Agentic AI is entering software engineering workflows, but empirical evidence on its transition from experimental capability to production use remains limited. We report a qualitative interview study with 16 practitioners from 12 companies, using a six-level maturity framework as an analytical lens. Reported production practices corresponded to Levels 1-3, while participants in four companies reported experimental capabilities beyond production-integrated use. Across the cases, four previously identified barriers recurred: context management, performance on proprietary content, non-determinism and qualification, and data confidentiality. We synthesize their interaction as a capability-deployment verification gap structured by two interdependent dimensions: information asymmetry and qualification absence. The study thereby characterizes reported adoption practices and explains what constrains further agentic automation in the represented industrial contexts.

翻译：暂无翻译