Enterprise teams are handing agents more autonomy than they have assurance to cover. A VB Pulse survey of 157 enterprise respondents at companies with 100 or more employees, reported by VentureBeat in July 2026, found 66% already permit some production deployment without human review, or are building systems intended to do so within twelve months. In the same survey, 5% said they fully trust the automated evaluations that would make those release decisions.
That is the gap worth worrying about. Not a shortage of evaluation. A shortage of evaluation anyone believes.
Passing the evals is not the same as not failing
Half of the organisations surveyed had shipped an AI agent or LLM feature that passed their internal evaluations and then caused a customer-facing failure. A quarter had seen it happen more than once.
The most cited reason for distrusting automated evaluation was not missing tests. It was that evaluations align poorly with real-world outcomes, named by 29%. Bias or inconsistency followed on 21%, lack of explainability on 18% and data leakage or privacy concerns on 17%.
That ordering matters. These teams are not describing a coverage problem. They have suites and the suites pass. What has broken down is the correlation between passing and behaving.
One caveat on the numbers. The sample was self-selected rather than a probability sample, so treat the figures as directional. The direction is at least consistent internally: enterprises of 2,500 employees and up reported both more autonomy, on 70%, and more customer-facing failures, on 54%, than smaller firms on 64% and 48%.
Where reliability actually comes from
A paper published on 19 July 2026 by Arunabh Dastidar of the Leni team takes this apart for a real production agent, measured across three public benchmarks. On SpreadsheetBench Verified the system reached 91.25% against a baseline of 80.25%, a gain of 11 points.
The useful finding is where that gain came from. Most of it came from scaffolding, routing and specialist models rather than from the verification step itself, whose isolated contribution was 1.5 points. Swapping the trained verifier for the base frontier model eliminated most of the rescues.
Read that as a warning about bolt-on assurance. A checker added at the end of a weak pipeline is not verification. The verification layer earned its keep only as a trained specialist inside a system designed to be checkable in the first place.
Three things to change before your next release
- Set the assurance bar by autonomy level, not by feature - a step that commits an action with no human review needs a higher evidence threshold than one that drafts something for a person to approve.
- Track whether passing predicts not failing - a stable pass rate alongside rising incident counts means your suite is measuring the wrong thing.
- Put verification inside the pipeline rather than after it - and staff it with a model trained for the check, not a general-purpose model bolted on at the end.
Autonomy is arriving faster than the assurance beneath it. Closing that gap is an evaluation design problem, not a tooling purchase.
EvalSpec does not monitor production. It gives you a structured pre-release evaluation: 70 test cases across five dimensions, scored by an LLM judge that you can review and override.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation