KPMG finds 49% of leaders scaled back AI agents over cost

KPMG published its Global AI Pulse for Q2 2026 on 24 June 2026. In it, 49% of senior leaders said they had scaled back AI agent deployments because operating costs outweighed the benefits. The survey covered 2,145 senior leaders across 20 countries at organisations with more than $50 million in revenue.

What KPMG found

KPMG's US pulse surveyed 204 leaders at billion-dollar companies. Only 26% of them have full real-time visibility of what AI costs to run at scale. A further 35% cite AI cost management and economic literacy, meaning token and inference pricing, as a barrier, while 54% have embedded cost reviews into their AI approval processes. KPMG also reported that employee resistance rose to 20% from 5% the previous quarter, and Forbes covered the findings on 9 August 2026.

Why agent costs are an engineering problem

An agent in production makes repeated model calls, and each one is billed in tokens. Without evals, your team can't tell whether a cheaper model or a shorter prompt still gives acceptable output, so spend grows with no clear trade-off to point to. Token budgets per workflow and caching of repeated calls are decisions engineers make in the codebase. So is routing simple steps to deterministic code. Some steps don't need a model at all, and a rules engine or a plain query is cheaper to run and easier to test.

What it means for engineering teams

KPMG's 26% figure suggests most large companies can't see what each workflow costs, which makes it hard to defend an agent in an approval review. With 54% now putting cost reviews into approvals, your team is likely to be asked for cost per completed task, failure rate and how both change when the model or prompt changes. Producing those numbers means logging token usage per run, tagging calls by workflow and keeping eval suites that track quality alongside spend. Teams that can show this data can keep agents running where they pay off and retire them where they don't.

What it means when you hire engineers

When you hire engineers to build or maintain agents, it's worth asking how they'd measure cost per workflow, when they'd cache and when they'd take a model out of a step entirely. Prompting skill alone won't close the gap KPMG describes around token and inference pricing. Engineers who join client teams through Miyagami go through two sixty-minute assessments in a real codebase, the first with AI tools off and the second with the candidate's own agents on. Both scorecards are shared with you, so you can see how an engineer works with and without agents before they join your team.

You can read more about how we vet engineers, including what each session covers.