CAIA in Practice: Field Evaluation of an AI-Assisted Support System for Text-Based Online Counselling

Asynchronous email counselling is a vital, low-threshold channel for mental health support, yet practitioners face significant service delivery challenges. This study introduces CAIA (Counselling AI Assistant), an AI-based tool suite designed to augment, rather than replace, human counsellors. Developed through a co-design process involving practitioners and researchers, CAIA provides seven LLM-driven functionalities—such as session summaries, social network mapping, and methodological suggestions—integrated via a non-intrusive, split-screen interface. The researchers conducted a four-week field evaluation with 34 professional counsellors who engaged with trained students enacting realistic case vignettes to assess how these tools fit into actual clinical workflows.

The field evaluation revealed that counsellors do not adopt all AI tools equally. Interpretive functionalities—specifically those offering hypotheses and methodological suggestions—achieved the highest engagement, with counsellors frequently consulting these panels while composing their own responses. In contrast, information extraction tools (like session summaries and situation overviews) were used primarily for rapid orientation or case handovers. A notable finding was the 'paradox of engagement' regarding hypotheses: even when counsellors were critical of the AI's output, they still valued the tool for stimulating professional reflection and providing fresh perspectives.

Professional autonomy emerged as the decisive factor for sustained acceptance. Counsellors preferred the non-intrusive design, which allowed them to choose when and how to engage with AI assistance. However, the study highlights a critical vulnerability: accuracy. When the AI produced hallucinations or incomplete information, it significantly undermined professional confidence. This was most evident with the Social Network Graph, which, despite its conceptual appeal, suffered from implementation errors that led to a decline in usage. The results suggest that for AI to be successfully integrated into mental health practice, it must prioritize high-fidelity outputs and support the counsellor's existing decision-making process rather than attempting to automate it.