Lily Poursoltan

Generative AI is increasingly embedded in professional workflows to generate drafts that humans remain responsible for evaluating, revising, or accepting before a task is complete. Yet it remains unclear when these drafts save time, when they impose higher verification costs, and whether organizations can predict this difference before a draft is offered. We address these questions through task–technology fit, proposing that human-in-the-loop generative AI changes the locus of fit. Because the system generates a distinct draft for each task instance, fit varies by interaction, depending on how closely the draft aligns with the final response the professional needs to send. We study this framework in patient portal messaging, where clinicians choose whether to begin from an AI-generated draft or from a blank response field. Using thirteen months of operational data from a large academic medical center, covering 12,030 clinician-message interactions and 369.5 clinician-hours of observed portal response work, we find that AI-draft initiation is associated with an approximately 20% reduction in clinician response time. This association is absent in unadjusted comparisons and emerges only after accounting for within-clinician differences and features observed before response initiation. A clinician-driven twelve-category task taxonomy reveals an ordered gradient: estimated time savings are largest for standardized tasks with reusable phrasing, smaller for tasks requiring patient-specific clinical interpretation, and not reliably observed for laboratory result questions. This boundary reflects an information-architecture constraint rather than model capability alone: when privacy-preserving design withholds result values and case-specific clinical context, the draft cannot reliably fit the task. Finally, a selective draft-offering design based on ex ante predicted fit is associated with reliable time savings in categories where uniform draft offering yields inconclusive estimates, while preserving message and clinician coverage. This study contributes by reconceptualizing task–technology fit for human-in-the-loop generative AI, extending the jagged technological frontier from a task-level boundary to an ordered gradient of interaction-level fit, and showing how information architecture and predicted fit shape selective deployment in professional workflows.