A VB Intelligence / VB Pulse survey of 107 enterprises finds that most organizations no longer place their bets on a single AI orchestration platform: the median enterprise now operates three different orchestration tools in parallel. This approach reflects not only a desire to avoid vendor lock‑in but also concern that no single vendor can be fully trusted for security and permissioning, so many firms prefer to retain those controls themselves.
Which platforms appear in stacks
Respondents most commonly reported Microsoft AI Foundry/Copilot Studio (present in 70% of stacks), OpenAI Agents SDK (68%), and Anthropic Claude Platform (47%). Other tools in use include Google Enterprise Agent Platform, LangChain/LangGraph, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex. Twenty‑two percent of builders augment vendor tools with custom in‑house orchestration.
This hybrid posture is expected to persist: 53% of respondents expect the primary control plane to be hybrid by the end of 2026. Fourteen percent expect a provider‑managed service, 13% plan a custom in‑house control plane, and 11% are betting on external platforms that abstract away model providers.
Frequent platform churn and future consideration set
More than two‑thirds of respondents plan to change platforms within the year: 15% within three months, 24% in three to six months, and 28% in six to 12 months. Anthropic’s Claude Agent SDK is a top tool under consideration — 43% of builders are exploring it. Around one‑third are evaluating Google’s Enterprise Agent Platform, 31% are focused on custom in‑house orchestration, and 25% are investigating OpenAI options.
Enterprises appear to be deliberately avoiding naming a single winner and are designing for a future where multiple orchestration platforms, models, and agents interoperate across a hybrid control plane.
Satisfaction and implementation challenges
On average, respondents rate their platforms 4.17 out of 5 for overall satisfaction. They rate ease of implementation lower (3.91/5) and value for money even lower (3.63/5). These metrics will be important to watch as orchestration platforms and AI roadmaps mature.
Where enterprises are spending
Buying decisions are guided by several priorities: flexibility (29%), security and permissions (17%), production reliability (15%), and control over agent execution (15%). Only 10% cite model gravity (native alignment with a state‑of‑the‑art base model) as important; 8% name ease of development, 4% total cost of ownership, and 2% latency and memory performance.
Spending patterns reflect emphasis on visibility, security and control: the largest budget shares go to agent monitoring and debugging (31%) and security and permissions enforcement (30%), with workflow tooling receiving 19%. This represents a shift from a wave a month earlier when workflow tooling led orchestration spending.
Enterprises are optimizing for task completion reliability (30%), multi‑step workflow management (27%), developer productivity (23%), and operational stability (13%). Only 7% list end‑user experience as a top priority, indicating orchestration and reliability are the main focuses at this stage.
The visibility problem
Builders’ primary worries about platforms are control and oversight. Top concerns include security and permissioning limitations (37%), vendor lock‑in (23%), limited visibility and observability (22%), and inflexibility around models and tools (16%).
In these early days of AI agents, many enterprises still struggle to control agent token use: one in five firms cannot stop a runaway agent’s spending in real time. To manage spending, 30% rely on native platform controls (built‑in budget caps or throttling) and 25% have built custom gateway middleware to intercept runaway agents. Twenty‑five percent use dynamic routing to move heavy work to lower‑cost models, while 21% still depend solely on reactive monitoring such as post‑hoc logs and therefore lack real‑time kill switches.
Notably, organization size makes little difference in fiscal control maturity: 18% of enterprises with more than 10,000 employees exercise only reactive control, compared with 23% of smaller organizations.
Most deployments are not yet true multi‑step agents
Builders assessed how agentic their stacks are and found steady but limited progress beyond chatbot‑like deployments. Only 2% report that 76–100% of their systems are advanced and largely autonomous; 14% say 51–75% are complex multi‑agent pipelines; and 47% report that 26–50% of their systems are true orchestration. On the other hand, 35% state that only 1–25% of their systems are real orchestration, most deployments remain basic assistants, and 3% are still deploying only chatbots.
These figures align with VB’s June Pulse survey, where 71% said a quarter or fewer of their deployed “agents” can autonomously complete multi‑step work, and just 10% reported agents deployed at scale.
Conclusion
Enterprises are actively building control planes and infrastructure for agents, but many have yet to solve real‑time cost control, visibility, and fully agentic multi‑step automation. The prevailing strategy is hybrid and multi‑platform: organizations are preparing for heterogeneous environments while prioritizing security, permissioning, and spending oversight.
