VB Intelligence surveyed people who work on AI at organizations with 100 or more employees, or for client organizations that size, in its September VB Pulse survey. We asked whether AI that passed pre-release evaluations went on to fail customers, whether AI agents may push changes to production with no human review, and how enterprises monitor agents in production.
Executive summary
Enterprises that let AI agents skip human review are twice as likely to report customer-facing AI failures. Among organizations that deploy autonomous agents and test AI before release, 80% of those letting agents ship some changes without a person's approval reported AI that passed its tests and then failed customers. Among organizations that require human review, 36% did.
We cannot quite say that skipping review causes the failures, however, because we did not ask whether a failure came from a change an agent shipped without review. In August, we should note, enterprises letting agents skip review and those requiring it reported no such differences in failure rates.
Among respondents whose organization deploys autonomous agents, 24% do not expect to permit fully automated deployment in the foreseeable future, down from 42% in August. The share of agent deployers who did not know whether their organization would allow fully automated deployment rose to 12% from 3%.
Across all organizations that run pre-deployment evaluations, 52% of respondents reported at least one failure after passing tests in the past 12 months, and 30% reported more than one. Another 12% of those respondents could not say whether a feature that passed its tests went on to fail customers, up from 2% in August.
Monitoring in production can show when an agent that passed its tests starts giving wrong answers. Yet only 28% of agent deployers name real-time automated quality checks on live agent output as their main monitoring approach. In August, 29% did, a statistically insignificant difference.
OpenAI's native evaluation tooling is used by more respondents than any other platform, at 32%, though that share fell from 59% in August. Most enterprises plan to add or replace evaluation tooling within a year: 68% plan to adopt a new, additional or replacement platform within 12 months.
Finding 1. 80% of enterprises that let AI agents skip human review report a customer-facing AI failure, against 36% of enterprises that require human review
The gap holds among respondents working on AI for their own organization, at 77% against 20%.
Letting an AI agent skip human review means the agent can push a change to production once the change passes automated tests, with no person approving it. With no reviewer, the automated tests carry more of the job of catching a bad change before customers see it.
A test set covers only the cases someone thought to write. A change can pass every test and still give customers wrong answers or break a workflow. A human reviewer is one more chance to catch a bad change before release.
We asked two questions. The first was whether the organization would currently allow an autonomous AI agent to deploy a change to production on automated test results alone, with no human review.
The second was whether, in the past 12 months, the organization had shipped an AI agent or LLM feature that passed its internal tests and then caused a customer-facing failure. Examples given were an incorrect output, a broken workflow or a quality incident.
We compared two groups of enterprises that deploy autonomous agents. Enterprises that let agents skip human review already allow unreviewed deployment for specific low-risk agents or changes. Enterprises that require human review do not allow it today, whether or not they plan to within 12 months.
Almost every enterprise deploying agents, 98%, tests AI before release, so the comparison covers those that do. We left out respondents who did not know their organization's policy.
Finding 1 — Customer-facing failure after passing tests, among enterprises deploying AI agents
Base: organizations that deploy autonomous agents, test before release and could say whether a failure happened; group sizes in parentheses. Four respondents requiring review, and none letting agents skip it, could not say whether a failure happened. Counted as reporting no failure, those four make the figures 80% and 33%.
Of the enterprises that let agents skip review, 80% reported a customer-facing failure after passing tests, against 36% of the enterprises that require human review. The difference is statistically significant.
Of the enterprises letting agents skip review, 48% reported more than one such failure, against 22% of those requiring review.
We did not ask whether a failure came from a change an agent pushed without review, so we cannot say that skipping review causes the failures. We also did not ask how much AI each organization ships.
In August, the two groups did not differ in a statistically significant way: 76% of enterprises letting agents skip review reported such a failure, against 75% of those requiring review.
The share of enterprises requiring review that reported such a failure fell to 36% in September from 75% in August, which is statistically significant.
Finding 2. 24% of enterprises deploying autonomous AI agents do not expect to permit fully automated deployment, down from 42% in August
Among enterprises running AI agents, 65% already let agents skip human review for some low-risk changes or are building toward it. That share has not changed in a statistically significant way since August.
A human reviewer can catch an agent's change that passes automated tests but would break something in production. The cost is slower releases and engineering time spent on approvals. Removing the reviewer lets changes ship faster, and leaves automated evaluation and production monitoring to catch what the tests miss.
We asked respondents whether their organization would currently allow an autonomous AI agent to deploy a code or system change to production based on automated evaluation results alone, with no human-in-the-loop validation.
Finding 2 — Would the organization allow an agent to deploy to production on automated evaluation alone?
Base: 103 qualified respondents. One answer each. The figures in the text leave out respondents whose organization does not deploy autonomous agents (base: 93).
Among September respondents whose organization deploys autonomous agents, 24% said their organization does not expect to permit fully automated deployment in the foreseeable future. In August, 42% of agent deployers gave that answer, and the difference between the two months is statistically significant.
Among the same September respondents, 45% of enterprises running AI agents already let agents skip human review for some low-risk changes. That share has not changed in a statistically significant way since August. Another 19% are engineering their pipelines to allow unreviewed pushes within 12 months. The share answering that they did not know rose to 12% from 3%, also a statistically significant change.
Finding 2 — Among respondents whose organization deploys autonomous agents, August → September
Bases: 118 August and 93 September respondents whose organization deploys autonomous agents. One answer each.
Among agent deployers, the share not expecting to permit fully automated deployment was 20% in July, 42% in August and 24% in September.
Finding 3. 52% of enterprises that test AI before release say AI passed the test but then failed customers, and 30% report more than one such failure
Another 12% of those respondents could not say whether such a failure had happened, up from 2% in August.
Pre-deployment evaluations test an AI agent or LLM feature against a fixed set of cases before release. A feature that passes can still fail on real customer requests the test set did not cover. A failure in production can mean wrong answers that customers act on, broken workflows, and engineering time spent finding and fixing the problem.
We asked respondents whether, in the past 12 months, their organization had deployed an AI agent or LLM feature to production that passed internal evaluations but then caused a customer-facing failure. Examples given were an incorrect output, a broken workflow or a quality incident.
Finding 3 — Customer-facing failure after passing evaluations, past 12 months
Base: 103 qualified respondents. One answer each. The figures in the text leave out respondents whose organization runs no pre-deployment evaluations (base: 101).
Among enterprises that test their AI before release, 52% said they shipped an AI agent or AI feature that passed its tests and then failed customers. That includes 30% that said it happened more than once. Neither share has changed in a statistically significant way since August.
In September, 12% of enterprises that test their AI before release said they do not track whether AI that passed its tests later failed customers, or could not say. That share rose from 2% in August, a statistically significant increase. An enterprise that does not track these failures cannot tell whether its tests are missing them.
We asked whether an organization had at least one such failure in a year, and whether such failures happened more than once. An enterprise that ships fifty agents has more chances to report a failure than one that ships two, so the answers count affected organizations and do not give a failure rate per agent.
Finding 4. Only 28% of enterprises build their AI agent monitoring around checking whether answers are right
Nearly half build it around whether agents run, how fast and at what cost, and an agent can pass all of those checks while giving customers wrong answers.
Production monitoring for AI agents can watch two different things. System health monitoring, such as trace logs and gateway metrics, shows whether an agent ran, how fast and at what cost. Quality monitoring checks whether the agent's answers are correct and can alert a team when quality drops. An agent can look healthy by every system measure while giving customers wrong answers.
We asked respondents how their organization's live production monitoring for AI agents is set up: to evaluate system health, or to evaluate the quality of what the agent produces.
Finding 4 — Production monitoring approach
Base: 93 respondents whose organization deploys autonomous agents. One answer each.
Only 28% of respondents whose organization deploys autonomous agents named real-time automated quality checks on live agent output as the approach their production monitoring is built around. Trace logging or gateway tracking is the main approach for 49%, and neither answer option was worded as checking output quality. In August, 29% of agent deployers named real-time quality checks, a statistically insignificant change. Respondents could name only one main approach, so some organizations may also run quality checks alongside it.
Finding 5. OpenAI's native evaluation tooling is the most widely used and most often primary evaluation platform, though its use fell to 32% of enterprises from 59% in August
Most enterprises plan to add or replace evaluation tooling: 68% plan to adopt a new, additional or replacement platform within 12 months.
An evaluation platform runs an AI agent against test cases, scores the results and keeps traces of each step the agent took. OpenAI builds evaluations and tracing into its developer platform, and Anthropic offers evaluation tools in its console. Independent tools such as Braintrust and Confident AI work alongside them. Without a platform, teams have to build their own way to store test cases and compare scores between releases.
We asked respondents which agent reliability or evaluation platforms their organization uses today, and then which one is their primary platform.
Finding 5 — Evaluation platforms in use and named primary
Base: 103 qualified respondents. Several answers allowed for use, so those shares sum past 100%; one answer for primary. The share of users naming a platform primary is shown for platforms with 10 or more users; each base is that platform's user count.
Use of OpenAI's native evaluation tooling was 32%, more than any other platform, against 17% for Braintrust and 17% for Anthropic's Claude Console. OpenAI's tooling was also named as the primary platform more often than any other, by 27% of respondents, while 12% use no dedicated evaluation tools.
Among respondents whose organization uses OpenAI's tooling, 85% named it as primary. Respondents could choose a primary platform only from the platforms they use. Of OpenAI's users, 42% use no other platform, so OpenAI was their only choice. Among OpenAI users who also use another platform, 74% named OpenAI as primary.
Enterprises that use OpenAI's evaluation tools look like everyone else. Compared with enterprises that do not use OpenAI tools, OpenAI users showed no statistically significant difference on any of five questions:
how often AI that passed its tests went on to fail customers
whether they let AI agents ship changes without human review
whether they build their monitoring around checking whether answers are right
whether they trust automated evaluation
whether they plan to switch evaluation platforms
In August, 59% of respondents said their organization uses OpenAI's tooling, and the fall to 32% is statistically significant. Use of Confident AI fell to 15% from 36%, and use of Opik rose to 7% from 1%, both statistically significant changes. We cannot say why these shares moved.
Most respondents plan to add or replace an evaluation platform: 68% plan to adopt a new, additional or replacement platform within 12 months, including 23% within three months. The share planning a change within three months was 35% in August, a statistically significant fall.
Among those planning a change, no platform under consideration leads by a statistically significant margin. Langfuse was considered by 24%, against 11% in August, but the change is statistically insignificant because it does not hold among respondents working on AI for their own organization.
Finding 5 — Platforms considered by respondents planning a change
Base: 70 respondents planning to adopt a new, additional or replacement platform within 12 months. Several answers allowed, so shares sum past 100%. Platforms named by fewer than five are not shown.
Asked which factor most influenced their choice of evaluation vendor, 41% of respondents with a named primary platform chose ease of integration, more than any other factor. Cost of evaluations and evaluation accuracy followed, at 22% and 18%, a statistically insignificant gap.
Finding 5 — Factor that most influenced vendor selection
Base: 91 respondents who named a primary platform. One answer each.
Respondents who named a primary platform rated it 4.12 out of 5 for overall satisfaction, 3.88 for ease of implementation and 4.02 for value for money. The overall satisfaction rating rose from 3.95 in August.
Finding 6. Only 17% of enterprises say they trust automated evaluation of their AI agents
The rest named a limitation. Data leakage, evaluation bias and poor alignment with real-world results were each named by 18% to 21%.
Automated evaluation uses scripts or another AI model to grade an agent's output, so a team can check far more cases than people could review by hand. The grades are useful only when they match what happens with real users. A team that does not trust its automated grades has to keep people reviewing output, which slows releases and costs reviewer time.
We asked respondents which limitation most reduces their trust in automated agent evaluations today. One of the answers was "we trust automated evaluation today."
Finding 6 — Limitation that most reduces trust in automated evaluation
Base: 103 qualified respondents. One answer each.
Among the limitations, data leakage or privacy concerns were named by 21%, evaluation bias or inconsistency by 19% and poor alignment with real-world outcomes by 18%. The gaps between those three are statistically insignificant.
Lack of explainability was named by 10% of respondents, against 24% in August, a statistically significant change.
What changed since the August Pulse
We asked the questions compared below in the same words in August and September. We count a change as statistically significant only when the test holds on the base named in the note and also among respondents working on AI for their own organization.
What changed in a statistically significant way
What changed: statistically significant
Statistically significant means the test in the methodology confirms the difference on both comparisons. Bases, August then September: agent deployer rows 118 and 93; agent deployers requiring review 69 and 36; failure rows 133 and 101; platform-use rows 136 and 103; satisfaction 115 and 91; other rows 140 and 103.
What did not change in a statistically significant way
What did not change: statistically insignificant
Statistically insignificant means the difference does not pass the test described in the methodology. Bases, August then September: agent deployer rows 118 and 93; agent deployers letting agents skip review 37 and 40; failure rows 133 and 101; platform-use rows 136 and 103; respondents planning a change 87 and 70; vendor factor and rating rows 115 and 91; other rows 140 and 103.
The bottom line: Enterprises that let AI agents skip human review are twice as likely to report a customer-facing AI failure
Among organizations that deploy autonomous agents and test before release, 80% of those letting agents skip human review reported AI that passed its tests and then failed customers. Among those requiring human review and able to say whether a failure happened, 36% did.
In August, by contrast, the two groups reported failure rates that did not vary in a statistically significant way, so the next Pulse will test whether the gap holds.
Among agent deployers, 24% do not expect to permit fully automated deployment in the foreseeable future, down from 42% in August. Only 28% of agent deployers name real-time automated quality checks on live agent output as the main approach of their production monitoring.
In the next Pulse, VB Intelligence will measure three things. The first is whether enterprises letting agents skip human review again report more failures after passing tests than enterprises requiring review. The second is how the share not expecting to permit fully automated deployment compares with September's 24%. The third is how the share of agent deployers naming real-time quality checks compares with September's 28%.
Respondent profile
| Whose organization respondents answered about | Respondents |
|---|---|
| Work on AI for their own organization | 54% (56) |
| Answered the screening questions about a client organization | 46% (47) |
Base: 103 qualified respondents. One answer each.
| Organization size | Respondents |
|---|---|
| 100–250 employees | 42% (43) |
| 251–1,000 employees | 23% (24) |
| 1,001–5,000 employees | 17% (17) |
| 5,001–10,000 employees | 11% (11) |
| 10,001 or more employees | 8% (8) |
Base: 103 qualified respondents. One answer each. For respondents answering about a client, the size is the client organization's.
| Role in AI purchasing | Respondents |
|---|---|
| Final decision maker | 35% (36) |
| Recommender or influencer | 43% (44) |
| User | 9% (9) |
| No involvement | 14% (14) |
Base: 103 qualified respondents. One answer each. For respondents answering about a client, the role is in the client's AI purchasing.
| Job level | Respondents |
|---|---|
| C-suite | 17% (18) |
| VP | 5% (5) |
| Director | 10% (10) |
| Partner or principal | 6% (6) |
| Manager | 37% (38) |
| Individual contributor | 17% (17) |
| Other | 9% (9) |
Base: 103 qualified respondents. One answer each.
| Industry | Respondents working on AI for their own organization |
|---|---|
| Healthcare or life sciences | 16% (9) |
| Technology or software | 16% (9) |
| Manufacturing | 13% (7) |
| Government or public sector | 11% (6) |
| Financial services | 11% (6) |
| Education | 5% (3) |
| Retail or e-commerce | 5% (3) |
| Professional services or consulting | 5% (3) |
| Other industries | 18% (10) |
Base: 56. One answer each. We did not ask respondents answering about a client organization for the client's industry.
Methodology
VB Intelligence fielded this VB Pulse survey in September 2026 and received 115 responses. Of those, 103 qualified for this report. Every figure is based on those 103 respondents unless a smaller base is given in the note under its figure.
We removed responses for these reasons. Seven respondents were analysts, academics or journalists, served customers of a technology vendor, or gave another role. Three said they do not work directly on AI systems. Two answered about a client organization without day-to-day or partial access to its AI systems.
Respondents who work on AI for client organizations answered the screening questions about the client where they spent the most time on AI in the past six months. From the evaluation questions onward, the survey asked about "your organization" and did not tell them whether to answer for the client or for their own firm. For those respondents, answers from that point on may describe either organization. Of the respondents answering about a client, 18 work as independent consultants or at boutique firms with fewer than 100 people, so some of their answers may describe a smaller organization.
With those 18 respondents left out, the Finding 1 comparison reads 79% against 32%, and the Finding 2 share is 21% against August's 42%. Both differences are still statistically significant. Other shares in the report move by 5 points or less.
One comparison does not hold without those 18 respondents. In Finding 5, ease of integration leads cost of evaluations as a vendor selection factor at 36% against 21%, a statistically insignificant gap.
Respondents answering about a client organization were more likely than in-house respondents to report a failure after passing evaluations, 67% against 41%. The Finding 1 comparison holds among in-house respondents, at 77% against 20%. Among respondents answering about a client, the gap is statistically insignificant.
We checked every qualified respondent for answers about their own organization that cannot both be true. We looked for a primary platform not listed as used, "no plans to change" alongside platforms under consideration, and "no dedicated tools" alongside a named tool. No qualified respondent gave any of those three pairs of answers.
Two respondents said their organization lets agents deploy on automated evaluation results alone, yet runs no pre-deployment evaluations. We flag them and keep them in the figures. They fall outside the Finding 1 comparison.
Leaving the two flagged respondents out moves Finding 2 shares by 1 point or less.
Some respondents gave the same score on all three platform ratings. Under our quality rules, one such flag is noted and the respondent is kept. The Finding 1 comparison holds when those respondents are left out.
Each base names the respondents a question applies to. Figures on unreviewed pushes and on production monitoring use respondents whose organization deploys autonomous agents. Figures on failures after passing evaluations leave out organizations that run no pre-deployment evaluations.
The Finding 1 comparison also leaves out respondents who could not say whether a failure happened, and respondents who did not know whether their organization lets agents skip human review. Platforms considered use respondents planning a change, and vendor selection factors and ratings use respondents who named a primary platform.
The platforms-in-use and platforms-considered questions allowed several answers. Every share is a share of respondents, so shares on those questions add up to more than 100%. Every single-answer question was checked, and none received more than one answer.
A difference between two shares, whether between two groups in the same month or between August and September, counts as statistically significant only when a statistical test gives p below 0.05. We use a two-proportion z-test, or Fisher's exact test when either group has fewer than 40 respondents. Average ratings are compared with a Mann-Whitney test.
To rank two answers within one question, we use an exact binomial test on single-answer questions and an exact McNemar test on questions that allowed several answers.
August did not route respondents who work for client organizations to separate screening questions, and did not screen out respondents who do not work directly on AI. So every August comparison was also run on September respondents who work on AI for their own organization, and we count a change as statistically significant only when both tests confirm it.
From the September Pulse, respondents whose answers about their own organization contradict each other are flagged and reported both ways, where August removed them. We reran every August comparison without September's two flagged respondents, and every result stays the same.
Some groups in this report have fewer than 40 respondents. The table below names each group and its base.
| Group with fewer than 40 respondents | Base |
|---|---|
| Agent deployers that test before release, could say and require human review | 36 |
| Of those requiring review: planning to let agents skip it within 12 months, and with no plans to | 17 and 19 |
| In-house agent deployers that test before release and could say, letting agents skip review and requiring it | 22 and 20 |
| Client-path agent deployers that test before release and could say, letting agents skip review and requiring it | 18 and 16 |
| Users of OpenAI's tooling, Anthropic's console, Braintrust, Confident AI, custom in-house tooling and LangSmith | 33, 18, 17, 15, 14 and 11 |
| Users of OpenAI's tooling who also use another platform | 19 |
| Users of OpenAI's tooling who deploy autonomous agents, and those who test before release and could say | 30 and 31 |
| Without the 18 independent and boutique respondents: agent deployers that test before release and could say, letting agents skip review and requiring it | 34 and 28 |
| In-house respondents planning a platform change | 34 |
| August agent deployers that tested before release, could say and let agents skip review | 37 |
| Respondents without identical ratings, letting agents skip review and requiring it | 26 and 30 |
Each base is the number of respondents in the group.
Percentages for groups this small are less precise; for a group of 38, a result could differ by about 15 percentage points either way from the true figure. We report a difference between groups only when a statistical test confirms it is unlikely to be due to chance. The test used for each comparison is listed in this section.
Finding 1 compares the 40 agent deployers that test before release and let agents skip review against the 36 that require human review, among those who could say. Fisher's exact test gives p<0.001.
When respondents who could not say are counted as reporting no failure, the comparison is 40 against 40, and a z-test gives p<0.001. The August comparison, 37 against 69, gives p=1.00 on Fisher's exact test.
The fall among enterprises requiring review in Finding 1 uses Fisher's exact test, p<0.001, and p<0.001 on in-house respondents.
The five comparisons of OpenAI users with other respondents in Finding 5 use Fisher's exact test, and none gives p below 0.05. The change in Finding 2 uses a z-test, p=0.006, and p=0.016 on in-house respondents.
The July figure in Finding 2 comes from the August report, which used respondents whose organization deploys autonomous agents in both months. Respondents are a self-selected group of VentureBeat readers and panel members. The figures describe these respondents and are approximate as a guide to the wider enterprise market.
The sample of 103 respondents gives figures near 50% a margin of about 10 percentage points either way. VentureBeat surveyed a separate sample each month, so month-to-month comparisons describe two different groups of respondents.
We compared many answers between the two months, so a few could pass the test by chance alone.
Published by VB Intelligence. © 2026 VentureBeat. All rights reserved.