One in Five Companies Unable to Control AI Agent Costs

As enterprises begin to harness AI agents at scale, challenges related to their "control" are coming to light. According to a survey conducted by VB Pulse covering 107 companies, a significant number of enterprises are unable to track and control AI agent token consumption—the costs incurred based on the volume of data processed by the AI—in real time. Notably, one in five companies cannot immediately halt agents that continue to operate beyond their budget, revealing a gap between cost management frameworks and operational reality.

The survey gathered data from practitioners directly involved in AI development, including software engineers, machine learning engineers, product managers, and VPs or directors in data, AI, and analytics divisions. AI agents refer to autonomous AI systems that execute tasks under human direction. As the adoption of such agents expands, securing "visibility" into actual costs becomes an increasingly critical challenge for enterprises.

A striking finding is that companies are strengthening their commitment to avoid dependence on any single vendor in orchestration—the infrastructure for managing multiple AI agents and models collectively. 85% of surveyed companies are using two or more orchestration tools simultaneously, with 64% operating three or more. Only 15% rely on a single tool. On a median basis, companies are simultaneously operating three platforms each.

Looking at specific usage patterns, Microsoft's AI Foundry/Copilot Studio leads at 70% of company stacks, positioning itself as the primary tool today. OpenAI's Agents SDK follows at 68%, with Anthropic's Claude Platform at 47%. Other platforms in use include Google's enterprise agent platform, LangChain/LangGraph, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex. Additionally, 22% of companies have built and operate their own proprietary orchestration foundations.

Regarding future direction, a shift toward diverse, multi-platform configurations is anticipated. 53% of respondents predict that by the end of 2026, primary control infrastructure will shift to hybrid models combining multiple approaches. Furthermore, over two-thirds of companies plan to change platforms within a year. Among tools under consideration, Anthropic's Claude Agent SDK leads at 43%, followed by Google's enterprise agent platform at approximately 33%, in-house development at 31%, and OpenAI's options at 25%.

Behind these trends lie not only the desire to avoid vendor lock-in to specific providers but also deep-seated concerns about security and access permission management. The survey revealed widespread sentiment that existing vendor-provided security and permission management mechanisms cannot be fully trusted, prompting companies to want control systems of their own. One perspective suggests that some enterprises' previous experience with becoming overly dependent on specific cloud providers during the early cloud era continues to influence today's distributed strategy.

AI agents remain an emerging technology, and which platform will become the standard is yet to be determined. The current diversified landscape across multiple tools can be understood as a rational response to this uncertainty. Conversely, cost visibility challenges will intensify as agent autonomy increases. Going forward, the ability to provide control and cost management capabilities in practical, actionable ways appears poised to become a central axis in platform selection decisions.