Only 6% of Firms Get Real AI Earnings Impact • The Weighted Average

McKinsey surveyed 1,719 professionals and business leaders for its 2026 State of AI report and produced the widest gap yet between how AI feels and what it earns. Eighty percent of respondents who use AI at work say it has improved their individual productivity. 37% say their organisation attributes at least some EBIT impact to AI — statistically unchanged from 2025. And just 6% clear McKinsey’s high-performer bar of attributing 5% or more of EBIT to AI while calling the impact significant, also flat year over year. Divide the two ends and you get the number this brief is built on: felt productivity outruns booked earnings by more than 13 to 1.

The gap is a measurement problem before it is a value problem

A thirteenfold spread between personal and institutional returns has two honest explanations, and they demand opposite responses. Either the productivity is real and leaking — absorbed into slack, longer queues, or work nobody bills — or the productivity is self-reported and inflated, which surveys of one’s own output reliably are. McKinsey’s own coauthor leans toward the first. Senior fellow Michael Chui told The Register that high performers are seeing “real ROI” but that it takes organisational change rather than tool deployment, adding: “It should not be surprising that it has taken time, because it is a reflection of trends we’ve seen with other technologies.”

The cost side has stopped being free. Twenty percent of respondents said AI-related operating costs have constrained their use of the technology — a constraint that barely registered when everyone was piloting, and one that follows the industry’s shift to usage-based AI pricing. That is the same squeeze this paper traced when OpenAI reinstated a five-hour usage cap on $20 Plus seats: vendors are done absorbing agentic consumption, and buyers are discovering their spend was never a fixed subscription.

Anthropic’s own telemetry explains why that cost line is climbing faster than headcount. Its June 2026 Economic Index report finds that a typical conversation mapped to a top-tercile occupation consumes 2.07 times as many tokens as one mapped to the bottom tercile, with users taking 1.53 times as many turns and enabling extended thinking more often. Put McKinsey’s constraint next to Anthropic’s gradient and the squeeze is arithmetic: the work most likely to move EBIT is also the work that costs roughly double per interaction, so the 20% of firms already cost-constrained are constrained precisely where the returns would come from. Cheap tokens buy explanations; expensive tokens buy the artifacts a P&L notices.

Meanwhile adoption keeps accelerating into that headwind. 40% of respondents at organisations above $1 billion in revenue say they are scaling AI agents, up from 27% last year. Nearly a third say their organisation decided against buying one or more software products or features, choosing instead to build the functionality in-house with agentic coding tools. That second figure is the most consequential line in the survey and the least discussed — and it lands against evidence that developers ship AI-written code they know carries defects.

Coding agents are quietly repricing enterprise software

Read the two agentic numbers together and a market shift appears. If roughly a third of large enterprises are now declining SaaS purchases they would previously have made, the addressable market for mid-tier business software is being reduced by internal engineering capacity — capacity that only exists because coding agents changed the cost of a bespoke internal tool. That is a bigger claim than “AI improves productivity,” and unlike EBIT attribution, it is a decision you can audit in your own procurement log.

It also explains part of the earnings gap. Software a company builds instead of buying shows up as avoided cost inside an engineering budget, not as an AI-attributed line in EBIT. A CFO looking for the AI contribution in the income statement will not find it in a renewal that never happened. The measurement architecture predates the change; the value may be present and unlabelled.

Now the counterpoint, which is substantial. McKinsey sells AI transformation consulting, and “on the road to ROI” is a conclusion its business model prefers to “no ROI.” The survey is self-reported, non-random, and asks executives to grade their own initiatives. Its headcount findings undercut its own credibility as a forecast: 39% of respondents expect AI-driven job cuts this year, up from 32% — yet McKinsey concedes 2025’s actual workforce reductions “fell well short of what respondents in last year’s survey had anticipated,” a pattern that matched AI-driven layoffs followed by quiet rehiring. Expectations in this dataset have a track record of running ahead of outcomes, which is precisely the pattern the entry-level data suggests is reshaping hiring at the bottom rather than the middle.

The operator verdict: stop trying to attribute EBIT and start instrumenting the two things that are measurable this quarter — software you did not buy because an agent built it, and inference spend as a share of the function it serves. If your firm is in the 63% reporting no EBIT impact, the useful question is not whether AI works but whether your workflows were redesigned or merely augmented, since McKinsey’s high performers are separated by organisational change rather than tooling. What would change this verdict is a 2027 survey where high performers break out of the 6% band; if that number is flat a third year running, the constraint is structural, not temporal. Today’s lead shows the same dynamic on the supply side, where efficiency gains are real but arrive on the vendor’s timetable.