Across 157 enterprises, organizations are granting AI agents increasing autonomy while losing confidence in the evaluations meant to govern that autonomy. Half have already deployed an agent that passed internal evaluations only to fail in front of a customer. Only one in twenty fully trusts automated evaluation today, and the most frequently cited weakness is that evaluations do not align with real-world outcomes. Nevertheless, two-thirds of organizations either already permit or are actively engineering toward deploying agent changes to production based on automated evaluation alone — with no human in the loop.

The result is what VentureBeat terms an "evaluation gap": the distance between how much autonomy enterprises hand their agents and how much they trust the tests designed to catch failures. The gap is partly structural. Unlike traditional software, where deterministic unit and integration tests can verify expected behavior, AI agents chain multiple reasoning steps, call external tools and APIs, and produce variable outputs for the same input — making it inherently difficult to define what a correct outcome looks like, let alone automate a test for it.

Survey Overview

This installment of VentureBeat Pulse Research — the Agentic Reliability & Evals tracker — examines how technical leaders measure agent performance, which reliability and evaluation platforms they use, how they select and trust them, what breaks in production, and how far they are willing to let agents operate without human oversight.

Methodology

VentureBeat conducted this survey as part of its ongoing Pulse Research series. Responses were filtered to organizations with 100 or more employees (n=157), drawn from a single survey fielded in June 2026. Because this is one wave rather than a pooled multi-month sample, the report is cross-sectional and does not infer month-over-month trends. Where questions allowed multiple selections, shares may sum to more than 100%.

The sample is senior and buyer-credible by role: 38% are final decision-makers for AI purchases, and another 34% are recommenders or influencers. Named titles include product and program managers (15%), consultants and advisors (10%), directors of engineering/IT (8%), and CIOs/CTOs/CISOs (8%), alongside a large "Other" category (37%).

By organization size, the sample is mid-market-weighted: 100–499 employees (37%) and 500–2,499 employees (27%) lead, followed by 2,500–9,999 (20%), 10,000–49,999 (10%), and 50,000+ (6%). Technology/Software is the largest represented industry at 23%, followed by Retail/Consumer (15%), Healthcare/Life Sciences (12%), and Manufacturing (10%).

At 157 respondents, the sample is large enough to be directionally informative but should be treated as a signal rather than a precise measurement. It is self-selected, not a probability sample, and skews toward the mid-market — making it best read as the perspective of organizations actively building agent evaluation practices rather than the largest operators.

Note: This survey was rebuilt for the June wave from the earlier "LLM observability and evaluations" survey. Because the questions and sample differ, no comparisons are made to the April–May data.

Finding 1: A Passing Eval Is Not a Working Agent

Organizations were asked whether, in the past 12 months, they had deployed an agent or LLM feature that passed internal evaluations but then caused a customer-facing failure. Half of those that run evaluations reported they had.

This is the report's defining statistic. Half of organizations (50%) have shipped an AI feature that cleared internal evaluations and then failed in front of a customer — whether through an incorrect output, a broken workflow, or a quality incident. A quarter have experienced this more than once. Only 36% report no such failure, while the remainder either run no pre-deployment evaluations (8%) or do not track root causes closely enough to know (6%).

The failure is precise and consequential: the evaluation indicated the agent was ready, and it was not. Every subsequent finding — how enterprises trust their evaluations, what they monitor, and how much autonomy they grant — is shaped by this experience.

Finding 2: Almost No One Fully Trusts Automated Evaluation

Respondents were asked which limitation most reduces their trust in automated agent evaluations. Only a sliver of enterprises reported no complaint at all.

Trust in automated evaluation is scarce and specific. Only 5% of organizations say they fully trust automated evaluation as it currently stands, meaning 95% identify a limiting factor. The most common limitation, cited by 29%, directly explains Finding 1: evaluations align poorly with real-world outcomes, passing agents that later fail. Bias or inconsistency follows at 21%, and a lack of explainability — enterprises cannot always determine why an evaluation reached its verdict — accounts for 18%. Another 17% cite data-leakage or privacy concerns within the evaluation process itself.

The tests designed to certify agents are not yet trusted to do so, which makes the autonomy trajectory described in Finding 3 especially striking.

Finding 3: The Autonomy Ceiling Is Rising Anyway

Organizations were asked whether they would allow an autonomous agent to deploy a code or system change to production based solely on automated evaluation results, with no human-in-the-loop validation.

This is the paradox at the heart of the report. Despite nearly universal distrust of automated evaluation (Finding 2), two-thirds of organizations (66%) either already allow zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to permit it within a year (33%). Only 22% rule it out for the foreseeable future.

The direction is unambiguous: enterprises are moving to let evaluations gate production autonomously — removing the human check — at the very moment they acknowledge those evaluations do not reliably match reality. The autonomy ceiling is rising faster than the assurance beneath it, creating the mechanism by which the false-confidence failures of Finding 1 will scale rather than shrink.

Notably, this push toward autonomy is not confined to smaller companies. Splitting the sample by company size, larger enterprises are slightly further along the path toward zero human review than smaller ones (70% versus 64%) and are slightly more likely to have shipped an evaluation-passing agent that then failed a customer (54% versus 48%). The assumption that large, regulated organizations hold the human in the loop longest is, in this sample, reversed. These figures are directional, drawn from 57 respondents at companies with 2,500+ employees and 100 from smaller organizations.

Finding 4: The Evaluation Stack Is Fragmented and Provider-Led

Respondents were asked which agent reliability or evaluation platform they primarily use. The market has no clear leader, and a substantial share has nothing dedicated.

The evaluation layer is early and unconsolidated. Provider-native tooling leads: OpenAI's native evals and traces (17%) and Anthropic's Claude Console evals (13%) together outweigh any independent platform. However, provider-native tooling is tied at the top by a notable response — 17% of enterprises use no dedicated agent-evaluation tooling at all, a striking gap for organizations shipping agents to customers. Specialist evaluation vendors — DeepEval (12%), Braintrust (8%), LangSmith, Weave, Promptfoo, Langfuse, and Arize — are scattered across single to low double digits, and 11% have built their own solutions. No independent platform has yet become the category standard, leaving most enterprises to evaluate agents with provider-native tools, home-grown scripts, or nothing at all. The fragmentation partly reflects the pace upstream: open-source agent frameworks such as LangChain, AutoGen, and CrewAI proliferated faster than evaluation standards could form, leaving each ecosystem to define its own notions of correctness.

Finding 5: Production Monitoring Rarely Watches Output Quality

Production monitoring for an AI agent can track two fundamentally different things. It can monitor whether the system is functioning — is the agent up and responding, did each request complete, how fast, at what cost, and with any errors. Or it can monitor whether the agent's output is correct — automated checks that evaluate the content of each response as it goes out: did the agent give the right answer, take the right action, and stay within policy.

This distinction matters because a confidently wrong answer is invisible to the first type of monitoring. The request completes, the response is fast, no error is thrown, and every functioning metric reads healthy.

When grouped by what is actually being monitored, the split is stark: 51% of organizations monitor only whether the agent is functioning, while 23% monitor whether its answers are correct. Counting ad-hoc reviewers and those who do not know, roughly three-quarters of organizations run no automated, real-time evaluation of output correctness in production. They can see that the system is up and what it costs, but they take the correctness of its answers on faith.

The technical barrier to closing this gap is significant: evaluating output correctness in real time typically requires a second model — often called an LLM-as-judge — to score each response as it goes out, a technique whose own reliability and consistency limitations echo the very concerns enterprises raise about automated evaluation in Finding 2.

This blind spot is the runtime counterpart to the pre-deployment gap in Finding 1: the same organizations engineering the human out of the deployment decision largely cannot see, in real time, when the deployed agent begins producing wrong answers.

Finding 6: Bought on Cost, Measured on Consistency

Organizations were asked what most influences their choice of an evaluation vendor and what they consider their primary measure of success. Both answers are pragmatic.

Enterprises buy evaluation tooling based on economics and trust it for repeatability. Cost of evaluations (28%) narrowly leads selection criteria, just ahead of ease of integration (27%) and evaluation accuracy (24%). Breadth of observability (13%) and vendor roadmap (4%) matter far less.

On what success looks like, more than a third (36%) name evaluation consistency — receiving the same verdict for the same behavior every time — well ahead of speed of experimentation (19%), reduction in failures (18%), production visibility (13%), and compliance (11%). The emphasis on consistency is revealing: before enterprises can trust an evaluation's verdict, they need it to be stable — the very property whose absence (bias and inconsistency) ranked among the top trust limitations in Finding 2. Satisfaction with current tooling is only moderate, averaging 3.8 on a five-point scale across overall satisfaction, ease of implementation, and value for money.

Finding 7: The Next Dollar Goes to Humans and Observability

Respondents were asked which reliability and evaluation investment will grow most over the next year. The funding is flowing toward closer oversight of agents — including through people.

The second-largest planned investment — behind only production observability — is human review workflows, at 26%. Read against Finding 1, this is the report's quietest contradiction: at the same time two-thirds of enterprises are engineering the human out of the deployment decision, more of them plan to increase spending on human reviewers (26%) than on the automated evaluation pipelines (16%) that would replace them. The zero-human trajectory and the human-review budget are rising within the same companies simultaneously. Indeed, only 8% report that their budget is not increasing.

Taken together, enterprises are hedging: building toward autonomy while investing to watch agents more closely and keeping humans available for the decisions that automated evaluation cannot yet be trusted to make.

Finding 8: A Tooling Reshuffle Is Coming

Organizations were asked whether they plan to adopt a new, additional, or replacement evaluation platform, and which ones they are considering. Few intend to stand pat.

The evaluation market is wide open. While 36% have no plans to change, a clear majority (64%) intend to adopt a new, additional, or replacement platform within twelve months, and 31% within the next quarter. The consideration set highlights where current usage is thinnest: Confident AI's DeepEval leads the platforms being evaluated (20%), ahead of OpenAI's native evals (13%) and Braintrust (9%) — with open-source specialists attracting more interest than their current footprint would suggest.

Given that many enterprises today rely on provider-native tools or nothing at all (Finding 4), this is less a defection than a first genuine wave of tooling adoption — the moment the evaluation layer begins to consolidate. Which platforms earn trust in a market where almost no one yet trusts automated evaluation remains the open question this series will continue tracking.

The Bottom Line: An Evaluation Gap That Autonomy Will Widen, Not Close

Organizations with 100 or more employees are granting AI agents more independence than they trust their evaluations to support. Half have already shipped an agent that passed its evaluations and then failed a customer. Almost none fully trust automated evaluation, primarily because it does not match real-world outcomes. Most monitor production for uptime and cost rather than for whether the agent's answers are correct. Yet two-thirds already allow, or are actively building toward, deploying to production on automated evaluation alone.

This dynamic is unfolding as regulatory frameworks begin formalizing expectations for AI testing. The EU AI Act, which entered into force in August 2024, introduces risk-based obligations including pre-deployment assessment and post-market monitoring for high-risk AI systems. The NIST AI Risk Management Framework, published in January 2023, provides voluntary guidance for measuring, managing, and monitoring AI reliability. Both frameworks underscore the same principle the survey data reveals: the ability to demonstrate trustworthy evaluation is becoming a requirement, not a preference.

The vendor market is early and unsettled: the most common primary evaluation tools are provider-native evals, tied with no dedicated tooling at all, and a clear majority plan to adopt or switch platforms within the year. The next dollar is going to observability and — pointedly — human review, suggesting enterprises sense the gap even as they engineer past it.

At 157 respondents in a single wave, this is a directional read, skewed toward the mid-market. The direction, however, is clear: autonomy is being granted on the strength of evaluations that the people granting it do not yet trust. The evaluation gap is not a coverage problem that more tests alone will close; it is a problem of evaluations that reflect reality and can be trusted to gate it. The open question for later waves is whether assurance catches up to autonomy — or whether false-confidence failures escalate from customer incidents into changes that deploy themselves.

Based on survey responses from 157 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. This is a directional read rather than a precise measurement — the sample is self-selected, not a probability sample, and skews toward the mid-market. Respondents include product and program managers, consultants and advisors, directors of engineering/IT, and CIOs/CTOs/CISOs, among other functions, across technology/software, retail/consumer, healthcare/life sciences, manufacturing, and other industries.