Structured Large Language Model Workflows for Motivational Interviewing in Health Behavior Change: Proof-of-Concept...

Motivational interviewing (MI) is an effective approach for supporting health behaviorchange, but face-to-face delivery is resource-intensive and difficult to scale. Rule-based conversational agents (CAs) can improve access; however, their scripted interactions and limited language flexibility constrain MI delivery. While large language models (LLMs) are increasingly being used for MI coaching, their conversational fidelity and quality compared with human coaches and rule-based CAs remain understudied.

This study aimed to describe the development of an LLM-based CA, Artificially Intelligent Motivational Interviewing (Aimi), orchestrated with structured workflows, and to evaluate its feasibility, conversational fidelity, and user perceptions during MI coaching interactions.

We developed Aimi using structured LLM workflows designed to enhance MI fidelity. We conducted a within-participants study, where 18 adults interacted with (1) Aimi, (2) a novice MI-trained human coach, and (3) a rule-based CA during live text-based role-play coaching sessions. Transcripts were independently evaluated by an MI expert using the Motivational Interviewing Skill Code, Version 2.0 (MISC-2), to assess MI competency and fidelity. Participants completed a user experience questionnaire to provide general feedback and to assess session alliance, dialogue relevance, empathy, engagement, linguistic quality, and perceived motivation to change. Feedback from users was thematically summarized and categorized under strengths and weaknesses for each approach.

Aimi achieved fidelity scores comparable to those of the novice human coach and higher than those of the rule-based CA on summary metrics, including higher reflection-to-question ratios (median 0.84, IQR 0.62-0.92 vs 0.62, IQR 0.42-0.74 vs 0.25, IQR 0.17-0.38), more complex reflections (median 66.67%, IQR 46.97%-76.92% vs 50%, IQR 34.38%-61.88% vs 0.00%, IQR 0%-50%), and greater elicitation of client change talk (median 90.83%, IQR 85.89%-100% vs 73.21%, IQR 63.10%-83.19% vs 66.67%, IQR 57.86%-81.94%). User experience ratings showed no significant differences across conditions. User feedback revealed distinct strengths and limitations across the coaching interactions. Participants described Aimi's interactions as personalized, fluid, and adaptive, though sometimes overly reflective and lengthy. The novice human coach was viewed as empathetic and supportive but slow to respond, whereas the rule-based coach was viewed as efficient and structured yet limited in depth and personalization.

This study demonstrates the technical feasibility of structured LLM-workflows for MI coaching and their capacity to maintain conversational fidelity comparable to that of a novice MI-trained human coach. Given the role-play paradigm, single-rater coding, and small convenience sample, these comparative findings should be interpreted as exploratory. Our findings serve as a foundational baseline for the development of scalable behavior change interventions in clinical settings.