Enterprise organizations are rapidly deploying AI agents to production despite significant gaps in evaluation trust and security controls, according to three reports from VentureBeat AI covering 365 total enterprises. According to the evaluation study of 157 organizations, half have shipped an agent that passed internal evaluations but subsequently failed in production with customers, and only one in twenty enterprises fully trusts their automated evaluations even as they grant agents increased autonomy.
Security concerns are materializing at scale. According to VentureBeat’s security survey of 107 enterprises, 54% have already experienced a confirmed AI agent security incident or near-miss. Despite these incidents, most organizations still allow agents to share credentials, with only about one-third giving each agent its own scoped identity.
The infrastructure challenges extend to how agents access business context. According to the context study of 101 enterprises, retrieval-augmented generation has become the default method for providing AI agents with business information, and provider-native retrieval solutions have quietly overtaken dedicated vector databases. However, VentureBeat reports that this infrastructure is being built faster than it can be trusted, with organizations continuing to develop fixes while deploying agents to production environments.