p = 0.062). Respondents from university/teaching hospitals ( n = 7) ranked Type G first in 14 of 33 responses, compared with 10 of 135 responses from secondary referral hospitals (OR 9.27, 95% CI 2.94–29.22; adjusted p < 0.001), although this estimate is imprecise because the subgroup is small. Conclusion This study proposes a veterinary-specific exploratory taxonomy of AI query types and characterizes how respondents’ first-choice preferences varied across clinical stages. These findings offer a preliminary, exploratory framework that may inform the future design of context-aware, stage-adaptive veterinary AI systems, pending cross-national replication and formal validation.