The agent evaluation gap: Enterprises trust AI agents more than the tests meant to catch failures

According to Pulse Research conducted by VentureBeat in June 2026, enterprises face a significant 'evaluation gap' regarding AI agent reliability. Data shows that over the past year, 50% of operating enterprises have deployed AI agents or LLM features that passed internal evaluations but triggered customer-visible failures in production environments (such as erroneous outputs or workflow interruptions). Among these, one-quarter of enterprises reported that such incidents occurred more than once. In contrast, only 36% of enterprises reported no such failures, while 8% do not conduct pre-deployment evaluations, and 6% could not confirm due to a lack of deep tracking into root causes. This finding indicates that 'passing an evaluation' does not equate to 'being usable,' and this disconnect has become industry norm.

The core pain point causing this phenomenon lies in the insufficient alignment between evaluation tools and actual performance. When asked about the main factors limiting trust in automated evaluations, 29% of respondents pointed to poor alignment between evaluation results and real-world outcomes. Additionally, 21% of enterprises attributed issues to bias or inconsistency, 18% cited a lack of explainability (i.e., not knowing the reason behind the evaluation's verdict), and 17% expressed concern over data leakage or privacy issues during the evaluation process. This means that although evaluations are viewed as a gatekeeping mechanism, their own credibility is extremely low, with 95% of enterprises identifying at least one limitation.

Despite extremely low trust in evaluation tools, enterprises are accelerating the granting of higher autonomy to AI agents. The survey shows that 66% of enterprises have already allowed zero-human-review (human-in-the-loop) automated deployment for low-risk agents, or are actively engineering their pipelines to achieve this goal within twelve months. Specifically, 34% of enterprises have implemented this strategy, while 33% are in the planning phase. Only 22% of enterprises explicitly stated that they will exclude zero-human deployment in the foreseeable future. This parallel state of 'low trust, high autonomy' constitutes the evaluation gap referred to in the report: the speed at which enterprises grant autonomy to agents exceeds the speed at which trust in their regulatory testing is established.

Notably, this trend is not exclusive to small and medium-sized enterprises; large enterprises are pushing forward even more strongly. Among large enterprises with more than 2,500 employees, 70% support or plan to achieve zero-human review, higher than the 64% among small and medium-sized enterprises. Meanwhile, large enterprises also experience a higher proportion of cases where 'evaluations passed but customers failed' (54% vs. 48%). This breaks the stereotype that 'large regulated organizations are more conservative,' indicating that larger institutions, while pursuing automation efficiency, also bear higher trial-and-error costs.

The technical foundation supporting this automation process appears weak and dispersed. In terms of evaluation tool selection, no single dominant player holds an absolute advantage. The most commonly used primary evaluation tool is the model provider's native evaluation (17%), which ties with the proportion of enterprises having no dedicated evaluation tools (17%). This fragmentation implies that many enterprises may rely on non-standardized or unverified methods to judge agent performance.

The situation is equally concerning regarding continuous monitoring in production environments. Only about a quarter (25%) of enterprises perform quality checks on real-time production traffic. Most enterprises seem to rely on static pre-deployment evaluations, neglecting dynamic runtime anomaly detection. The lack of monitoring capabilities, combined with the aforementioned high failure rates and low trust levels, further amplifies the risk exposure of automated deployments.