Enterprise AI agents face a reality-alignment gap: 50% fail in production after passing evals

The agent evaluation gap is widening as enterprises grant AI agents more autonomy than the tests meant to govern them can support. VentureBeat's Pulse survey of 157 organizations (100+ employees) found that half have, in the past year, deployed an agent that passed internal evaluations only to cause a customer-facing failure. A quarter have seen this happen multiple times. Trust in automated evaluation remains perilously thin: just 5% of respondents fully trust it, and the top cited limitation (29%) is that evaluations do not align well with real-world outcomes. In short, a passing eval is not a working agent.

Yet the direction of travel is toward greater autonomy. Two-thirds of organizations already permit fully automated, zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to allow it within twelve months (33%). The evaluation stack itself is fragmented and immature: the most common primary tools are model providers' native evals tied with having no dedicated tooling at all (17% each). Only about a quarter of enterprises run real-time quality checks on live production traffic. The autonomy is arriving faster than the assurance, creating a dangerous gap between how much freedom agents have and how little trust organizations place in the safety nets that are supposed to catch failures.