Enterprises are deploying AI agents at a pace that has left governance controls in the dust — and they know it. Five parallel surveys fielded by VentureBeat Research in June 2026, covering every layer of the agentic stack, reveal a stark reality: organizations are rushing agentic systems into production while retrofitting the identity, evaluation, cost, context, and orchestration controls that should have been in place from the start. The pattern is consistent across all five reports, with 57 to 68 percent of enterprises planning to switch or add vendors within 12 months for each control layer, and roughly a third acting within the quarter. This is not a story of ignorance — it is a story of deliberate risk-taking in pursuit of competitive advantage, and the market is now scrambling to catch up with its own ambitions.
The Five Controls Enterprises Are Rushing to Retrofit
VentureBeat Research measured five distinct control layers that an enterprise must build before it can trust an AI agent with production workflows. These are identity, evaluation, cost telemetry, the context layer, and orchestration. Each layer addresses a specific failure point: identity governs which agent is allowed to do what, and under whose credentials. Evaluation determines whether an agent’s output is accurate and reliable. Cost telemetry tracks what each agent actually costs to run — something most enterprises cannot yet do with precision. The context layer supplies the business data, definitions, and metrics that agents draw on when they answer questions or take actions. Orchestration coordinates multi-step agent work, ensuring that sequences of tasks execute correctly and that failures are handled gracefully. Each of the five reports focuses on one of these controls, and the aggregate findings paint a picture of systemic underinvestment followed by urgent remediation.
Most Deployed “Agents” Are Chatbots Wearing the Label
The surveys reveal a critical definitional gap. Seventy-one percent of enterprises said that a quarter or fewer of their deployed “agents” can complete multi-step work autonomously. Only 10 percent reported that true multi-step agents constitute the majority of their deployments. These respondents are positioned to know: 81 percent recommend or decide AI purchases at their companies. A single-prompt chatbot that requires human review for every answer needs none of the advanced controls that a true agent requires. A multi-step agent, by contrast, needs all of them — and most enterprises cannot confidently say which type they have deployed. This finding, detailed in the Agentic Orchestration report, suggests that many organizations are overstating their agentic maturity. The gap between marketing language and operational reality is wide, and it has implications for budgeting, staffing, and security.
Autonomy Is Outrunning the Evaluations That Should Gate It
Perhaps the most alarming finding concerns the relationship between agent autonomy and evaluation trust. Two-thirds of enterprises either already allow an agent to push code or system changes to production based solely on automated evaluation results — with no human review — or are actively engineering toward that capability within 12 months. Yet only 5 percent of those same enterprises fully trust the evaluations that would make that call. This is not a small disconnect; it is a chasm. Half of all enterprises shipped an agent in the past year that passed internal evaluations only to cause a customer-facing failure. The implication is clear: internal evaluation benchmarks are not aligned with production reality. Before removing human review from any workflow, the Agent Reliability & Evals report advises testing evaluations against production outcomes rather than internal benchmarks. The industry is building autonomous systems on evaluation foundations that are not yet trustworthy.
What is the agent evaluation gap? The agent evaluation gap is the discrepancy between how well an agent performs in internal tests and how it performs in production. VentureBeat Research found that half of enterprises have experienced a customer-facing failure from an agent that passed internal evaluations, indicating that current evaluation methodologies are insufficient to guarantee reliability. Closing this gap requires moving beyond static benchmarks to continuous monitoring and feedback loops that reflect real-world conditions.
Credential Sharing Is a Root Cause of Security Incidents
Identity management for AI agents is in a primitive state. The Agentic Security & Identity report found that 69 percent of companies let at least some of their agents share credentials — multiple agents operating under one API key or service account. This practice is directly correlated with security incidents. Organizations that allow credential sharing anywhere experienced a security incident or near-miss at a rate of 63.5 percent (47 out of 74), compared with 40.9 percent (9 out of 22) at companies where every agent has its own scoped identity. The remedy is straightforward: scoped identity for every agent, starting with those that touch production systems. Yet even as enterprises acknowledge the risk, many continue to share credentials out of convenience or because their existing infrastructure does not support granular identity management at scale. The market for agent identity solutions is wide open, and the switching intent data suggests that vendors who can deliver simple, scalable scoped identity will find eager customers.
GPU Utilization Hovers at 50 Percent or Less
Hardware spending is another area where governance is lagging deployment. More than eight in 10 enterprises that run their own GPUs reported utilization of 50 percent or less. Only 44 percent rigorously track what their AI compute actually costs and what it returns. The AI Infrastructure & Compute report argues that the number worth chasing first is not more GPUs — it is the utilization and per-workload cost of the ones already running. Enterprises are buying infrastructure faster than they can measure its cost-effectiveness. This pattern mirrors earlier cycles in cloud computing, where overprovisioning was common before FinOps practices matured. The same maturation is now needed for AI infrastructure, but the speed of agent deployment is outpacing the discipline required to manage it efficiently.
Agents Answer Confidently from Data Nobody Governs
The context layer — the business data and definitions that agents draw on — is perhaps the most overlooked control. The Context Layers / RAG report found that 57 percent of enterprises traced a confident, wrong agent answer in the past six months to missing or inconsistent business context. Wrong metrics, stale definitions, absent documents — these were the root causes. Most of those enterprises saw the problem more than once. An agent that confidently gives a wrong answer because it is operating on outdated financial definitions or incomplete product catalogs is not an AI problem; it is a data governance problem. Governing the definitions agents answer from — metrics and entities first — must come before scaling the agents that depend on them. The market for context layer solutions is nascent, but the need is acute. Enterprises are beginning to invest in knowledge graphs, metadata management, and RAG pipelines that keep agents anchored to trusted data.
No Incumbent Has Won Any Control Layer
A striking finding across all five surveys is the lack of entrenched incumbents in any control layer. The default tools today are the built-in capabilities that ship with the major AI platforms enterprises already use — think of the identity and evaluation features embedded in cloud AI services or the orchestration tools bundled with foundation model providers. But switching intent is high across the board, running highest in orchestration itself, where 68 percent plan to adopt, add, or replace platforms within 12 months, and 34 percent within the quarter. The surveys did not ask whether that money moves toward the platforms’ built-in tools or toward the specialists challenging them. That open question defines the next four quarters of this market. Vendors who can demonstrate measurable improvements in reliability, security, or cost efficiency will have a window of opportunity as enterprises actively shop for replacements.
Historical Context: How We Got Here
The current state of agent deployment mirrors earlier technology cycles. In the early days of cloud computing, organizations moved workloads to public clouds before they had cloud security or cost management in place. The same pattern repeated with mobile apps, where security and privacy controls lagged deployment. Now it is happening with AI agents. The difference is the pace: agents can make decisions and take actions in seconds, compounding risk faster than earlier technologies. The regulatory landscape is also different. While GDPR and other privacy regulations provided a framework for data governance, there is no equivalent for agent autonomy. Enterprises are effectively writing their own rules as they go, and the surveys show that many of those rules are not yet good enough.
Market Implications: A Window for New Vendors
The switching intent data signals a market in flux. Every control layer is seeing active evaluation of alternatives. For identity, vendors that offer granular, policy-based agent identity management are likely to gain traction. For evaluations, companies that can production-validate agent behavior and provide continuous monitoring will fill a clear gap. Cost telemetry tools that integrate with existing FinOps platforms and provide per-agent cost breakdowns will be in demand. Context layer solutions that connect to enterprise data catalogs and maintain freshness are needed. And orchestration platforms that can handle multi-step workflows across diverse agent types will find buyers. The next 12 months will determine whether the major AI platforms can extend their built-in tools to cover these needs adequately, or whether a new wave of specialized vendors will capture the market.
Strategic Recommendations for Enterprises
For enterprises already deploying AI agents — or planning to — the research offers a clear roadmap. First, audit what you are calling an agent. If it is a chatbot with human-in-the-loop, do not invest in advanced orchestration and full autonomy controls. If it is a true multi-step agent, invest in all five control layers. Second, prioritize identity. Scoped credentials for every agent, starting with production-facing ones, will immediately reduce security risk. Third, align your evaluation strategy with production outcomes. Before removing human review, validate that your internal benchmarks predict real-world performance. Fourth, govern context. Ensure that the data your agents draw on is authoritative, current, and consistent. Fifth, measure compute utilization and cost per workload. Do not buy more GPUs until you understand what the existing ones are doing. Finally, prepare for vendor switching. The market is fluid, and the tools you start with are unlikely to be the tools you end with.
The Open Question: Platform Versus Specialist
The surveys leave one question unanswered: Will enterprises migrate toward the built-in tools of the big AI platforms, or will they adopt specialized vendors for each control layer? The answer will depend on how quickly the platforms mature their offerings and how urgent the pain points become. If an enterprise is already running a major cloud AI platform, it may prefer to use that platform’s identity and evaluation tools rather than integrate a third party. But if those tools prove insufficient — as the survey data suggests they often are — the switching intent will materialize into actual vendor changes. The next year will be a proving ground. The enterprises that move first to close their governance gaps will be the ones that can safely scale their agent deployments. Those that delay will face the accumulating costs of incidents, inefficiency, and lost trust. The research is clear: the foundation for trustworthy AI agents is not yet built, but the blueprint is now available. The question is which organizations will choose to follow it.
