Personality-Adaptive Conversational AI for Emotional Support: A Simulation Study Integrating Big Five Detection with Zurich Model Regulation

LLM-based conversational agents generate fluent responses but remain limited in adapting their supportive style to individual personality and emotional needs. We address this gap through a theory-grounded personality-adaptive conversational AI pipeline that performs turn-by-turn Big Five trait detection and applies Zurich Model-aligned behavioural regulation, orchestrated with PROMISE (a model-driven framework for state-based LLM orchestration). We generated 20 simulated dialogue sessions (120 assistant turns total) comparing regulated and non-adaptive assistants across two boundary-condition Big Five profiles (all traits set to +1 vs. -1), evaluated using a structured LLM-based evaluator complemented by author-conducted qualitative review. The system achieved near-perfect implementation fidelity (98.3% detection accuracy; 100% regulation adherence). Personality Needs Addressed were met consistently under regulation (100%) but rarely at baseline (8.3%; Cohen’s d = 4.42, p < 0.001), reflecting a binary architectural difference rather than a graded dose-response improvement. Emotional tone and relevance did not differ between conditions (d = 0.000 and d = 0.183, ns), confirming selective enhancement of personalisation without degrading generic conversational quality. These findings establish a reproducible detection → regulation → evaluation framework for future empirical validation of personality-adaptive AI-augmented mental health support with human participants.