KPMG: 49% of firms scaled back AI agent rollouts as costs soared

Nearly half of large organizations have scaled back, narrowed, delayed, or paused AI agent deployments because operating costs began to outweigh the value produced, according to a KPMG survey released on June 24, 2026 KPMG. The report, which covered 2,145 senior leaders across major enterprises, signals a sharp reversal in the pace of agent adoption after years of rapid expansion.

The Numbers Behind the Pullback

Of the organizations that took action, 24% scaled back or narrowed their agent deployments, while 25% delayed or paused further rollout KPMG. The breakdown reveals a split between those actively retreating from agent investments and those taking a wait-and-see approach.

Even more striking: only 7% of surveyed organizations reported that their deployments had reached established return on investment KPMG. This suggests that the ROI bar remains stubbornly high for most enterprises, despite rapid improvements in agent frameworks and model efficiency over the past 18 months.

What's Driving the Retreat

The primary culprit appears to be the gap between token costs and measurable business impact. As organizations moved beyond proof-of-concept deployments into production workflows, the cumulative expense of running agents continuously—particularly those requiring frequent API calls, long context windows, or multi-step reasoning—began to strain budgets. At the same time, many organizations struggled to isolate and quantify the value agents were generating, leading to difficult conversations about whether the investment was justified.

The timing matters: this survey was conducted amid a period when inference pricing had begun to stabilize after several years of rapid cost reduction. While per-token costs continued to fall, they were no longer declining fast enough to offset the operational overhead of deploying agents at scale across distributed teams and workflows.

What This Means for the Agent Economy

The KPMG finding contradicts the narrative of unstoppable agent adoption that dominated tech coverage through 2025. Instead, it reveals a market correcting toward realistic expectations: agents are useful in specific, high-impact scenarios, but they are not universally profitable in every business process.

For vendors in the agent space—including framework developers, model providers, and agent-ops platforms—the pullback will likely accelerate demand for better measurement tools, cost-optimization frameworks, and use-case guidance. Organizations that paused deployments are not abandoning agents; they are demanding proof that they work.