1. Executive Summary

The promise of AI agent autonomy has collided with a brutal budgetary reality. According to the most recent VB Pulse data, collected from 107 large enterprises, one in five respondents admits to lacking effective mechanisms to stop in real time an AI agent that incurs uncontrolled spending. This figure is not a minor operational anomaly; it is the symptom of a broken trust architecture between engineering teams and commercial orchestration platforms.

The problem is not solely the dollar cost per token, but the absence of emergency stop mechanisms (kill switches) and granular spending policies that are applied consistently across multiple frameworks. The industry has responded with a strategy of "hybridizability": 85% of companies use two or more orchestration tools, and 64% operate with three simultaneously. This plurality, far from being an aesthetic choice, is a defense against the lack of trust in the security and permission capabilities of individual vendors. For chief technology officers, data governance leaders, and platform architects, this report is a wake-up call. The question is no longer which language model is smarter, but who controls the spending tap when an autonomous agent decides, on its own, that it needs to process one million tokens to complete a trivial task. The answer, according to the data, lies not in a single vendor, but in a proprietary control layer that companies must build and demand.

2. Deep Technical Analysis

The root of the real-time control problem lies in the execution architecture of modern agents. Unlike traditional chatbots, which process a request and return a response, an agent operates in a reasoning-action-observation loop. Each iteration of this loop can generate multiple API calls, external tool invocations, and vector database queries. The cost is not linear; it is exponential when the agent enters a "chase" cycle toward a poorly defined objective.

VB Pulse data indicates that Microsoft AI Foundry/Copilot Studio is present in 70% of technology stacks, closely followed by the OpenAI Agents SDK (68%) and the Anthropic Claude Platform (47%). This coexistence of three distinct ecosystems creates a telemetry challenge. Each platform has its own cost dashboard, its own token metrics, and, crucially, its own spending limits. An engineering team operating with three tools cannot easily aggregate total spending, let alone establish a unified budget that applies in real time.

The 22% of respondents building their own orchestration reveals a fundamental technical distrust. These teams are not only seeking to avoid vendor lock-in; they are seeking to implement permission policies that vendors do not offer natively. For example, the ability to interrupt an agent loop based on the content of intermediate output, or to limit the number of calls to a specific tool per minute, are features that often require a custom control proxy. Without this layer, an agent that receives an ambiguous instruction can interpret it as an order to iterate indefinitely, accumulating costs without any human intervention.

Visibility is the other major technical failure. Vendor dashboards offer aggregated metrics with a delay of minutes or hours. For an agent that can consume thousands of dollars in tokens within seconds, this data latency is unacceptable. The technical solution emerging in the most mature companies is the implementation of real-time telemetry proxies that intercept every API call, calculate the marginal cost instantly, and compare it against a "burst budget." If the threshold is exceeded, the proxy can inject a "pause" instruction into the agent loop or revoke API credentials within milliseconds. Furthermore, model fragmentation exacerbates the problem. With GPT-5.6 Sol, Claude Opus 5, and Gemini 3.7 Flash available simultaneously, model routers decide which model to use for each subtask. If the router is not configured with per-model cost limits, it can send trivial tasks to the most expensive and sophisticated models, driving up the bill without the agent technically "failing." Cost governance, therefore, must be integrated into the routing layer, not just the orchestration layer. Finally, permission security is a critical factor. An agent with access to paid tools (such as third-party APIs or cloud computing services) can perform irreversible actions. Companies reporting a lack of real-time control are those that have not implemented a "read-only mode" for agents in testing phases, or that have not segmented credentials by risk level. Trust in the vendor is low, and VB Pulse data confirms that this distrust is the primary driver of the multi-platform strategy.

3. Industry Impact and Market Repercussions

The inability to curb an agent's spending has direct implications for the return on investment (ROI) of AI. Companies that cannot control costs in real time are reluctant to scale their autonomous agent deployments. This slows adoption in high-value use cases, such as the automation of complex business processes or proactive customer support, where an uncontrolled agent could generate significant losses within minutes.

For platform vendors, this trust crisis is both an opportunity and a threat. Microsoft, with its deep integration into Azure, has the advantage of offering cost policies at the enterprise subscription level. However, OpenAI's presence in 68% of stacks indicates that technical teams prefer the OpenAI SDK for its flexibility, despite Microsoft being the primary investor. This dynamic creates tension: customers want OpenAI's innovation with Microsoft's governance, and neither company currently offers a perfect integrated solution. Anthropic, for its part, leads in future adoption intent. Its focus on agent safety and interpretability (with models like Claude Opus 5 and Claude Fable 5) resonates with teams that prioritize control. However, its reliance on AWS as the official cloud (with Amazon as the primary investor of $4 billion) and Google's minority investment ($2 billion) create a complex alliance ecosystem that could slow innovation in cross-platform governance tools. The market for agent observability and governance tools is booming. Companies offering cost control proxies, permission management, and unified telemetry are seeing growing demand. The fact that 22% of companies build their own orchestration suggests there is a market gap for "control layer" solutions that are vendor-agnostic. Industry analysts point out that the next competitive battle will not be between language models, but between agent governance platforms. Regulatory pressure is also increasing. In the European Union, the AI Act requires transparency and human control over high-risk systems. An agent that spends money without supervision could be considered a governance failure, with potential fines. Multinational companies operating in Europe are therefore more motivated to solve this problem than their counterparts in other regions, which could create a competitive advantage for vendors offering robust and auditable control tools.

4. Expert Perspectives and Strategic Analysis

The technical consensus among platform architects is clear: cost control responsibility cannot be entirely delegated to the model provider. Engineering teams must assume that any agent can fail and design fault-tolerant systems. This implies implementing "agent budgets" that are defined per task, not per month. An agent processing invoices must have a spending limit per batch of invoices, and if it exceeds it, it must stop and escalate to a human.

VB Pulse data suggests that the most successful companies in cost control are those that have designated an "AI cost owner" within the platform team. This person has the authority to review agent execution logs, identify anomalous spending patterns, and adjust model routing policies. Without this figure, cost control becomes a diffuse responsibility that no one fully assumes. A recommended strategy is the implementation of "degradation modes." If an agent approaches its budget, the system can automatically switch from a premium model (like Claude Opus 5) to a more economical model (like Gemini 3.7 Flash or Llama 4) for the remaining subtasks. This technique, known as "model cascading," allows the task to be completed without exceeding the budget, albeit with potentially lower quality. Companies that have not implemented this strategy are those that report uncontrolled expenses. Permission security also requires a "least privilege" approach. Agents should not have access to all company tools by default. Instead, they should obtain dynamic permissions based on the specific task. For example, a customer support agent should not have access to the billing API unless the task explicitly requires it. Implementing an "authorization proxy" that validates each call to an external tool is a recommended practice that reduces the risk of irreversible actions. Finally, experts emphasize the importance of real-time observability. Cost metrics must be available on the same dashboard as performance metrics (latency, error rate). An agent that is slow and expensive is a double problem. Correlating these metrics allows teams to quickly identify if an agent is stuck in a retry loop, which is often the main cause of uncontrolled spending. Investment in distributed tracing tools is essential to achieve this visibility.

5. Future Roadmap and Predictions

In the next 12 to 18 months, platform providers are expected to integrate more native and granular cost controls. Microsoft, OpenAI, and Anthropic are under pressure to offer API-level "emergency switches" that allow customers to remotely pause an agent. The standardization of these mechanisms will be a key differentiator in enterprise contract renewals.

By the end of 2027, we predict that most large enterprises will have consolidated their orchestration strategy into two main platforms, rather than three. The current trend of using three tools is unsustainable in the long term due to operational complexity. However, consolidation will not be towards a single provider, but towards a "dual-core" where one platform manages customer-facing agents and another manages internal business process agents. The emergence of de facto standards for agent telemetry is imminent. Initiatives like OpenTelemetry are expanding their scope to cover AI agent events, including cost per token and routing decisions. Companies that adopt these standards early will have a competitive advantage in terms of portability and control. Providers that do not adopt them risk being marginalized in purchasing evaluations. On the model front, the trend towards more efficient models (like Gemma 4 12B or Llama 4) will allow companies to run lower-cost agents for routine tasks, reserving premium models for high-value tasks. This segmentation of the model market will reduce pressure on budgets but will require even more precise governance to ensure the correct model is used for the correct task.

6. Conclusion: Strategic Imperatives

In conclusion, the strategic analysis of One in five companies cannot stop uncontrolled AI agent spending in real-time underscores a critical transformation in modern software architecture and executive-level decision-making. The speed of innovation not only demands evaluating the raw performance of new technologies but rigorously quantifying economic efficiency, latency in production environments, and the interoperability of corporate infrastructures.

For Chief Technology Officers (CTOs) and architecture teams, the strategic imperative lies in avoiding exclusive reliance on single vendors (vendor lock-in), implementing robust enterprise data governance mechanisms, and designing agile systems capable of diverting workloads based on operational complexity. Competitive advantage will belong to organizations that execute this integration with technical discipline and long-term vision.