A Novel Workflow for Artificial Intelligence-Enhanced Patient Messaging Services

Abstract Current artificial intelligence (AI) integration in patient messaging often relies on generating AI drafts for clinician review, yet this workflow has achieved limited effect. This study aimed to describe and evaluate a novel, clinician-first workflow for patient messaging where AI enhances a clinician-generated draft. Using 268 patient questions from public data, we compared physician-only responses, AI-only responses, and AI-enhanced responses. Responses were ranked on overall preference and the CREATE TRUST framework. AI-enhanced responses were significantly preferred overall, ranking first in 38.8% of evaluations (average rank 1.69; p < 0.01), outperforming both AI-only (27.6%; 2.29) and physician-only (25.5%; 2.11) responses. AI-enhanced responses ranked highest for Understandable (44.5%) and Tailored (39.4%). AI-only responses ranked highest for Thorough (71.0%) and Empathic (69.8%), while physician-only responses ranked highest in Authentic (90.9%; all p < 0.01). Safety analysis identified consequential additions in 3.36% and omissions in 1.12% of AI-enhanced messages. A novel workflow—based on the Clinical Action Support framework—where AI enhances a clinician's draft may offer an improved approach to AI implementation in patient messaging services.