NEW YORK — What happened
A cross-industry financial survey covering 140 enterprise IT departments found that corporate pilot programs for autonomous tool-using agents exceeded their projected cloud inference budgets by an average of 320% in the first half of 2026. As a result, nearly a third of surveyed firms have paused deployment.
Why it matters
Investors poured billions into agentic automation startups on the promise that AI agents would perform complex, hours-long office workflows for pennies. The economic reality is that an agent attempting to correct its own errors, inspect user interfaces, and execute recursive tool loops consumes staggering volumes of tokens, often costing more than human labor for repetitive administrative work.
Deep dive
The root cause of the budget explosion is recursive tool calling. In a typical customer-support triage setup, a traditional conversational bot uses roughly 800 tokens per interaction. In contrast, an autonomous agent tasked with investigating an order discrepancy, querying an ERP system, interpreting an inventory database, and verifying shipping logs generates an average of 24,000 tokens across multiple internal chain-of-thought and verification passes. If the agent hits an unexpected error or loops while parsing a malformed table, single transactions have been observed burning through over 150,000 tokens before failing.
Report check
Startup pitches routinely assert that autonomous agents cut administrative task costs by 90%. While this holds true for trivial one-step workflows, accounting data across financial services and healthcare logistics proves that complex multi-step agents currently cost up to 40% more per resolved ticket than hybrid human-assisted pipelines once API bills and monitoring costs are tallied.
Open questions
Will inference costs decline quickly enough to rescue the economics of autonomous agents, or will enterprises mandate hard token caps that reduce agent success rates? Will businesses retreat to deterministic workflow automation (RPA) augmented by small, fine-tuned task-specific models?