A new wave of research from VentureBeat reveals a critical disconnect in enterprise AI deployment: organizations are granting their agents increasing autonomy while simultaneously distrusting the evaluations meant to catch failures. The Agentic Reliability and Evals survey, fielded in June, found that half of enterprises have already shipped an agent that passed internal evaluations only to cause a customer-facing failure. Yet only 5% of organizations fully trust automated evaluation today, and the most cited weakness is that evaluations do not align with real-world outcomes. Two-thirds of enterprises, however, are already allowing or engineering toward fully automated, zero-human-in-the-loop deployment.

What You Need to Know

Enterprise AI teams are racing to deploy autonomous agents, but the evaluations designed to ensure reliability are failing to catch real-world problems. This gap between autonomy and trust is creating a risky dynamic where failures are common and confidence in automated checks remains low. The data comes from a survey of 157 organizations with 100 or more employees, conducted by VentureBeat in June.

The Evaluation Failure Rate Is Stubbornly High

The survey’s most striking finding is that half of organizations (50%) have, in the past year, deployed an agent or LLM feature that passed internal evaluations and then triggered a customer-facing failure. A quarter have experienced this more than once. Only 36% report no such failure, while the rest either skip pre-deployment evaluations or lack tracking. The core problem is clear: passing an evaluation does not guarantee a working agent.

This experience shapes everything else. Enterprises that have seen a passed eval lead to a production incident are far less likely to trust automated checks going forward. Yet the pressure to ship remains high. The result is a tension between velocity and verification that many organizations are still struggling to resolve.

Trust in Automated Evaluations Remains Fragile

Only 5% of enterprises say they fully trust automated evaluation in its current form. The remaining 95% cite specific limitations. The most common complaint, at 29%, is that evaluations align poorly with real-world outcomes. Bias or inconsistency follows at 21%, and a lack of explainability at 18%. These gaps directly explain the failure rate: if evaluations do not mirror production conditions, they cannot catch the failures that matter.

  • Poor real-world alignment: 29% of enterprises say evals don’t match actual customer scenarios.
  • Bias or inconsistency: 21% report evaluations that vary unpredictably across runs or inputs.
  • Lack of explainability: 18% cannot understand why an eval passed or failed an agent.

The tooling landscape is equally fragmented. The most common primary evaluation tools are the model providers’ native evals, tied with having no dedicated tooling at all, each at 17%. Only about a quarter of enterprises run real-time quality checks on live production traffic. This immaturity compounds the trust problem.

Autonomy Is Outpacing Assurance

Despite low trust, enterprises are pushing forward. Two-thirds of organizations (66%) 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 autonomy is arriving faster than the assurance. The evaluation gap — the distance between how much autonomy enterprises grant their agents and how much they trust the tests meant to govern that autonomy — is widening.

This pattern suggests a high-risk strategy. Organizations are betting that speed will outrun failure, but the data shows that failures are already frequent. Without better evaluation alignment, the gap will continue to produce incidents that erode customer trust and increase operational costs.

Why This Matters

The implications extend beyond individual failures. Enterprise AI adoption is accelerating, and agents are being given more decision-making power in areas like customer support, code generation, and data analysis. If evaluations cannot reliably predict real-world performance, the consequences include damaged brand reputation, financial losses, and regulatory exposure. For technical leaders, the message is urgent: invest in evaluation frameworks that mirror production conditions, or risk scaling failures alongside autonomy. The survey from June makes clear that the current trajectory is unsustainable.