VentureBeat Research published on July 10, 2026, the results of a June survey of 573 technical leaders at companies with at least 100 employees. The study covered five parallel surveys across different layers of the agentic stack?agentic stack: The set of software layers needed to build, deploy, and control AI agents — from the model through orchestration to monitoring.. The central finding: enterprises are deploying AI agents before building the control mechanisms needed to manage them — and they are doing so knowingly.
Key takeaways
- 86% of enterprises running their own GPUs report utilization of 50% or less
- Only 44% rigorously track the actual costs and returns of their AI compute
- Half of enterprises deployed an AI agent that passed internal evaluations but then caused a customer-facing failure
- 69% of companies allow credential sharing among agents — those companies experience security incidents at nearly twice the rate
- 57% of companies traced at least one confident, wrong agent answer to missing or inconsistent business context
The most expensive hardware is idle
Eighty-six percent of enterprises running their own GPUs report utilization of 50% or less. This is buy-side measurement — from the companies that bought and maintain the hardware. The data enters direct dialogue with the ongoing Wall Street debate over whether AI infrastructure investment is overbuilt.
The problem is compounded by a measurement gap: only 44% of companies rigorously track what their AI compute actually costs and returns. The rest are working from estimates. Yet the purchasing process continues: 45% of respondents name an AI-specialized cloud (CoreWeave, Lambda, Crusoe, Nebius) as the option they are most likely to evaluate in the next 12 months — even though fewer than 2% actually use one today. At the same time, 32% of companies are considering non-Nvidia accelerators (AWS Trainium, Google TPUs, AMD), and 28% are looking at next-generation Nvidia GPUs. The data points to a single practical conclusion: before signing new compute contracts, measure what each inference actually costs on hardware already in the building.
Agents or chatbots in disguise?
Seventy-one percent of enterprises report that no more than a quarter of their deployed "agents" can complete a multi-step task autonomously — the rest are chatbots responding to single prompts. Only 10% of companies say true agents represent the majority of their production deployments.
The data matters because widely cited industry studies paint a very different picture: Gartner predicts that 40% of enterprise applications will be integrated with task-specific agents by the end of 2026, while Writer's 2026 survey reports that 97% of executives say their company deployed AI agents in the past year. The key difference: those surveys asked whether companies have deployed something called an AI agent. The VentureBeat survey asked the people running those deployments a harder question: of the agents you have in production, how many can complete a multi-step task without a person driving each step? The gap between those definitions determines what control mechanisms are actually needed — and what bills will arrive.
Control loses pace with deployment
Thirty-four percent of enterprises already allow AI agents to push changes to production environments based on automated evaluations alone, with no human review. Another 33% are actively building those pipelines for the next 12 months. The paradox: only 5% fully trust the automated evaluations that would make that decision.
The distrust is earned. Half of enterprises deployed an agent that passed internal evaluations and then caused a customer-facing failure — and one in four experienced that scenario more than once. The most commonly cited weakness in current evaluation systems: "poor alignment with real-world outcomes" — 29% of responses. After an agent ships to production, only 23% of companies run real-time quality checks on the answers it generates.
Security: whoever holds the keys carries the risk
Sixty-nine percent of companies allow credential sharing somewhere in their agent fleet — multiple agents operating under one API key or service account. The correlation with security incidents is clear: companies with credential sharing anywhere in their fleet experienced a security incident or near-miss at a 63.5% rate, versus 40.9% at companies where every agent has its own scoped identity.
Why does this matter?
The VentureBeat survey provides a rare measurement from the buyer side — not from technology vendors or analysts. It establishes something that industry forecasts consistently obscure: the gap between declared agent deployment and genuine autonomy is large, while cost and operational discipline lags well behind deployment pace.
For companies investing in GPU infrastructure, the 86% figure is a direct management signal: before the next hardware purchase, it is worth knowing what each inference costs on the equipment already owned. For companies building agent-driven workflows, the evaluation gap between internal benchmarks and production outcomes has real financial and reputational consequences — half of respondents have already experienced it.
The results also confirm a structural gap in the market: there are no established vendors of AI agent control tooling. In each of five technical layers, the most common answer was either the model provider's built-in tools (Anthropic, OpenAI, Google, Microsoft, AWS) or no dedicated tooling at all. The agent control tooling market remains largely open.
What's next?
- VentureBeat will present full reports from all five trackers at VB Transform, July 14–15, 2026, in Menlo Park — results will include action recommendations for each layer of the agentic stack
- The next survey wave (Q3 2026) will measure whether companies followed through on their stated plans: assigning scoped identities to agents, testing evaluations against production data, increasing GPU utilization
- Growing availability of open-weight?open-weight: A model whose weights are publicly available to download and run, though its license may restrict how it is used. models (GLM-5.2 under MIT, Hy3 under Apache 2.0) may shift vendor dependency dynamics — 51% of companies expect to move to a hybrid control model by the end of 2026
