AI Failures Drive 85% of Burned Enterprises Toward No Approval Deployment
VentureBeat finds 49% of enterprises saw tested AI fail customers, while 85% of burned firms pursue deployments without human approval during July 2026.
Summary
On August 18, 2026, VentureBeat published VentureBeat Intelligence’s self-selected July survey of 108 enterprises with at least 100 employees, down from 157 in June. It found 49% had at least one customer-visible AI agent or LLM feature failure in the previous year after internal approval, versus 50% in June across 265 responses; 24% reported repeats. Complete trust in automated evaluation rose from 5% to 13%, while alignment concerns fell from 29% to 19%. Trust was 4%, 2 of 53, among burned firms versus 24%, 10 of 41, among unburned firms.
Despite failures, 85% of burned enterprises pursued deployment without human approval versus 61% of unburned firms; rejection was 11% versus 24%. Overall, 37% already allowed limited autonomous changes and 30% were preparing within a year, totaling 67%, unchanged from June. Of 106 monitoring responses, 26% used live semantic assertions, 26% traces and 24% gateway metrics; among 40 already automating approvals, only 28% checked answer correctness. Raindrop.ai CTO Ben Hylak said Fortune 100 companies are shrinking eval sets as MCPs and subagents multiply, favoring anomaly detection before and after production.
Primary platforms were OpenAI evals and traces at 18%, Confident AI DeepEval 17%, Braintrust 15%, Anthropic Claude Console and Workbench 12%, no dedicated tool 12%, and internal tools, Promptfoo and LangSmith at 6% each. Braintrust rose from 8% to 15% and DeepEval from 12% to 17%; no platform fell five points. Integration led buying at 39%, up 12 points, ahead of accuracy at 28% and cost at 23%, down from 28%; switching intent fell from 64% to 56%. Investment favored human review at 31%, observability 30%, evaluation pipelines 19% and safety 16%; 6% reported no increase, while burned firms prioritized review 38% versus 24%.
Positives
- Braintrust’s primary platform share rose from 8% to 15%, the survey’s largest and only statistically significant vendor gain.
- OpenAI appeared in 31% of evaluation stacks, DeepEval in 27%, Braintrust in 22% and Anthropic tooling in 20%.
- Internal tooling reached 14% of stacks, while Weights & Biases Weave and open-source Langfuse each reached 11%.
- Average satisfaction reached 3.9 of 5; evaluation consistency led success metrics at 38%, followed by fewer failures and regressions at 20%.
- Human review attracted increased investment from 31% overall and 38% of burned enterprises, indicating recognition that automated checks need a backstop.
Risks & concerns
- 49% of enterprises reported at least one customer-visible failure after internal testing, and 24% had experienced the outcome more than once.
- Only 28% of the 40 enterprises already allowing limited no-approval deployment automatically checked the meaning and correctness of live answers.
- 85% of burned enterprises pursued no-approval deployment, potentially increasing total incidents if deployment volume rises while failure rates remain constant.
- July’s 108-person, self-selected sample used cross-tabs of 40 to 68 respondents, so VentureBeat describes the findings as directional rather than representative.
- Technology participation fell nine points to 14%, while retail and consumer participation rose four points to 19%, changing the month-to-month industry mix.
- Vendor adoption figures do not establish which evaluation platform produces more reliable agents, and the survey did not measure total incident volume.