More than half of enterprises have already pulled the plug on or postponed autonomous AI agent deployments because of monitoring expenses, according to a survey of over 300 enterprise IT decision-makers in North America and Western Europe commissioned by Apica and conducted by Omdia/Informa TechTarget. The report finds that 59% of organizations have terminated or delayed agentic AI projects due to telemetry costs, with finance departments—not engineering teams—making the final call to cancel programs. The deployments most frequently affected involve high-stakes applications including cybersecurity, compliance, and fraud detection.
The scale of the problem is already visible in corporate data volumes. Fifty-four percent of enterprises have watched their telemetry volume triple in just the past year, with AI and machine learning workloads accounting for 43% of that expansion—the single largest driver. Companies now spend an average of $3.17 million on observability, a figure climbing 28% year over year with no apparent limit. Looking ahead, businesses anticipate an average 9.5X jump in telemetry data within two years, with 44% of organizations expecting increases ranging from 6X to 100X. Not surprisingly, 83% rank AI observability as a top priority for the coming year.
"They knew they couldn't afford to keep going on the same trajectory, so they had no choice but to cancel certain AI programs," Andi Mann, chief product and technology officer at Apica, tells The New Stack, describing a large bank's recent struggle to track AI spending. Mann says he's witnessed a recurring pattern where AI initiatives consume budgets originally allocated elsewhere, forcing outages, downtime, penetration attacks, and DDoS protection to compete for the same resources. The report notes that nearly two-thirds of enterprises are only somewhat prepared for the shift to managing autonomous agents, despite 35% claiming widespread deployment.
The report explains that autonomous agents generate far more telemetry than conventional applications because each agent task can produce a top-level trace, multiple model calls, retrieval operations, tool invocations, retries, and loops—and when one agent delegates work to another, it adds another branch to the trace. Each model generates token, latency, cost, and provider data, while every tool call creates records for arguments, results, status, and downstream activity. Identifiers like tool_name, agent_id, and trace_id introduce high cardinality, making data harder to aggregate and more expensive to index, with costs multiplying at every stage. Unlike traditional applications, agents can call multiple models and tools, repeat tasks, or expand workflows in ways that are difficult to predict, making both capacity planning and monitoring budgets nearly impossible to forecast accurately.
The report recommends intervening earlier in the telemetry pipeline rather than sending all data to expensive central analytics platforms. According to the authors, a pipeline-first architecture can sample repetitive successful events while retaining failures, retries, policy violations, and unusually slow traces, then enrich records with agent, session, model, tool, token, and estimated-cost data before deciding what deserves ingestion. Enterprises with a telemetry pipeline are 50% more likely to be prepared for agentic AI data growth, and among mature agentic AI organizations, pipeline adoption is the distinguishing factor: these companies are 80% more likely to have avoided the operational cost challenges hampering their peers. Sixty-eight percent of enterprises plan to evaluate changes to their observability stack within six months, while almost a quarter say existing vendor relationships won't significantly influence those decisions. "This has already become a board-level decision," Mann says. "Ultimately, it's a choice about how smart you can afford to make your business." The tension between innovation velocity and infrastructure economics will likely define which organizations can sustain autonomous systems at scale, and which will need to retreat to narrower, more predictable workloads.