Artificial intelligence-supported assessment is increasingly entering educational practice through dashboards, automated feedback systems, learning analytics, image-based analysis, and generative AI tools. In primary education, these developments raise a distinctive problem: assessment outputs are produced by computational systems, but their consequences are experienced by young children whose learning, confidence, participation, and academic identities are still developing. This chapter argues that the central question is not whether AI can generate assessment information, but how such information should be interpreted, ethically filtered, and translated into pedagogical action in primary classrooms. The chapter contributes to existing debates by developing a teacher-mediated and child-centred framework for responsible AI-supported assessment. Rather than treating AI as an autonomous evaluator or as a neutral source of evidence, the chapter positions AI outputs as provisional indicators that require professional judgement, contextual interpretation, and ethical scrutiny. The discussion examines pedagogical opportunities such as early identification, progress monitoring, differentiated support, and formative feedback, while also critically addressing algorithmic bias, opacity, datafication, misclassification, privacy, developmental vulnerability, and the risk of premature labelling. The proposed framework connects five components: evidence generation, teacher interpretation, ethical filtering, pedagogical decision-making, and continuous review. By showing how these components function as an interpretive cycle rather than as a simple checklist, the chapter clarifies how AI-generated outputs can support, but not replace, teacher judgement. The chapter concludes that AI-supported assessment can serve primary education only when it remains subordinate to pedagogical purposes, child dignity, and human responsibility.