By admin 
The landscape of enterprise Artificial Intelligence (AI) deployment is undergoing a significant transformation, moving away from a singular reliance on one orchestration platform towards a more diversified, multi-vendor strategy. Contrary to the expectation of a dominant platform emerging to manage complex AI workflows, the typical enterprise is now actively utilizing three distinct orchestration platforms simultaneously. This deliberate fragmentation is not an accidental outcome of vendor competition, but rather a strategic response driven by a fundamental lack of complete trust in any single vendor to govern the entirety of their AI operations. This crucial insight, derived from VB Pulse data, highlights a deep-seated concern within organizations regarding vendor capabilities in security, permissioning, and overall control.
The primary motivation behind this multi-platform approach extends beyond the conventional desire to avoid vendor lock-in and maintain operational flexibility, although these remain significant factors. A pervasive undercurrent of uncertainty and even distrust surrounds the security and permissioning functionalities offered by current AI orchestration vendors. Enterprises are increasingly asserting their need to implement and enforce their own stringent security protocols and access controls, rather than ceding this critical responsibility to third parties. This desire for granular control is shaping how organizations build and manage their AI infrastructure.
While Microsoft’s AI Foundry/Copilot Studio currently leads in primary adoption, holding a significant presence in 70% of enterprise AI stacks, the future trajectory is being heavily influenced by Anthropic. Anthropic’s Claude Platform is emerging as a dominant force in terms of what enterprises are actively considering for future adoption, leading by a considerable margin. However, despite these platform shifts, enterprises continue to grapple with persistent challenges. Foremost among these are issues related to token usage and a distinct lack of comprehensive visibility into the true cost of agent deployment and operation.
These findings are part of an ongoing, in-depth analysis by VB Intelligence, meticulously examining the practical realities of AI deployment and utilization within enterprises. This research delves into the platforms organizations are choosing, the decision-making criteria guiding these choices, their prioritization of AI capabilities, their expectations for AI performance, their strategies for cost management, and ultimately, whether their deployed "agents" are truly autonomous and capable of complex, multi-step tasks, or if they remain fundamentally sophisticated chatbots merely labeled as agents.
VB Intelligence’s methodology involves gathering direct feedback from individuals on the front lines of AI development and deployment. This includes invaluable insights from software and machine learning engineers, product and program managers, and VPs and directors of data, AI, and analytics. These are the builders and strategists who are actively navigating the complexities of enterprise AI, providing a ground-level perspective that complements broader market trends.
Concerns Around Retaining Visibility and Control Take Center Stage
The data from a survey encompassing 107 enterprises paints a clear picture: agentic orchestration has decisively adopted a pluralistic approach. The overwhelming majority of organizations are actively eschewing commitment to a single model. A substantial 85% of enterprises are leveraging two or more orchestration tools, with a significant 64% actively managing three. Only a modest 15% of organizations are currently operating with a single orchestration platform.
Examining the current stack, Microsoft AI Foundry/Copilot Studio is the most prevalent, appearing in 70% of deployments. Close behind is OpenAI’s Agents SDK, utilized by 68% of enterprises. Anthropic’s Claude Platform follows, integrated into 47% of AI architectures. Beyond these market leaders, builders are also incorporating Google’s Enterprise Agent Platform, the open-source LangChain/LangGraph framework, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex to varying degrees. Furthermore, a notable 22% of builders are augmenting vendor-provided tools by developing and deploying custom, in-house orchestration solutions, underscoring the demand for tailored control.
This trend towards hybridability is not a fleeting phenomenon; it is projected to accelerate. More than half of the respondents (53%) anticipate that the primary control plane for their AI operations will be hybrid in nature by the end of 2026. Looking further out, 14% expect to rely on provider-managed services, 13% are planning for a fully custom in-house control plane, and 11% are placing their bets on external platforms that offer abstraction layers away from direct model providers, further diversifying their control mechanisms.
In alignment with this multi-platform strategy, a significant majority of respondents—more than two-thirds—plan to alter their orchestration platform choices within the next year. Specifically, 15% are considering changes in the immediate three months or sooner, 24% plan to make adjustments within three to six months, and 28% are looking at a six to twelve-month timeframe. Among the platforms generating considerable interest for future adoption, Claude Agent SDK from Anthropic is a top contender, with 43% of builders actively exploring its capabilities. Roughly one-third of respondents are evaluating Google’s Enterprise Agent Platform, another 31% are focused on developing custom in-house orchestration solutions, and 25% are investigating OpenAI’s evolving suite of tools.
This deliberate avoidance of a single "winner" appears to be a learned lesson from the challenges and limitations experienced during the early days of cloud computing, where vendor lock-in became a significant impediment to agility. Enterprises today are strategically designing their AI futures to accommodate a collaborative ecosystem where multiple orchestration platforms, diverse AI models, and various agents can seamlessly interoperate across a unified, albeit hybrid, control plane.
Despite the complexities of managing multiple platforms, overall satisfaction with the currently deployed tools remains relatively high, with respondents rating them an average of 4.17 out of 5. However, when it comes to the ease of implementation, satisfaction dips to 3.91 out of 5, and value for money receives a lower rating of 3.63 out of 5. These metrics suggest that while organizations are finding utility in their current orchestration solutions, there is considerable room for improvement in terms of operational efficiency and cost-effectiveness. As orchestration platforms mature and AI roadmaps evolve, it will be crucial to monitor these satisfaction scores for insights into vendor progress and enterprise needs.
Where Enterprises are Strategically Allocating Their AI Investments
The logic behind enterprise AI purchasing decisions has evolved into a complex interplay of several critical factors. While flexibility remains a paramount consideration, cited by 29% of respondents, other key drivers are gaining prominence. Security and permissions are now viewed as crucial, with 17% of enterprises prioritizing these aspects. Production reliability (15%) and the ability to exert granular control over agent execution (15%) are also significant factors influencing buying decisions. Interestingly, "model gravity"—the inherent alignment of an orchestration platform with a state-of-the-art base model—is a relatively minor consideration for most, identified as important by only one in ten respondents. Ease of development is cited by 8%, total cost of ownership by 4%, and latency and memory performance by a mere 2%. This indicates a strong bias towards operational robustness and governance over raw model performance in the current enterprise AI strategy.
The allocation of spending further reflects this enterprise emphasis on visibility, security, and control. Builders are directing the largest portions of their investment towards agent monitoring and debugging (31%) and security and permissions enforcement (30%). Workflow tooling accounts for another significant 19% of spending. This represents a notable shift from VentureBeat’s prior analysis conducted a month earlier, where workflow tooling had outright led orchestration spending. This recalibration suggests that enterprises are now prioritizing the foundational elements of secure and observable AI operations before expanding their investment in complex workflow automation.
When it comes to optimizing their AI deployments, enterprises are largely focused on achieving task completion reliability (30%), efficient multi-step workflow management (27%), and enhanced developer productivity (23%). Operational stability rounds out the top priorities at 13%. Notably, end-user experience is currently a low priority for only 7% of respondents. This indicates that at this stage of AI adoption, many organizations are still heavily invested in establishing the core orchestration infrastructure and capabilities, with the refinement of end-user experiences taking a backseat. The implication is that once robust and reliable AI workflows are firmly in place, the focus is likely to shift towards optimizing the user interface and overall interaction for the end consumer. This phased approach suggests a pragmatic understanding of the AI development lifecycle, where foundational capabilities must be solidified before advanced user-centric features are prioritized.
The Pervasive Visibility Problem in AI Agent Operations
The most significant concerns for builders when selecting AI orchestration platforms revolve around achieving comprehensive control and oversight. There is a strong reluctance to grant vendors the power to restrict an organization’s ability to monitor and understand the precise actions and operations of their AI agents on any given platform. Key factors at the forefront of these concerns include security and permissioning limitations (cited by 37% of respondents), the specter of vendor lock-in (23%), limited visibility and observability into agent behavior (22%), and inflexibility concerning the integration of diverse models and tools (16%). This indicates a strong preference for open, transparent, and customizable AI environments.
Compounding these concerns, in these nascent stages of AI agent development and deployment, enterprises continue to struggle with effectively controlling agent token usage. A concerning one in five organizations still lacks the capability to halt runaway agent spending in real-time, leaving them vulnerable to unexpected and potentially substantial cost overruns.
To mitigate these financial risks, builders are employing a variety of strategies. Thirty percent of respondents rely on native platform controls, such as built-in budget caps and throttling mechanisms provided by the orchestration platforms themselves. An additional 25% have opted to develop custom gateway plumbing, essentially implementing proxy middleware solutions designed to intercept and manage potentially excessive agent activity.
A quarter of respondents utilize dynamic routing strategies, a technique that involves intelligently offloading computationally intensive tasks to lower-cost AI models. However, 21% of enterprises still depend solely on reactive monitoring methods, such as analyzing post-hoc logs. This group of organizations lacks any real-time "kill switches" or immediate intervention capabilities to prevent cost escalations, relying instead on retrospective analysis to identify and address issues.
An intriguing finding emerges when examining the correlation between organization size and fiscal control maturity. Unlike previous observations, the size of an enterprise appears to have minimal impact on its ability to effectively manage AI spending. A comparable 18% of enterprises with 10,000 or more employees exercise only reactive control over their AI costs, mirroring the 23% of smaller organizations that also rely solely on post-hoc monitoring. This suggests that the challenge of implementing robust, real-time cost control mechanisms for AI agents is a universal one, affecting organizations of all scales. Clearly, while enterprises are acutely aware of the financial implications of AI agent deployment, many have not yet fully instrumented their technology stacks with the necessary tools and processes to effectively rein in expenditure.
The Gradual Evolution Towards True Multi-Step Agents
When builders were asked to provide an honest assessment of their current technology stacks, a consensus emerged: the concept of "agents" is steadily progressing beyond simple chatbots that have been superficially enhanced with advanced terminology. The journey towards truly autonomous and multi-step capable agents is ongoing, with a spectrum of deployment maturity observed across enterprises.
The survey data reveals a nuanced distribution of AI agent capabilities. A small but significant segment of respondents (2%) report that between 76% and 100% of their AI systems are advanced, largely autonomous, and capable of complex operations. A larger group, comprising 14% of respondents, indicates that 51% to 75% of their systems are dedicated to complex, multi-agent pipelines, suggesting a high degree of interconnectedness and collaboration between AI components. Furthermore, 47% of organizations report that 26% to 50% of their systems are engaged in true agentic orchestration, signifying a substantial investment in building sophisticated AI workflows.
On the other end of the spectrum, 35% of respondents indicate that only 1% to 25% of their systems represent true orchestration, with the majority of their deployed AI solutions still functioning as basic assistants. A small, yet present, 3% of organizations are still exclusively deploying rudimentary chatbots, indicating that some enterprises are in the very early stages of AI adoption. This finding aligns with VentureBeat’s June Pulse survey, which indicated that 71% of respondents reported that a quarter or fewer of their deployed "agents" were capable of autonomously completing multi-step tasks, and only one-tenth claimed to have deployed agents at scale.
While there is no doubt that enterprises are actively constructing the control planes and underlying infrastructures necessary for advanced AI agents, the realization of the "true agentic wave"—marked by widespread adoption of sophisticated, autonomous, and multi-step AI capabilities—appears to be on the horizon for many, rather than a present reality. The current focus on platform diversification, security, and cost control lays the groundwork for this future, but the ultimate realization of highly autonomous AI agents remains a work in progress.
