Survey: Companies are building AI agent systems faster than they deploy them

In short: A new survey suggests many companies are investing in systems to run AI “agents,” but most of what they have deployed today still acts like basic chatbots.

What's going on

VentureBeat Pulse Research surveyed 101 organizations with 100 or more employees about how they build and manage AI agents. Here, an “agent” means software that can carry out a task in multiple steps, like booking travel by checking rules, comparing options, and confirming details, not just answering a question.

The survey found companies are standardizing on big AI provider platforms for what is often called “agent orchestration,” which is the layer that coordinates steps and tools (like a manager coordinating a team). Anthropic’s Claude platform was the primary choice for 40% of respondents, followed by Microsoft at 18% and OpenAI at 13%.

Companies said they pick these platforms mainly because of the underlying AI model they want to use. They also judge success by whether the agent reliably finishes multi-step work. But when asked to assess what they have actually deployed, 71% said that a quarter or fewer of their “agents” are truly multi-step systems. Many are still single prompt chatbot wrappers.

Companies also showed caution about depending too much on one provider. By the end of 2026, 51% expect a “hybrid control plane,” meaning they will split control between a provider’s tools and their own external layer (like using a landlord’s security system, but keeping your own locks too). Only 6% expect to fully hand control to a provider-managed service.

What to watch

Cost control is a weak spot. More than a quarter of respondents said they cannot stop a runaway agent in real time before costs pile up, which matters as more of these systems move from testing into everyday use.

Source: Venturebeat