This chapter reflects on the role of agent-based modelling (ABM) in the AI FORA project, which sought to explore how Artificial Intelligence (AI) for social welfare assessment decisions might be made more context-sensitive and better aligned with societal values. We discuss the development of bespoke ABMs for three case studies—Spain, Estonia, and Germany—and how they were used to support participatory workshops and serious games in pursuit of ‘better AI’. By simulating decision-making rules and their effects, the models helped to surface dynamics that are not always visible in practice. They also helped to inform the design of serious games by enabling rules and parameters to be refined in advance. The three case studies differed in their aims, scope and access to stakeholders, which shaped both the resulting models and insights generated. Our experiences highlight the importance of early and sustained stakeholder engagement, with careful mapping and relationship-building to help ensure that models reflect real-world knowledge and that resulting participatory games yield meaningful insights. Together, the case studies establish a prototype for an approach to tailoring and improving AI systems that is grounded in stakeholder engagement and responsive to the ethical, political, and cultural dimensions of social assessment and public service provision.