VB Pulse AI agents pass tests but half of enterprises ship failures

Enterprise AI is shipping autonomous agents faster than it can verify them. According to the June 2026 VB Pulse survey of 157 qualified enterprise respondents at companies with 100 or more employees, half of enterprises have deployed an AI agent or LLM feature that passed internal evaluations and still caused a customer-facing failure, with one in four reporting more than one such failure. The sample is self-selected rather than a probability sample, which the source itself flags, so the numbers should be read as directional. The disconnect is structural: 66% of respondents already permit some production deployment without human review or are building systems to do so within 12 months, while only 5% fully trust the automated evaluations that drive release decisions. The organizations moving fastest toward zero-human deployment are also the ones shipping agents that fail customers more often.

The reason a passing evaluation is not the same as a working agent lies in what agents actually do. Traditional software tests usually check whether a defined input produces an expected output. An agent chooses its own sequence of steps, calls tools, retrieves data, modifies state and produces different trajectories from one run to the next. The source describes the failure mode in concrete terms: an agent may retrieve the correct account but update the wrong field, draft a valid refund request and send it without approval, or call five tools successfully before a sixth step leaks sensitive information or leaves a workflow incomplete. None of these failures look like an agent that lacks capability. They look like an agent that lacks consistency. The survey's breakdown of distrust in automated evaluation supports that reading: 29% of respondents cite poor alignment with real-world outcomes as the top concern, followed by bias or inconsistency at 21%, lack of explainability at 18%, and data leakage or privacy concerns at 17%. Enterprises are not saying the scoring is too slow or too expensive. They are saying the score does not predict what happens when a real customer or business process encounters the agent in production.

National and vendor-level guidance already treats this as a known problem. NIST's Generative AI Profile notes that measurements gathered in controlled environments may not transfer cleanly to deployment, because agent behavior changes with prompts, users, context and operating conditions. The profile calls for field testing, post-deployment monitoring and clear processes for escalating failures. Anthropic's guidance on agent evaluation draws a related distinction: between measuring whether a system succeeds at least once across repeated attempts and whether it succeeds every time. The first metric is a capability test. The second is a consistency test. When the workflow is customer-facing or operational, only the consistency reading is actually load-bearing. A model that occasionally produces an excellent answer may still be unacceptable for production if the same task fails unpredictably on the next attempt. The implication is that a single passing score is closer to a launch lottery than to a release verdict.

The size of the company appears to amplify the gap rather than close it. The source reports that larger enterprises, those with 2,500 or more employees, are moving toward zero-human deployment fastest, at 70% versus 64% for smaller companies. The same group is also shipping more agents that subsequently fail a customer, at 54% versus 48%. The pattern is not that bigger companies are better at assurance. It is that bigger companies are more aggressive at removing humans from the loop, and the customer-facing failure rate rises in proportion. The source argues that zero-human operation should be earned through demonstrated reliability and bounded by the consequences of failure, not granted by default. Low-risk actions like drafting internal summaries or categorizing documents can tolerate broader autonomy. Financial transactions, customer communications, code deployment, access-control changes and data deletion need stricter thresholds, repeated consistency tests, policy checks, rollback mechanisms and clear human escalation paths, according to the source. The current survey data does not separate which of those categories the failing agents actually touched, which is the breakdown enterprises would need to allocate assurance work where it pays off.

The source frames repeatability as a first-class evaluation metric. The argument is that the same scenario needs to run multiple times, with phrasing and context varied, tool failures tested, and the final business outcome checked for correctness even when the route changes. The evaluation set, the source adds, has to evolve as well: every production incident should become a permanent regression test, and customer escalations, failed tool calls, incorrect approvals and data-handling mistakes should feed back into the pre-deployment suite rather than staying isolated support cases. None of that is in the survey itself, and the source does not report how many enterprises have actually built that feedback loop. What the survey does show is the gap such a loop is meant to close: half of enterprises have shipped a passing agent that failed a customer, and only 5% fully trust the automated evaluations that produced the green light. The economic incentive to move toward greater autonomy is real, and the survey does not dispute that. The data does not yet show whether the assurance layer catches up at the same pace as deployment.

The source frames the next year as a retrofit cycle, with budget shifting toward identity, evaluation, cost, context and orchestration tools that make agent deployments governable. That framing is consistent with the survey's 66% figure, but the survey itself does not show whether retrofit tools actually close the assurance gap or simply provide a checklist for the same untested agents. The 157-respondent sample is self-selected, and the source flags that explicitly, which means the failure rates may overrepresent companies motivated to discuss AI agent problems. The directional read still holds: the organizations moving fastest toward zero-human deployment are also the ones shipping agents that fail customers more often, and the automated evaluations that sign off on those agents are trusted by almost no one. Removing the human from the loop does not remove the uncertainty. It moves the uncertainty into a production decision the organization has already declined to verify before deployment.