A VentureBeat Research survey of 157 enterprises (July 2026) exposes a structural industry problem: half of companies have deployed an AI agent that passed internal evaluations and then failed in production. Meanwhile, two-thirds of those same organizations are building toward zero-human-in-the-loop deployments — even though only 5% fully trust automated evaluation today.
Key takeaways
- 50% of surveyed companies shipped an AI agent that passed evals and then caused a customer-facing incident
- Only 5% of organizations fully trust automated agent evaluation
- 66% already allow zero-human-in-the-loop deployments or are actively building toward them
- Top evaluation weakness cited: poor alignment with real-world outcomes (29%)
- Most common evaluation tool: OpenAI native evals — tied with having no dedicated tooling at all (17% each)
The gap between test and reality
The report identifies what the authors call an "evaluation gap?evaluation gap: The mismatch between the autonomy granted to AI agents and the actual reliability of the tests meant to justify that autonomy.": the distance between the autonomy enterprises grant AI agents and the trust they place in the tests meant to legitimize that autonomy. Half of organizations (50%) confirmed that in the past 12 months, they deployed an agent or LLM feature that passed internal evaluations and then caused a customer-visible incident. One in four experienced this more than once.
Only 36% reported no such incident. The remaining 14% either do not run pre-deployment evaluations (8%) or do not track root causes closely enough to know (6%).
Trust in evaluation is near zero
Asked about the main weakness of automated evaluation, respondents pointed primarily to poor alignment with real-world production conditions (29%), bias and inconsistency (21%), and lack of explainability (18%). Data leakage concerns in the evaluation process were raised by 17%.
Just 5% said they fully trust automated evaluation — meaning 95% of companies have concrete objections to the tools they are simultaneously using to gate production. Tools from OpenAI and Anthropic lead among evaluators used.
Autonomy is growing faster than trust
The paradox at the center of the report: despite such low trust levels, the industry is moving toward more agent autonomy, not less. 34% of companies already allow no-human-in-the-loop?human-in-the-loop: An operating model in which a human approves or oversees an AI system’s decisions before they take effect. deployments for low-risk agents. Another 33% are actively building pipelines toward that mode within 12 months. Only 22% rule out full deployment automation for the foreseeable future.
Notably, larger organizations (2,500+ employees) are further along the zero-human path (70% vs 64% for smaller firms). The assumption that large, regulated companies are more cautious does not hold in this data.
Evaluation tooling: early and fragmented
The enterprise evaluation stack is scattered and dominated by model-provider-native tooling. OpenAI native evals lead at 17%, tied with companies having no dedicated tooling at all (also 17%). Anthropic Claude Console native evals have 13%, DeepEval 12%, custom in-house solutions 11%.
| Tool | Share |
|---|---|
| OpenAI native evals | 17% |
| No dedicated tooling | 17% |
| Anthropic Claude Console | 13% |
| DeepEval | 12% |
| Custom in-house | 11% |
The market is very early: 64% of companies plan to change or expand their evaluation stack within a year, and 31% within the next quarter.
Where the next budget goes
Planned investments signal industry awareness of the problem: 30% of respondents cited production observability tooling as their top priority, and 26% cited human review workflows. The latter is particularly telling — companies simultaneously building zero-human pipelines plan to increase spending on human review. Only 16% plan to invest primarily in automated evaluation pipelines.
Why this matters?
The report precisely documents a tension that was previously visible only anecdotally: companies are deploying AI agents faster than they can reliably test them. This is not a tooling problem that more tests will fix on its own — the authors stress that evaluations simply do not reflect production conditions, which is a fundamental problem.
The consequences are direct for companies building agent-based products: a passing eval score does not equal a working agent in production. Half of organizations had to learn this at their customers' expense. As autonomy scales — without human review of every deployment — these incidents will scale proportionally.
The data also shows that the evaluation tooling market is exceptionally early: no dominant vendor, 17% of companies with no dedicated tooling, and model-provider-native evals leading. This is open space for consolidation — and a signal that assessing AI agent quality is far harder than the pace of deployments would suggest.
What is next?
- 64% of companies plan to change or expand their evaluation stack within 12 months — the next wave of specialized tooling adoption is coming
- VentureBeat announced ongoing Agentic Reliability & Evals tracker waves — comparable data will follow in future quarters
- Anthropic Claude leads among tools being considered for adoption (20% of firms evaluating it), which may break the current OpenAI-vs-no-tooling tie at the top
Sources
VentureBeat — The agent evaluation gap
