How telemetry pipelines keep AI agent costs under control
The New Stack Megan Carnegie
AI agents are blowing up telemetry bills, and 59% of firms have delayed or killed deployments over the cost. The fix is pushing controls upstream before data hits pricey observability tools.
Based on reporting by The New Stack, Megan Carnegie — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
As companies move AI agents out of the lab and into production, a less glamorous bill is starting to bite: telemetry. These systems are messy by design. They loop, retry, branch, call tools, and generate a lot more data than a normal app request. That makes their monitoring harder to plan and much harder to pay for.
The pressure is already changing decisions. In a survey of more than 300 enterprise IT decision-makers in North America and Western Europe, commissioned by Apica and run by Omdia/Informa TechTarget, 59% said they had delayed or terminated an agentic AI deployment because of monitoring costs. Andi Mann, Apica’s chief product and technology officer, says the finance side is often the one cutting projects when the bills get fuzzy. He described a large bank that could not pin down what it was spending on AI programs and ended up cancelling some of them.
The numbers behind the trend are ugly. Most enterprises, 54%, say telemetry volume has tripled in the past year, with AI/ML workloads driving 43% of that growth. The average observability bill now sits at $3.17 million, up 28% year over year. No wonder 83% of respondents rank AI observability as a top priority for the year ahead.
What makes agents so expensive is not just volume but shape. A single customer-support task can create traces, model calls, retrieval steps, tool calls, retries, and loops. If that agent hands work to another agent, the trace grows again. Add identifiers like tool_name, agent_id, and trace_id, and the data gets harder to aggregate and more expensive to index.
Mann’s argument is simple: stop shipping all of that raw data straight into the priciest storage and analytics systems. Instead, make decisions earlier in the pipeline. Sample routine successes, keep failures and weird slow traces, redact sensitive data, and route records based on what they’re actually for. The companies doing that are better prepared for agentic AI growth, and the research in the Apica study says they’re also more likely to avoid the cost trouble hitting everyone else.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI that rarely gets a glossy demo: the invoice. The industry keeps talking about autonomous agents like they’re free-thinking coworkers, then acts surprised when the logs behave like a gold mine for the observability vendor. Europe’s regulators will love the “collect everything” habit even less than finance does.
Read more about this at: The New Stack