Recent research highlights that many businesses have implemented AI agents without establishing the necessary controls for effective management, often with full awareness. This conclusion stems from five separate surveys conducted by VentureBeat Research in June, encompassing all critical aspects of AI deployment. As a response, these businesses are now looking to enhance their oversight standards and are allocating budgets for that purpose. Across the various control dimensions assessed, between 57% and 68% of respondents indicated plans to switch to new vendors or incorporate additional ones within the next year, with about a third expressing intentions to move forward within the upcoming quarter.
VentureBeat Research evaluated five essential controls an organization must establish to trust an AI agent: identity, evaluation, cost telemetry, the context layer, and orchestration. Identity determines which capabilities an agent is permitted to execute, based on designated credentials. Evaluation assesses the quality of the agent’s outputs. Cost telemetry monitors the operational costs associated with each agent. The context layer provides the critical business data and definitions the agents rely on for their responses. Lastly, the orchestration control plane organizes multi-step tasks effectively. Each of the reports focuses on one of these control mechanisms.
Most AI “agents” currently used are essentially chatbots. A significant 71% of organizations reported that only a quarter or fewer of their “agents” can independently perform complex tasks, and just 10% indicated that genuine agents comprise the majority of their deployments. Those responding to the surveys are well-positioned to know, as 81% hold roles that involve recommending or deciding on AI purchases. Simple chatbots that require human oversight for every response do not need the controls outlined in the remaining reports. On the other hand, a true multi-step agent necessitates the full array of controls, yet many companies struggle to specify which type they have implemented. (Further details are available in the Agentic Orchestration report.)
Trust is lagging behind the autonomy of AI evaluations. Over two-thirds of companies either permit an agent to implement code or system changes based solely on automated evaluation results—without human review—or are working towards doing so within the next year. However, only 5% of these organizations fully trust the evaluation systems in place that enable such actions. Alarmingly, half of the enterprises have deployed agents that passed initial evaluations but later resulted in failures affecting customers in the past year. This highlights the necessity of validating evaluations against real-world outcomes before delegating workflow decisions to machines. (Detailed findings can be found in the Agent Reliability & Evaluations report.)
Entities enabling credential sharing among agents experience more frequent security breaches. Approximately 69% of companies permit some of their agents to share credentials, allowing several agents to operate under a single API key or service account. Organizations implementing credential sharing have reported security incidents at a rate of 63.5%, compared to 40.9% in businesses that assign individual, scoped identities to each agent. To improve security, it is essential to adopt scoped identities for all agents, especially those that interact with production systems. (For further insights, see the Agentic Security & Identity report.)
Suboptimal use of hardware resources is common. Over 80% of enterprises managing their own GPUs noted that they operate at 50% capacity or lower. Additionally, only 44% diligently track the expenses and returns associated with their AI computing. Rather than increasing the number of GPUs, organizations should first focus on improving the utilization rates and cost-effectiveness of their existing hardware. (Complete data is available in the AI Infrastructure & Compute report.)
Agents can deliver incorrect answers with misplaced confidence. In the past six months, 57% of enterprises have traced incorrect yet assertive responses from agents back to deficiencies in their own business context—such as inconsistent metrics, outdated definitions, or missing documentation. Many of these businesses have experienced this issue multiple times, emphasizing the importance of governing the definitions on which agents base their answers, particularly metrics and entities, before expanding their deployment. (Relevant findings are detailed in the Context Layers / RAG report.)
There currently are no dominant players in any of these control areas. Most defaults are linked to the integrated tools provided by major AI platforms that businesses have already adopted. The intent to switch vendors is particularly high in the orchestration domain, where 68% of organizations plan to either adopt new platforms or alter existing ones within the next year, and 34% intend to do so in the upcoming quarter. However, the surveys did not explore whether this investment will be directed towards the built-in tools of these platforms or toward specialized providers, leaving an open question for the market in the coming year.
About this research
VentureBeat Research conducted five parallel surveys in June 2026 as part of its VB Pulse program, focusing on various aspects such as Agentic Orchestration (101 participants), Agent Reliability & Evaluations (157), Agentic Security & Identity (107), AI Infrastructure & Compute (107), and Context Layers / RAG (101), totaling 573 qualified respondents from organizations with a minimum of 100 employees. The samples were self-selected, and while some results should be considered indicative, each report provides a detailed methodology note. The consistent trends across all surveys strongly suggest a unified direction in the findings. VentureBeat sponsors both this research and the VB Transform conference, where these results were presented.
