A new survey of 108 enterprises shows that 49% experienced AI agents or features that passed internal testing but then failed in production, yet companies are still moving faster to remove humans from deployment decisions rather than slower.
Trust in automated evaluation jumped to 13% in July from 5% in June, while concerns about misalignment between tests and real-world performance dropped from 29% to 19% — suggesting companies are gaining confidence in their testing methods even as evidence mounts that those methods are missing critical failures.
What happened
A VentureBeat Intelligence survey of 108 enterprises found that 49% of respondents said an AI agent or LLM-powered feature that passed company testing later caused a customer-facing problem; nearly a quarter (24%) reported this happened more than once. Despite this, trust in automated evaluation rose to 13% in July from 5% in June, and concerns about misalignment between tests and real-world results fell from 29% to 19% month over month.
Why it matters
Companies that have already experienced AI failures in production are paradoxically accelerating their removal of humans from deployment decisions rather than slowing down, even as they acknowledge the gap between testing and real-world performance. This suggests enterprises may be doubling down on automation despite evidence that their evaluation methods are not catching problems before customers encounter them.
What to watch
The survey tracks a widening disconnect: rising confidence in automated evaluation methods even as nearly half of enterprises report their testing failed to prevent customer-visible failures. The research indicates a critical phase in enterprise AI rollout where the gap between evaluation and production performance remains unresolved.
Ask the AI about this article →
The survey reveals a paradox at the heart of enterprise AI deployment: companies are experiencing measurable, repeated failures in production despite their testing processes, yet they are responding by accelerating automation rather than strengthening human oversight. The data shows that 49% of enterprises have already been burned by AI agents or features that passed internal evaluation but failed when customers encountered them — a gap that undermines the reliability of the testing process itself.
Yet trust in automated evaluation is climbing sharply. In just one month, confidence jumped from 5% to 13%, and the share of companies worrying about misalignment fell from 29% to 19%. This shift suggests that enterprises are not learning to distrust their evaluation methods in the face of failures; instead, they appear to be interpreting the problem differently — or shifting responsibility away from human decision-makers. The fact that the survey notes companies are "moving faster toward removing humans from deployment decisions, not slower" even after experiencing real-world failures indicates a systematic move toward full automation, regardless of evidence that evaluation is incomplete.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Israeli startup DataAgent Ltd
SK Hynix presented a custom HBM concept at SEMICON Taiwan 2026, where compute functions are placed in the base…
The U.S. Department of Defense announced on August 31 that it has deployed ChatGPT Mil, a customized version o…
Nvidia reported earnings that were both remarkable and boring, reflecting its focus on avoiding a consolidated…
Anthropic has agreed to a $35bn cloud-computing contract with Lambda, a Nvidia-backed cloud provider
The Consumer Affairs Agency said Tuesday it will use generative AI to analyze about 900,000 annual consultatio…
