McKinsey Says AI Still Isn't Showing Up in the Books

44% of the organizations McKinsey surveyed say they're deploying AI enterprise-wide. A year earlier, that number was 38%. In the same report, 37% say AI has contributed positively to their operating profit. Last year it was 39%.

One number gained six points. The other one didn't move. The report is titled "On the road to ROI."

Who answered, and about what

The survey ran from May 4 to June 8, 2026, and pulled in 1,719 responses across 97 countries. These aren't all executives: McKinsey describes participants at all levels, from the C-suite down to individual contributors. 36% work at companies with more than $1 billion in revenue. Responses are weighted by each country's share of global GDP.

The rest of the report tells the story of a year of real acceleration. Among companies above that $1 billion mark, the share deploying agents at scale climbs from 27% to 40%. Nearly a third of respondents skipped buying at least one piece of software because coding agents could build it in-house instead. 60% plan to raise their AI budget next year, and 20% already say operating costs, tokens included, have started capping how much they can use it.

All of that moved by several points in twelve months. Only one number didn't, and it's the one tied to money actually made. The way Microsoft and Meta split their capex showed what AI costs the companies building it. This report measures what it earns the companies buying it.

A plateau, not a decline

Going from 39% to 37% isn't a drop. At sample sizes like these, the margin of error runs roughly two points in either direction, and weighting by each country's GDP share widens it further. The two numbers are statistically indistinguishable, and McKinsey says as much itself: "essentially unchanged." The firm is right to say it.

The question's wording shifted slightly too: "any level of EBIT impact" in 2025, "at least some EBIT impact" in 2026. The two samples aren't identical either, 1,993 participants across 105 countries last year versus 1,719 across 97 this year. McKinsey got ahead of the objection and dealt with it in a footnote: the firm isolated the 552 people who answered both editions, and that subgroup's results line up with the full sample's.

So the plateau holds up. It's the one point where the report and any critical read of it actually agree. What gets interesting is what comes next.

What disappeared between the two editions

The 2025 figure of 39% arrived with its own chart and its own base: 1,753 respondents, all from organizations that use AI regularly in at least one function. The 2026 figure of 37% has neither. It lives in the running text, and it doesn't appear in any of the report's eighteen exhibits.

There's nothing improper about that on its own. A report picks its own illustrations, and the editorial choices shift between editions. But a number you can reconstruct and a number you're asked to take on faith don't read the same way, especially when it's the one carrying the title's whole promise.

Same story with what McKinsey calls "AI high performers," organizations that attribute a meaningful share of their profitability to AI. They're 6% this year, 6% last year. Except the threshold isn't phrased the same way: "more than 5%" of EBIT in the 2025 notes, "5% or more" in the 2026 notes. Whether the underlying question changed is anyone's guess, the firm doesn't say. In headcount terms, that group shrinks from 109 to 92 people.

An 80% productivity gain, but among whom

The report's most-quoted number sits elsewhere: 80% of respondents say AI has improved their personal productivity, and half say it helps them make better decisions. That's a real result, and it's a new one. New in the strict sense: this question wasn't part of the previous edition, so there's no year-over-year comparison to make, in either direction.

The exhibit's footnote spells out who was actually asked: only respondents who use AI in their own work. Adding up the headcounts given by category gets you a base of 1,490 people, 654 of them C-suite. Nearly every other respondent claiming a productivity win is an executive describing their own workday.

It's the same scale mismatch that turns up in every survey on this topic, one we already flagged in what productivity studies are really measuring when they talk about AI. Counting applause in the room has never once filled the register: both things get measured, just not in the same place, and not by the same people.

The same panel already got it wrong, in the same direction

The report contains its own reality check, and it's a useful one. Last year, 32% of respondents expected a headcount cut of at least 3% over the following year. A year later, 14% report an actual decrease. McKinsey puts the two numbers side by side itself: less than half of what had been forecast actually happened.

This year, 39% expect a reduction over the next twelve months. The forecast keeps climbing while the measurement of the last forecast keeps shrinking. McKinsey states this plainly, and credit where it's due for that: the firm publishes the gap between what its panel predicted and what it later observed.

That's the strongest argument for reading a report like this as a sentiment gauge rather than a dashboard. The intentions get measured carefully. They just tend to run ahead of the facts by a cycle or two.

What the firm says about it, in its own words

The report's closing line is more honest than its headline. In McKinsey's own words: "Organizations' conviction in AI is growing faster than the immediate financial returns they can attribute to it." Conviction is outpacing the returns anyone can actually pin on it.

The Register, which first flagged the gap between the headline and the underlying data, reports that the firm reached back out after its article ran to clarify what it meant by that road to ROI. Michael Chui, a senior fellow at McKinsey QuantumBlack and one of the report's coauthors, told the outlet by email that part of the return is already showing up, that he expects more of it over time, and that this is a journey rather than a destination. He added that nobody should be surprised if it takes years, since earlier technologies followed the same curve.

He's probably right about that. Still, the road the title refers to comes down, for now, to a single number that hasn't moved.