Most Firms Claim an AI Return Without Knowing the Cost

← Back · ACT II · Chapter 7: Value Is Not Promised; It Is Proven

Ninety-eight percent of executives report a return on their AI. About a quarter can see what it cost.

On July 28 EY published the fifth wave of its US AI Pulse Survey. The instrument polled 534 US decision-makers at senior vice president and above, fielded April 24 to May 17, 2026, with a margin of error of four points. Ninety-eight percent of respondents whose organization is investing in AI say they have seen a positive return on investment from their AI initiatives.

The same release reports something else about the same people. Eighty-two percent say their organization is concerned about AI token usage and related costs. Ninety-eight percent of those using tools that consume tokens say those costs have caused the organization to reconsider its approach. And only sixty-four percent say their organization actively monitors token usage and has clear budgetary guardrails on what it spends.

Return on investment is a ratio. It has a numerator and it has a denominator, and the denominator is what the thing cost. More than a third of these organizations are not watching the fastest-moving cost line in the deployment. They are not reporting a return. They are reporting an impression of one, and the impression is arriving on desks that will fund the next wave against it.

Two Instruments, One Window

KPMG measured the same weeks from the other direction. Its Global AI Pulse Q2 2026 surveyed 2,145 C-suite and business leaders across twenty countries, territories and jurisdictions. The US cut surveyed 204 C-suite leaders at organizations with a billion dollars or more in revenue, fielded April 28 to May 25, 2026.

That US release reports sixty-six percent of organizations with monitoring dashboards and sixty-one percent with AI approval processes in place. It also reports that only twenty-six percent have full, real-time visibility into what their AI systems cost to operate, and that only thirty-six percent have implemented direct token or usage controls.

These are not the same respondents. They are the same stratum, in the same weeks, answering about the same spend. One instrument asked executives whether they got a return and ninety-eight percent said yes. The other asked whether they can see what the thing costs, and twenty-six percent could. A buyer who cannot separate what a system costs from what it is sold for is already working from the vendor’s arithmetic.

Read the Ceiling, Not the Multiple

KPMG went further and segmented on the variable. On its own page for the Q2 Global AI Pulse, the firm reports that organizations with full visibility into AI operating costs are five times more likely to report established ROI than those without, at fifteen percent versus three percent.

The multiple is what gets quoted, and the ceiling underneath it is the part worth reading. In a 2,145-leader global sample, the best-instrumented group establishes a return fifteen percent of the time. That is what the number looks like when someone checks it, against ninety-eight percent from the same population when nobody does.

Wharton Human-AI Research and GBK Collective found the same divide in a study published in October, drawn from about 1,400 people at US companies with more than a thousand employees and surveyed in July. Forty percent of the managers and executives were very confident their organization could track its total spend on AI once the costs folded into other software, licenses and cloud services were counted. The report describes everyone else as “attempting to prove a return against a number they are less sure of.” In the confident group, fifty-six percent reported strongly positive returns, against ten percent of the rest. The authors add that the confident respondents are also more enthusiastic about AI in general, so some of the difference may reflect “conviction rather than capability.”

McKinsey’s annual survey reads the same way from further up the house. McKinsey put its 2026 State of AI survey in the field from May 4 to June 8 and drew 1,719 responses across 97 nations. Thirty-seven percent of respondents attribute at least some EBIT impact to their use of AI, about the same share as a year earlier. Six percent attribute five percent or more of EBIT and call the impact significant, and that share has not moved either. Eighty percent say AI has improved their own productivity. The distance between the last figure and the first two is the whole problem. Individual improvement is easy to feel and enterprise return is hard to trace, and only the second one survives a renewal conversation.

The plainest version of the gap sits inside one survey. The CFO Survey, run by Duke University with the Federal Reserve Banks of Richmond and Atlanta, asked nearly 750 senior finance executives at the end of 2025 how much AI had raised their labor productivity that year. On average they said 1.8 percent. The same executives also estimated what AI had added to their revenue and to their headcount, and the productivity those two answers imply is 0.6 percent. The economists who analyzed the responses read most of the difference as timing, with revenue arriving later and some quality gains not yet showing up in it, and both figures are self-reported. Even allowing for that, the gain the finance chiefs reported was three times what their own revenue and headcount answers support.

The Standard the Buyer Will Not Turn Inward

The Value Standard asks the buyer to define value before the seller does, and to hold the seller to evidence the buyer specified. Buyers are getting better at this. They discount the vendor case study, they read the footnote on the analyst chart, and they know a benchmark drawn from a vendor’s own telemetry is not a benchmark.

Then they turn to their own program and produce the exact artifact they just refused. A favorable outcome, asserted by the party with an interest in the outcome, with no traceable cost basis underneath it. The vendor at least has to survive a procurement process. The internal claim goes straight to the board.

Value drift is usually described as the gap between what a seller promised and what a buyer can prove it received, and that is where it is easiest to see. It does not start there. It starts wherever a claim enters the record without a mechanism that could have falsified it, and a buyer’s own claim about its own program qualifies. An enterprise that cannot compute its own denominator has no standing to demand one from anybody else, and no way to know whether the AI it is buying is working.

What This Costs at the Next Renewal

A buyer who believes their AI is returning value negotiates from that belief. The renewal arrives, the vendor raises the price against a value story the buyer has already told internally, and the buyer has nothing with which to contest it. Ninety-eight percent is a concession made in advance of the conversation it will be used in.

What fixes it is instrumentation the buyer holds on both sides of the ratio, in place before the claim is made rather than after someone challenges it. KPMG and Wharton both found the dividing line in the same place, between organizations that can see what their AI costs and organizations reporting a return they cannot check.

Proof over promise applies to the promises an enterprise makes to itself.

A 2026 survey of 2,648 C-suite executives shows the same gap from procurement’s side: cost control is back on top as its primary contribution, and confidence in the function fell.

References: EY US AI Pulse Survey, fifth wave, July 28, 2026 (n=534, fielded April 24 to May 17, 2026). KPMG Global AI Pulse Q2 2026 (n=2,145, twenty countries). KPMG US AI Quarterly Pulse Survey, June 24, 2026 (n=204, fielded April 28 to May 25, 2026). McKinsey, The state of AI in 2026: On the road to ROI, August 25, 2026 (n=1,719, 97 nations, fielded May 4 to June 8, 2026). Baslandze et al., Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives, NBER Working Paper 34984, March 2026 (The CFO Survey with supplemental surveys, n=748, fielded November 11, 2025 to mid-January 2026). Wharton Human-AI Research and GBK Collective, The Adaptive Advantage: Who’s Winning with AI and How, October 2026 (managers and above n=1,007, entry and mid-level n=409, fielded July 9 to 29, 2026).

Keep reading: ACT II · Chapter 7: Value Is Not Promised; It Is Proven