Businesses have deployed AI agents ahead of the controls needed to manage them – and make them aware. That’s the average finding across five similar surveys VentureBeat Research fielded in June, covering all layers of the agent stack. Now those businesses are restructuring to meet their standards, and they’re budgeting for it: In each of the five surveys we measured, 57 to 68% of businesses plan to change vendors or add new ones within 12 months, and about a third, depending on the layer, plan to move within a quarter.
VentureBeat’s research measured five controls a business must build before trusting an agent: ownership, testing, cost telemetry, context layer, and orchestration. Identity controls which agent is allowed to do what, under whose credentials. Evaluation determines whether the agent’s work is good. Cost telemetry tracks how much each agent costs to operate. The context layer provides the business data and definitions agents use when responding. And the orchestration control plane coordinates the multi-step agent activity. Each of our five reports measures one of those controls.
Most used "ambassadors" they are labeled chatbots. Seventy-one percent of businesses said a quarter or less of their investment "ambassadors" they can complete a multi-step task on their own; only 10% say true agents do most of what they do. These respondents were put in the know: 81% recommend or decide to purchase AI for their companies. A one-prompt chatbot with human readability for every response doesn’t need controls measured by four other reports. A true multi-step agent needs all of them – and most businesses can’t tell which one they’ve used. (Full results: Agentic Orchestration report.)
Independence is beyond trust in the analysis that leads you. Two-thirds of businesses may already be allowing an agent to push code or system changes to production on automated test results on their own, without human review, or are engineering them within 12 months. Only 5% fully trust analytics to make that call – and half of companies have sent an agent who passed internal audits and caused customer-facing failures in the past year. Before removing a human review from any workflow, the test is compared to production results instead of internal benchmarks. (Full results: Agent Trust and Evals report.)
Companies that allow agents to share information are getting hit more often. Sixty-nine percent of companies allow at least some of their agents to share information – most agents work under a single API key or service account. Organizations that allow data sharing anywhere had a security incident or near miss rate of 63.5% (47 of 74), compared to 40.9% (nine of 22) in companies where each agent has their own identity. Maintenance is a scoped property for all agents, starting with those that affect production systems. (Full results: Security and Identity Report.)
The most expensive hardware in the building operates at half capacity or less. More than eight in 10 businesses running their own GPUs reported 50 percent utilization or less, and only 44% carefully track how much their AI costs and what it returns. The value worth chasing first isn’t more GPUs – the usage and cost per load of those that are already running. (Full results: AI infrastructure and accounting report.)
Agents respond confidently to non-dominant data. Fifty-seven percent of companies have traced a positive, negative agent response in the past six months to their missing or inconsistent business context — incorrect metrics, outdated definitions, missing documents — and many have seen it happen more than once. Controlling agents of interpretations that respond from – metrics and organizations first – must come before measuring the agents that depend on them. (Full results: Content Layers / RAG report.)
No layer has a fixed function: Automation today is the built-in tools shipped with major AI platforms that businesses already use. Changing intent is at the forefront of orchestration itself, with 68% planning to adopt, add, or change platforms within 12 months and 34% within a quarter. Our surveys did not ask in which direction the money goes – towards the built tools of the field or the professionals who challenge them – and that is an open question for the next four parts of this market.
About this study
VentureBeat Research included five similar surveys in June 2026 under its VB Pulse program: Agentic Orchestration (101 respondents), Agent Reliability & Evals (157), Agentic Security & Identity (107), Infrastructure & Compute for AI (107), and Content Layers / RAG totaled 573 respondents (573 respondents) with 100 employees or more. The samples are self-selected, and some findings should be read in a straightforward manner; each report carries its own methodology note. That’s a pattern that supports more than any one percent direction: each survey, independently, points in the same direction. VentureBeat produces both this research and VB Transform, the conference where these reports originated.
