KPMG survey finds 49 percent of enterprises scaled back AI agents over costs

Nearly half of large enterprises have pulled back on AI agents, and the reason is not capability. It is the bill.

What the survey found

KPMG’s Global AI Pulse for Q2 2026 surveyed 2,145 senior leaders across 20 countries at organizations with more than $50 million in annual revenue. 49 percent said they had scaled back AI agent deployments because operating costs outweighed the benefits.

The pullback is not a retreat from AI itself. In the same survey, 79 percent still rank AI as a top investment priority, and average AI spending held steady at $188 million. What changed is that the bill arrived and many teams could not explain it: roughly a third of leaders point to limited understanding of AI cost structures, including how token pricing actually works, as a barrier to deploying agents.

A separate KPMG pulse of 204 US leaders at billion-dollar companies adds a sharper data point. Only 26 percent said they have full, real-time visibility into what their AI costs to run at scale.

The demo-to-production gap

The pattern the numbers describe is consistent. Agents are cheap to demo and expensive to operate, and the gap between those two states is where the pullbacks are happening.

An agent pilot runs a handful of tasks on a bounded budget. A deployed agent runs continuously, chains model calls, retries failures and consumes tokens at a rate that few procurement processes were built to forecast. When a third of senior leaders say they do not fully understand the pricing model of the thing they are buying, cost overruns stop being surprising and start being structural.

The visibility figure may be the most actionable one. If only about a quarter of billion-dollar companies can see their AI running costs in real time, most organizations are discovering the economics of agents after the fact, in the invoice. That sequencing, spend first, understand later, is the mechanism behind the 49 percent, and it suggests the next phase of enterprise AI adoption will be won as much by cost instrumentation as by model quality.