Nearly half of executives pulled back on AI agents. Half of what?

A number went around this week. KPMG’s Q2 2026 AI Pulse research reports that 49% of leaders scaled back AI agent deployments because operating costs outweighed the benefits. It is a good number. It is specific, it is counterintuitive against the general enthusiasm, and it comes from a serious firm.

It is also being retold in a way the underlying research does not support, and the retelling is more interesting than the number.

The two figures are from two different surveys

The version travelling fastest pairs the headline with a second statistic, usually in one sentence: 49% scaled back because cost beat benefit, but only 26% have real‑time visibility into what AI costs to run.

Read that sentence and you form a view. They cut because they cannot see.

We went to the primary sources.

Two instruments, two populations, two revenue thresholds an order of magnitude apart. And the US release that carries the 26% contains no scale‑back percentage at all.

That does not make either number wrong. Both may be perfectly sound measurements of the populations they sampled. It means the sentence joining them is doing work neither survey did. A global mid‑market‑and‑up population and a US billion‑dollar population are not the same companies, so they cut because they cannot see is a hypothesis someone formed in the retelling, not a finding either instrument reports.

We would like it to be true. It is a tidy story and it matches our own priors, which is exactly why it is worth checking.

Nobody defines the thing being counted

In a set of claims whose entire content is a percentage, we could not find a definition of “AI agent” in the coverage or on the survey pages we reached. The nearest thing is functional: systems that run long tasks, call other software, and check their own work. That describes a category. It does not draw a boundary, and a respondent deciding whether their document‑classification pipeline counts will not find the answer there.

“Scaled back” is not operationalized either. It appears alongside “delayed” as though they were the same event. They are not. A deployment postponed a quarter for budget timing and a deployment killed after a cost review are different outcomes with different implications, and one of them is not a pullback at all.

This matters more than it sounds. When the definition is loose, the number measures the respondents’ interpretation as much as their behavior, and comparisons across quarters or against other surveys stop meaning what they appear to mean.

The most interesting figure is the one nobody is quoting

The global research reports that organizations with full visibility into AI operating costs are five times more likely to report established ROI than those without: 15% versus 3%.

Five times is the headline. Look at the absolute numbers.

Even among organizations that can see what their AI costs, 85% do not report established ROI. Among those that cannot, 97% do not. The relative gap is real and large. The absolute picture is that established ROI is rare in both groups, and cost visibility does not appear to be sufficient for it.

We have no view on why. We only note that “5x more likely” and “85% still don’t have it” describe the same two numbers, and only one of those framings is travelling.

Why we think this may not be a cost story at all

Here is the reading we find most plausible, offered as a hypothesis and labeled as one.

The pullbacks are being described as a cost problem. But the supporting evidence is about not being able to see cost: a quarter to a third of organizations report full visibility, a third cite limited understanding of cost structures including how token pricing works, and visibility correlates with ROI.

An organization that knows a workflow costs €4,000 a month to run and generates €3,000 of value has a cost problem, and cancelling is the correct decision, arrived at correctly.

An organization that cancels because the bill surprised them has an instrumentation problem. The deployment might have been fine. Nobody could tell, so it died in the budget review.

Those look identical in a survey response and they are entirely different failures. The first is a market discovering an economic boundary. The second is a market that has not yet built its meters. We would bet on the second being the larger share, and we would want to be shown.

What would change our mind

• The full Global AI Pulse instrument showing that the 49% and the cost‑visibility questions were asked of the same respondents, which would make the causal reading a legitimate cross‑tabulation rather than an inference.
• A definition of “AI agent” tight enough that two respondents in the same company would answer the same way.
• Evidence that the scaled‑back deployments were measured before cancellation, which would move it from an instrumentation story to a genuine economics story.
• Longitudinal data. One quarter is a snapshot, and the same research reports AI as a top investment priority for 79% of leaders, up from 74%. Pulling back on some deployments while raising overall commitment is not retreat. It is triage, and triage is what a maturing market looks like.

Why we bothered

We publish our own research with a stated error rate and preregistered hypotheses, so we try to hold other people’s numbers to what we would want applied to ours. That means we are not saying KPMG got this wrong. We have not seen the full instruments, we surveyed nobody, and the fieldwork is theirs.

What we are saying is narrower and, we think, more useful: between an instrument and a timeline, a number picks up a story it did not arrive with. Two populations became one. A correlation became a cause. A relative multiple replaced an absolute base rate. None of that required anyone to be careless. It only required the number to be interesting.

That happens to nearly every statistic that travels. It is worth knowing which ones you are holding.