Telemetry pipelines emerge as critical tool to manage exploding AI agent costs

The shift from AI experiments to live agent deployments is triggering an unexpected crisis: observability costs are becoming unmanageable. Unlike traditional applications with predictable data flows, autonomous agents generate telemetry in unpredictable ways, making both monitoring bills and system behavior difficult to forecast. The result is a wave of delayed or cancelled AI projects, often killed not by engineering teams but by finance departments unable to justify the spend.

The scale of the problem#

Research from Apica surveyed over 300 enterprise IT decision-makers across North America and Western Europe, revealing that 59% of organizations have already halted or postponed agentic AI deployments specifically due to monitoring costs. Many of these cancelled projects represent high-stakes use cases in cybersecurity, compliance, and fraud detection—exactly the areas where visibility matters most.

The financial pressure is mounting fast. Most enterprises report that their telemetry volume has tripled in the past year, with AI and machine learning workloads accounting for 43% of that spike—by far the dominant driver. Companies now spend an average of $3.17 million annually on observability, with spending growing 28% year-over-year. At the same time, organizations anticipate a staggering 9.5X increase in telemetry data over the next two years, and 44% expect growth of 6X to 100X.

"Imagine if your credit card bill went up more than nine times," notes Andi Mann, chief product and technology officer at Apica. "This isn't a gentle ramp; it's a skyscraper, and it's prompting panic."

Why agents create so much data#

The difference lies in how agent workloads generate telemetry. A single customer-support task might trigger a top-level trace plus multiple model calls, retrieval operations, tool invocations, retries, and loops. When an agent delegates to another agent, that complexity multiplies further. Each model logs token counts, latencies, costs, and provider details. Each tool call records arguments, results, status, and downstream effects. High-cardinality identifiers like agent_id and trace_id inflate indexing costs at every step.

This creates a fundamental mismatch: while 35% of enterprises claim widespread agentic AI deployment, nearly two-thirds acknowledge they are only moderately prepared for it. The infrastructure and monitoring systems designed for human-driven workloads cannot easily absorb the volume, variety, and unpredictability that autonomous agents introduce.