Enterprise AI agent deployment is running far ahead of the governance systems needed to manage it, according to a wide-ranging June survey from VentureBeat Research. The study of more than 570 technical leaders found that companies are deploying autonomous AI agents even as basic controls around security, cost, and reliability remain unfinished.
Fifty-four percent of companies reported an agent-related security incident or near-miss in the past 12 months. Twenty-seven percent said they only discover what an agent costs when the invoice arrives, with no per-agent budget or spending ceiling in place. The findings cut across five layers of the AI stack: identity, evaluation, cost telemetry, context, and orchestration.
The data also reveals a massive hardware utilization gap. Eighty-six percent of enterprises running their own GPUs report utilization at 50 percent or less — a finding that directly undercuts the narrative that the AI infrastructure buildout is efficiently deployed. Only 44 percent of companies rigorously track what their AI compute actually costs and returns; everyone else is estimating.
Despite the underutilization, enterprise procurement continues. Forty-five percent said they are likely to evaluate an AI-specialized cloud provider in the next year, while roughly a third are considering non-Nvidia accelerators as a hedge. The survey suggests enterprises should measure what they already own before committing to new compute contracts.
On the agent front, most deployed systems are still basic. Seventy-one percent of enterprises said a quarter or fewer of their agents can complete multi-step tasks independently — the rest are essentially single-prompt chatbots. Only 10 percent reported that true multi-step agents form the majority of what they run.
The broad picture is an industry that has rushed to deploy AI capabilities and is now scrambling to retrofit the management and governance layers that should have been in place from the start.
Source: VentureBeat
