Enterprise IT deployed autonomous AI agents ahead of the necessary security and evaluation controls, leaving organizations scrambling to retrofit governance layers. According to VentureBeat Research’s June 2026 VB Pulse surveys of 573 organizations, 57 to 68 percent of enterprises plan to switch or add vendors across five core control layers within the next 12 months.
Most deployed tools are chatbots wearing an agent label. Survey data shows that 71 percent of enterprises report a quarter or fewer of their deployed systems can complete multi-step workflows independently. Only 10 percent state that true agents comprise the majority of their operational stack.
A single-prompt chatbot interacting with a human reviewer needs almost none of the complex guardrails required by fully autonomous systems. A true multi-step agent executing tasks asynchronously demands rigorous oversight across identity, evaluation, cost tracking, context layers, and orchestration.
The Five Control Layers of Agentic Infrastructure
VentureBeat Research broke down enterprise agent maturity into five distinct control layers that organizations must establish to run reliable architectures:
- Identity: Governs permissions, determining which agent can execute specific tasks under defined credentials.
- Evaluation: Measures the accuracy and quality of agentic work before it impacts production systems.
- Cost Telemetry: Tracks granular runtime expenses for every individual agent workload.
- The Context Layer: Supplies enterprise databases, API definitions, and canonical business documentation.
- Orchestration: Manages the control plane coordinating multi-step autonomous workflows.
No single layer currently features an entrenched software incumbent. The default options remain built-in tools shipped directly by major cloud platforms. Orchestration exhibits the highest churn intent: 68 percent of surveyed enterprises plan to adopt, add, or replace platforms within a year, with 34 percent targeting a switch within the current quarter.
Autonomy Outrunning Reliability and Evals
Two-thirds of enterprises either already permit an agent to push code or system changes directly to production based solely on automated evaluation results, or they are actively engineering toward that capability within 12 months. Yet, only 5 percent fully trust the underlying evaluations making those deployment calls.
Half of the surveyed organizations shipped an agent that passed internal evaluation benchmarks only to cause a customer-facing failure over the previous year.
Sixty-nine percent of companies permit at least some of their autonomous agents to share credentials, forcing multiple software instances to operate under a single API key or service account.
Organizations allowing credential sharing experience security incidents or near-misses at a rate of 63.5 percent (47 out of 74 respondents). Conversely, companies enforcing scoped identities for every agent reported incidents at a significantly lower rate of 40.9 percent (9 out of 22 respondents).
Underutilized Hardware and Ungoverned Data Context
More than eight in ten enterprises running dedicated graphics processing units report utilization rates of 50 percent or less. Only 44 percent rigorously track the actual costs and returns generated by their AI compute workloads.
Fifty-seven percent of enterprises traced a confident, incorrect agent response in the prior six months to missing or inconsistent business context, such as outdated metrics, stale definitions, or absent documentation. Most of those organizations experienced this data drift multiple times.
Governing the underlying data definitions—prioritizing business metrics and entity structures—must precede any large-scale deployment of agents that rely on retrieval-augmented generation pipelines to function accurately.
What This Means for Enterprise IT
The data from the June 2026 VB Pulse surveys points to a unified structural reality: enterprise IT built out the cognitive tier of AI before engineering the administrative floor.
