A statistic lit up social media. Nearly half of executives have delayed or scaled back AI agent deployments after costs exceeded benefits, according to KPMG.
Prediction market Polymarket amplified the number alongside its market putting the odds of an AI bubble burst by year end at roughly fifteen percent.
Start with where the number comes from.
It is drawn from KPMG’s Global AI Pulse for the second quarter of 2026, a survey of 2,145 senior leaders across 20 countries at organizations with more than $50 million in annual revenue. In it, 49 percent of leaders said they had scaled back AI agent deployments because operating costs outweighed the benefits.
AI remained a top investment priority for 79 percent of leaders, up from 74 percent the prior quarter, with average AI spending holding steady at $188 million.
The share of organizations describing AI as part of everyday work jumped to 22 percent from 13 percent in the first quarter, the largest single-quarter move at any stage of KPMG’s maturity curve.
In Asia-Pacific, 81 percent of companies reported that AI is already delivering meaningful business value, up from 69 percent three months earlier.
Companies pulling back on agents are not exiting but they are looking at how they will redesign them.
Why The AI Agent Bills Exploded
To understand what is happening, you need to understand how AI is priced.
Most vendors have shifted from flat subscriptions to usage-based pricing, billed in tokens. A token is a small chunk of text, roughly a word fragment. Every question an AI system reads, every answer it writes, and every step it takes consumes tokens, and companies pay per unit, the way they pay for electricity.
Pilots were cheap and often subsidized. And sometimes companies cannot even ROI.
Agents changed the math because they work differently than chatbots. They run long tasks, call other software, and check their own work, and every one of those steps is metered. When GitHub Copilot moved to usage-based billing on June 1, one Visual Studio Magazine writer tracked his first day under the new meter and projected a $180 monthly bill on a plan that had been a flat $10, driven by a single long, tool-heavy session.
Most companies cannot even see the meter.
In KPMG’s companion US pulse survey of 204 leaders at billion-dollar companies, only 26 percent report full, real-time visibility into what AI costs to run at scale. In the global survey, a third of leaders cite limited understanding of AI cost structures, including how token pricing works, as a barrier to deploying agents.
This is finance catching up with engineering.
What Leaders Should Do With The AI Agent Data
The question becomes what should you do before you start to scale.
- Install a meter before you scale. The organizations pulling ahead have AI cost dashboards, which 53 percent of leaders globally now report, and they see spending as it happens, not at invoice time.
- Make token economics a leadership literacy. Treat AI spend the way you treated cloud spend a decade ago, as a discipline with owners, forecasts, and unit costs per workflow.
- Put cost review inside the approval loop. In the global survey, 54 percent of organizations have embedded cost reviews into AI approval processes, so no agent scales without a projected cost-to-value case.
- Rephase rather than retreat. Concentrate investment where returns are strongest, which is exactly what KPMG says organizations are already doing.
The fuller dataset shows a market growing up, with less open-ended experimentation and more financial discipline, and budgets following results instead of promise.
The companies scaling back agents today are mostly clearing room to scale what works tomorrow. The bill came due. Reading it carefully is not a crash. It is AI agents reaching adulthood.
