Enterprises push AI agents into production despite weak eval trust

VentureBeat’s Pulse Research surveyed 157 organizations with at least 100 employees and found a clear operational mismatch: enterprises are giving AI agents more authority while saying the evaluation systems behind those decisions are unreliable. Half of respondents said an agent or LLM feature cleared internal checks in the past year and still caused a customer-facing failure, while 66% already permit or are building toward production deployment based only on automated evaluations.

The survey, conducted in June 2026 as VentureBeat’s Agentic Reliability & Evals tracker, is not a probability sample and should be read directionally. The respondent pool skews mid-market, with 37% from companies with 100 to 499 employees and 27% from companies with 500 to 2,499 employees. VentureBeat said 38% of respondents were final decision-makers for AI purchases, with another 34% influencing or recommending purchases.

Internal tests are not matching production

The main finding is not that companies lack tests. It is that the tests are not predicting real customer outcomes. VentureBeat reported that 50% of organizations running evaluations had deployed an agent or LLM capability that passed internal review and later failed in front of a customer. A quarter said this had happened more than once. Another 36% reported no such incident, while 8% said they do not run pre-deployment evaluations and 6% said they do not track root cause closely enough to know.

Confidence in automated evaluation remains thin. Only 5% of organizations said they fully trust automated evals today. The largest complaint, cited by 29%, was poor alignment between evaluation results and real-world performance. Other concerns included bias or inconsistency at 21%, limited explainability at 18%, and privacy or data-leakage risk at 17%.

Autonomy is still moving ahead

Despite that lack of trust, enterprises are reducing human review in deployment workflows. VentureBeat found that 34% already allow fully automated production deployment for low-risk agents, with no human in the loop, and another 33% are engineering systems to support that within 12 months. Only 22% said they rule out that model for the foreseeable future.

The pattern was not limited to smaller companies. In VentureBeat’s split, companies with 2,500 or more employees were slightly more likely than smaller firms to be on the path toward zero-human review, 70% versus 64%. They were also slightly more likely to have seen an eval-passing agent fail a customer, 54% versus 48%. VentureBeat cautioned that those cuts are directional, based on 57 respondents from larger companies and 100 from smaller ones.

The tooling market is still unsettled

The evaluation stack looks early and fragmented. The most common primary tools were OpenAI’s native evals and traces, used by 17%, tied with having no dedicated agent-evaluation tooling at all. Anthropic’s Claude Console evals followed at 13%. Specialist tools including DeepEval, Braintrust, LangSmith, Weave, Promptfoo, Langfuse and Arize were spread across single-digit to low double-digit shares, and 11% of respondents said they built their own systems.

Production monitoring is also focused more on system health than answer quality. VentureBeat said 51% of organizations monitor whether agents are functioning, such as uptime, latency, errors and cost. Only 23% run automated checks on whether live outputs are correct. That means many organizations can see whether an agent responds, but not whether its response is wrong.

Buying criteria are pragmatic. Cost led vendor selection at 28%, followed by ease of integration at 27% and evaluation accuracy at 24%. The top success metric was consistency, named by 36%, ahead of experimentation speed, fewer failures, production visibility and compliance. Satisfaction with current tooling averaged 3.8 out of five across overall satisfaction, implementation and value for money.

The market may shift quickly. VentureBeat found that 64% plan to adopt, add or replace an evaluation platform within a year, including 31% within the next quarter. DeepEval led consideration at 20%, followed by OpenAI native evals at 13% and Braintrust at 9%.

Spending plans show the hedge. Production observability was the top planned investment area, while human review workflows ranked second at 26%, ahead of automated evaluation pipelines at 16%. Enterprises are preparing for more autonomous agent deployment, but their budgets suggest they still expect people to catch what the software misses.

This story draws on original reporting from VentureBeat.