Enterprise AI survey finds production failures persist after internal approval

VentureBeat's July survey of 108 enterprises found that 49% had seen an AI agent or LLM feature pass company testing and later create a customer-visible problem. Among respondents reporting such an incident, 85% were pursuing limited or planned deployments that can change systems or push code without human approval. The survey says 26% used automated inline quality assertions for live output checks, while its sample was self-selected and should be read directionally.