In 2025, 305 million people required humanitarian assistance, yet 75–85% of those with mental health disorders in low- and middle-income countries remain without care. While artificial intelligence (AI) is increasingly proposed to address this gap, humanitarian settings present unique challenges, including fragmented data systems, cultural and linguistic diversity and the central role of relational care. This perspective article examines the evolving AI landscape for Mental Health and Psychosocial Support (MHPSS) and introduces the “Humanitarian AI Paradox,” where widespread informal use of AI is already outpacing institutional strategy and governance. We outline four key developments shaping feasibility, including agentic AI, the emergence of an agentic workforce, the growing availability of open-weight models and the collapse of inference costs. Using the Inter-Agency Standing Committee (IASC) pyramid, we map the AI applications across four levels of care, identifying focused, non-specialised support (Level 3) as the most immediate opportunity to assist frontline workers through supervision and protocol guidance. We then examine domain-specific risks, including cultural and linguistic misinterpretation and higher operational costs for non-English languages due to token-based pricing. Finally, we propose a set of research needs necessary to support the safe, equitable and context-appropriate use of AI in humanitarian MHPSS.