Industrial visual inspection operates at the intersection of perception, decision-making and physical consequences. Despite advances in automated vision and large foundation models, industrial inspection remains fundamentally unsolved: the assets being inspected are domain-specific and evolving, labeled data is scarce, environments drift and inspection decisions must be made under uncertainty with safety constraints. In this talk, we argue that these challenges require a shift from single-pass perception to an agentic physical AI approach. Drawing on industrial deployments, including automated welding inspection, we identify the structural limits of classical pipelines and generic foundation models. We then introduce a human-centric, agentic physical AI framework that treats inspection as a closed-loop cognitive process, integrating memory, geometry-aware alignment, uncertainty-aware reasoning and safe adaptation over time. This architecture is inherently device-agnostic and domain-specific, enabling future inspection systems that scale across assets and environments while remaining interpretable, auditable and aligned with industrial practice.