- The paper presents a values-centered conversational AI that integrates SMART goal-setting and ACT techniques to support non-clinical well-being in high-stress academic environments.
- It employs a stateful dialogue architecture, reflective dashboard, and calendar integrations, validated by expert and student studies with high usability scores.
- The study highlights the benefits of continuous goal tracking and ethical data handling, while noting limitations like short deployment duration and context sensitivity.
GROW: A Technical Synthesis and Assessment of a Goal-Oriented Conversational AI Coach for Student Well-Being
System Motivation and Conceptual Foundations
GROW operationalizes a structured, values-centered conversational AI designed for university student well-being in high-stress academic environments. The paper delineates limitations in contemporary digital mental health tools which typically foreground CBT-oriented symptom reduction, fail to integrate values-based reflection and longitudinal follow-up, and neglect workflow integration into students’ daily routines. GROW's architecture addresses these gaps by explicitly embedding the SMART goal framework and ACT principles—particularly values clarification and committed action—into an LLM-powered coaching interface, dashboard, and peripheral notification infrastructure.
GROW is neither a clinical platform nor a diagnostic instrument. It explicitly positions itself as a non-clinical, supportive intervention aligned with evidence on non-judgmental, values-driven reflection as an enabler of motivation and well-being, drawing from extensive ACT literature and empirical meta-analyses on goal-setting efficacy within student populations. The system strategically avoids clinical scope creep, making the boundaries of engagement explicit and integrating robust escalation policies to institutional and crisis mental health resources.
System Design and Architecture
The system comprises:
- Conversational AI Coach: Implements a stateful dialogue architecture, transitioning users from rapport-building and personality assessment (OCEAN model), to domain-specific values elicitation (modified BEVS dialogue), through iterative SMART goal formation, and into ongoing follow-up leveraging ACT-derived micro-interventions for psychological flexibility.
- Reflective Dashboard: Aggregates and visualizes user-specific progress metrics, check-in cadence, goal evolution, and communication style summaries. It enables students to review historical goal engagement and values-action alignment with coarse-grained, non-diagnostic cues.
- Calendar and Notification Integrations: Goal check-ins and coaching interactions are embedded into Google Calendar with customizable timing, supplementing asynchronous email reminders, thereby optimizing for minimal friction and maximal ecological validity in adherence.
Interactional logic is phase-gated. Synchronous user input is parsed and augmented with phase-specific prompting; outputs are validated and persisted via pseudonymous identifiers, maintaining architectural separation from PII. Backend implementation employs few-shot prompt chaining and strict state-machine boundaries to avoid prompt contamination and conversational drift—critical for controlled, predictable user experiences.
Empirical Evaluation
Practitioner Study
The formative phase involved five domain experts (clinical psychologists, student coaching professionals). The findings showed:
- Usability scores (SUS: 83.5 ± 5.2) classified GROW as yielding 'Excellent' usability.
- High perceived system clarity, with the ACT/SMART integration conferring legitimacy and structure absent in generic chatbots.
- Positive ratings for features facilitating reflection and follow-through (goal tracking, values alignment, conversational summaries).
- Persistent emphasis on ethical boundaries and anonymized architecture as precondition for deployment in student-facing scenarios.
Salient recommendations included making underlying psychological frameworks transparent but accessible, reducing clinical jargon, and enhancing explanation of operational boundaries.
Student Deployment
A one-week deployment with n=30 undergraduates yielded:
- Above-threshold usability (SUS: 77.5 ± 5.2; 'Good-Excellent' range).
- Strong endorsement (80–87% "very/somewhat helpful") for the conversational coach across both goal pursuit and perceived well-being.
- Dashboard progress visualizations and check-in consistency cues were reported as primary drivers of engagement and motivation.
- Reflexive features (communication style, values alignment) were effective as introspective prompts.
- Users experienced higher trust when the system exhibited contextual memory and specific follow-up, and disengagement when responses became repetitive.
- Barrier reduction for self-disclosure attributed to the non-human, non-evaluative nature of the agent, though some privacy reticence persisted.
PERMA-Profiler metrics revealed modest yet directionally positive changes (notably +0.45 week-over-week in Engagement, Accomplishment, Positive Emotion).
Technical and Theoretical Implications
The study empirically substantiates:
- The role of attentional continuity and explicit memory in building user trust and engagement in goal-oriented conversational agents.
- The effectiveness of treating SMART goal-setting as dynamic, revisable commitments embedded in workflow, contrasting one-off static instantiations.
- Dashboard visualizations and lightweight feedback act as effective behavioral cues; reflection is catalyzed by these representations, not conversation alone.
- ACT and SMART serve as embedded design patterns for interaction structuring, not as overt frameworks taught to users, avoiding overwhelming the non-clinical user base.
A key limitation is the short deployment duration (1 week) and absence of a randomized control condition, precluding causal inferences regarding well-being outcomes. Prompt design also suffered from some context insensitivity and repetition, evidencing known constraints of template-based LLM system instantiations.
Design and Deployment Considerations
The findings imply:
- Conversational AI for well-being should prioritize longitudinal engagement via attentive follow-up, explicit representation of user history, and integration into daily workflows.
- Non-clinical scope and data boundaries must be explicit, both for ethical compliance and to set user expectations.
- Reflection-support technology in academic populations benefits from values-based goal alignment and actionable progress tracking, with emotional support serving a secondary, supporting role.
Conclusion
GROW exemplifies the potential for LLM-powered systems to support student well-being via structured, adaptive, and workflow-integrated goal setting, reflection, and accountability. The hybridization of ACT and SMART as latent interactional scaffolds, deployed in non-clinical academic contexts, demonstrates strong usability and high acceptability among both practitioners and students. Theoretical contributions reside in the design of continuity-focused, non-diagnostic conversational systems and the operationalization of values-action alignment outside clinical paradigms. Open questions remain regarding long-term engagement dynamics, optimal feedback granularity, and comparative efficacy in multi-arm trials. Continued technical refinement—including context-aware prompting and adaptive feedback—is warranted for scaling and sustained real-world impact.