Review: McKinsey's State of AI 2026 - the number that did not move
McKinsey and QuantumBlack published The state of AI in 2026: On the road to ROI on 25 August 2026. The annual survey tends to end up in board packs, so it deserves a careful read.
What the report claims
Organisations are scaling AI more broadly, agentic AI is growing mainly in large enterprises, and coding agents are starting to change build-versus-buy decisions. Individual employees report clear benefits. The share of organisations that attribute any EBIT impact to AI, however, has not moved in a year.
The headline figures:
37% of respondents attribute at least some EBIT impact to AI, about the same as last year. High performers (at least 5% of EBIT attributed to AI, and impact described as significant) remain about 6% of respondents.
44% report AI scaling across the enterprise, up from 38%. Among organisations with more than $1 billion in revenue, 40% are scaling AI agents, up from 27%. Smaller organisations are flat at 22%.
32% say their organisation decided against buying at least one software product or feature because it could be built in-house with agentic coding tools.
About 20% say AI operating costs constrained their use of AI, while 60% expect to invest more over the next year.
80% say AI improved their own productivity. 14% of respondents at organisations using AI report an AI-related workforce decline over the past year, less than half of the 32% who expected one in the previous survey. 39% now expect declines in the coming year.
How the evidence was produced
An online survey, in the field from 4 May to 8 June 2026, with 1,719 participants in 97 nations. 36% work at organisations with more than $1 billion in annual revenue. Responses are weighted by each nation's share of global GDP. A panel of 552 people who answered in both 2025 and 2026 was used to check the year-over-year workforce comparison.
What holds up
The flat EBIT figure is the most credible number in the report, precisely because it flatters nobody. A publisher that sells AI transformation services has little to gain from reporting that enterprise-level impact stood still. The workforce comparison is also better grounded than most, because it was checked against a panel of people who answered in both years. The gap between the 32% who expected cuts and the 14% who report them shows how far headcount forecasts can overshoot.
What does not
Everything is self-reported. EBIT attribution, productivity gains and workforce effects are respondents' estimates, not audited figures.
Small cells. With 1,719 responses across 97 nations, any cut by industry or region, and above all the 6% high-performer group, rests on few answers. The article gives no confidence intervals and no breakdown for Europe or Japan.
A change of method. McKinsey notes that the 2026 survey measured impact at business-function level rather than aggregating from individual use cases. In our reading that limits like-for-like comparison of function-level figures with earlier editions.
Correlation, not cause. The high-performer story (redesign workflows, commit leadership, pursue growth as well as efficiency) describes what those organisations do. The survey cannot show that these habits produce the results.
What to do with it
Set board expectations with the 37%. Individual productivity gains do not show up in the P&L without workflow redesign and explicit measurement. If a business case rests on time saved per employee, it should say how that time turns into revenue or lower cost.
Put AI operating cost on the engineering dashboard now. About one in five organisations already limits use because of it, and high performers report cost constraints on coding agents about three times as often as others.
Revisit build-versus-buy for small internal tools, with eyes open. Budget for the maintenance, security and ownership of what agents build. A tool that was cheap to create still needs an owner.
Our verdict: a useful annual benchmark, as long as it is read as a survey of perceptions. Its most honest number is the one that stayed where it was.