Enterprises deployed AI agents before building the controls to manage them, and they did it knowingly. That is the central finding from VentureBeat Research’s June survey of 573 technical leaders at companies with 100 or more employees.

The numbers paint a picture of an industry spending heavily on infrastructure it cannot fully use and labeling chatbots as agents it cannot properly govern.

86% GPU Utilization Below Half Capacity

Eighty-six percent of enterprises running their own GPUs report utilization of 50% or less, according to the survey. Only 44% rigorously track what their AI compute actually costs and returns. Everyone else is estimating.

The measurement gap compounds the waste. While Wall Street debates whether the AI buildout is overbuilt, the enterprises doing the buying have already answered: the most expensive hardware in their buildings sits mostly idle.

That has not slowed the shopping. Forty-five percent of surveyed enterprises say they plan to evaluate AI-specialized clouds (CoreWeave, Lambda, Crusoe, Nebius) in the next 12 months, but fewer than 2% currently use one. Meanwhile, 32% are considering non-Nvidia accelerators (AWS Trainium, Google TPUs, AMD), compared to 28% looking at next-generation Nvidia GPUs.

Most “Agents” Are Chatbots in Disguise

Seventy-one percent of enterprises say a quarter or fewer of their deployed “agents” can complete multi-step tasks without a human driving each step. Only 10% report that true agents constitute the majority of what they run. The rest are single-prompt chatbots repackaged under the agent label.

Gartner predicted that 40% of enterprise applications would integrate task-specific AI agents by end of 2026, up from under 5% in 2025. The same firm also warned that calling AI assistants “agents” is a common misconception it labeled “agentwashing.”

VentureBeat’s survey asked a harder question than most industry surveys: not whether companies have deployed something called an AI agent, but how many of those deployed agents can actually complete multi-step work autonomously. The gap between the adoption headlines and the operational reality is significant.

Governance Deployed After the Fact

The survey found enterprises are now retrofitting governance across five control layers: agent identity and credentials, output evaluation, cost telemetry, business context management, and orchestration control planes. Roughly six in ten plan to switch or add vendors in each layer within 12 months.

Fifty-four percent reported an agent security incident or near-miss in the past year. Sixty-nine percent allow credential sharing somewhere in their agent fleet, where multiple agents operate under one API key or service account. Organizations with credential sharing experienced security incidents at a 63.5% rate, versus 40.9% where every agent has its own scoped identity.

Two-thirds of enterprises either already allow agents to push code to production based on automated evaluations alone (34%) or are engineering toward it (33%). Five percent fully trust those evaluations. Half shipped an agent that passed internal evaluations and then caused a customer-facing failure in the past year.

The Business Context Problem

Fifty-seven percent traced at least one confident, wrong agent answer to missing or inconsistent business context: wrong metrics, stale definitions, absent documents. Only 25% run a governed semantic layer in production. Another 34% are building one. Forty-one percent have not started.

What Changes Next

The survey signals a shift in where value accrues in the agentic stack. The debate is moving from “which model is best” to “who controls the control plane.” Fifty-one percent of enterprises plan hybrid control planes combining provider-native and external orchestration by end of 2026. Vendor lock-in overtook security as the top concern about provider-controlled orchestration between spring and June survey waves, a shift VentureBeat attributes partly to the three-week loss of Claude Fable 5 access following a June 12 Commerce Department export order.

For orchestration and compliance tooling companies, the data suggests the market is opening faster than anyone expected. For enterprises, the takeaway is more sobering: measure what you own before buying more, scope every agent’s identity individually, and govern your business definitions before scaling the agents that depend on them.