Generative Agent-Based Simulations of Workforce Wellbeing
Authors: Arash Khajeh, Vindula Jayawardana, Tatiana Lau, Luiza Santos, Kate Sieck Published in: HHAI 2026, Frontiers in Artificial Intelligence and Applications, Vol. 423
Workforce wellbeing is dynamic and path-dependent, yet most workforce analytics remain correlational, offering limited support for comparing interventions or surfacing the underlying mechanisms. We present a hybrid human–AI agent-based simulation framework in which a large language model (LLM) generates monthly state transitions under human-designed empirical and structural constraints, enabling controlled counterfactual replay of organizational shocks. Across two shocks (overwork, promotion/bias) and four intervention arms, combined-mode interventions dominate under promotion/bias, and communication alone outperforms structural alone under bias. Subgroup analysis reveals a cross-shock vulnerability reversal by financial strain that mechanism analysis on the agents’ monologues traces to scarcity-mindset framing under both shocks and identity-invalidation framing only under unfair outcomes — the latter suppressible by procedural transparency.