49% of companies running AI agents have already scaled one back, and KPMG's 2026 survey found only 26% could see what it was costing them

Gartner has been saying for over a year that AI agents would move out of pilot projects and into the daily operating layer of the business application, and 2026 is roughly on that schedule: its August 2025 forecast put task-specific agents inside 40% of enterprise applications by the end of this year, up from under 5% the year before [1]. Adoption is not the open question anymore.

What KPMG's quarterly AI Pulse survey found in its second 2026 round is the part that forecast does not show. Among the 2,145 senior leaders it polled across 20 countries, 79% still call AI a top investment priority even if a recession hit in the next year, up five points from the quarter before [2][3]. In the same survey, close to half, 49%, said they had already scaled back, delayed or narrowed an AI agent deployment because the running cost had outrun the value it delivered [3][4].

Those two numbers describe the same leaders at the same moment: still committed to AI as a category, already pulling back on specific agents that got too expensive to keep running. The reason sits in a third figure. Only 26% of the organizations surveyed said they have full, real-time visibility into what running AI at scale actually costs them [4][5]. Most are not cutting an agent because it failed. They are cutting it because nobody priced it accurately until the bill arrived, and a business running the same kind of agent without a finance team watching the meter is exposed to the same gap, just sooner.

KPMG runs the Global AI Pulse as a quarterly check-in, not a one-off report, which is part of what makes the Q2 2026 wave worth reading against the Q1 one. The sample, 2,145 C-suite and senior business leaders across 20 countries and territories, is large and consistent enough quarter to quarter that a five-point move, from 74% to 79% on AI staying a top priority through a recession, reads as a real shift in confidence rather than noise ^[2][3]. Commitment to AI as a line item is not softening. If anything, it is hardening.

The 49% figure sits inside the same report, and it does not contradict the 79%. Coverage of the survey breaks the pullback into two roughly even groups: about a quarter of leaders narrowed the scope of an agent deployment, and a similar share delayed or paused one outright, with both citing cost against value as the reason ^[3][4]. Neither group walked away from AI. They turned an agent on, watched what it did to a line item, and pulled the lever available to them, scope it down or pause it, rather than cancel the AI program itself.

The 26% visibility figure explains why the correction happens after the fact instead of before it. 66% of the leaders surveyed said they already have some kind of AI cost monitoring dashboard in place, but only 26% said that visibility was full and real-time ^[4][5]. A dashboard that updates weekly, or that tracks spend by department rather than by the specific agent and the task it is running, shows the shape of last month's bill. It does not show, while an agent is mid-run, that a retry loop or an unexpected spike in call volume is about to double what that task normally costs.

The mechanism is specific to agents, not AI generally, and it is worth separating from a familiar budget line. A seat-based software license costs the same whether an employee uses it ten times a day or never opens it. An agent's cost moves with what it actually does: every model call, every tool it invokes, every retry after a failed step, adds to the bill in real time ^[6]. That is the upside case for agents, a business pays only for work actually done, and it is also exactly why the bill can move faster than a monthly dashboard would catch. A support agent that starts handling twice the ticket volume during a sale, or a sales agent that starts calling a research tool on every lead instead of one in five, can double its running cost inside a week with nobody deciding that should happen.

None of this is an enterprise-only problem once the mechanism is this specific. A business running a single AI chat widget, lead-routing agent, or email assistant pays the same way, per call and per token, that the organizations in KPMG's survey do. What it typically does not have is the finance function that at least some of those larger organizations used to notice the pullback was needed. The KPMG numbers describe companies with the staff to eventually catch the gap between what an agent was supposed to cost and what it does cost. A smaller operator without that staff usually finds out the same way those leaders did, just later and with less warning: the invoice, not a dashboard.