The rapid aging of the global population presents challenges in adopting new technologies, especially for the elderly, whose cognitive and social changes may delay their interaction with advanced information systems. This study explores how anthropomorphism and interaction modes impact elderly users’ experience engaging with generative AI agents. Ten participants aged 52–72 interacted with three anthropomorphic AI agents across various tasks using voice messages and video calls. Semi-structured interviews were conducted before and after the tasks to assess their expectations and experiences. The results show that discrepancies between before-use expectations and actual experiences significantly affect users’ perceptions of the AI’s usefulness, aligning with the Expectancy Violations Theory (EVT). Interaction modes, especially that of video chat, influenced participants’ satisfaction, emotional attachment, and privacy concerns. The study highlights the importance of incorporating social presence and emotional cues in AI design to improve elderly users’ engagement, trust, and overall satisfaction. These insights are essential for creating generative AI systems tailored to the elderly, supporting their long-term integration into senior communities.