How to Read AI Statistics Without Being Fooled

"78% of organisations use AI." That is what Stanford's 2025 AI Index reported for 2024, against 55% the year before. The figure is correct and striking, but it answers only one question: how many organisations have tried it? It says nothing about the other: how much did they gain?

First: adoption is not impact

"Use" is a wide word. It covers a trial in one department, a pilot that never left the lab, and deep integration into the core of the work. The same index gives a quieter picture of impact: among respondents using AI in service operations, 49% reported cost savings, and most said these were under 10%. Among those using it in marketing and sales, 71% reported revenue gains, and the most common level was under 5%.

In McKinsey's latest survey (May and June 2026, with 1,719 participants in 97 countries), 37% said AI had some effect on their earnings before interest and taxes, about the same as the year before, while the share scaling AI across the whole organisation rose from 38% to 44%. It seems that expansion is outpacing measured impact.

Second: the average hides the spread

When a tool is said to "raise productivity", ask: for whom? In a study of 5,179 customer-support agents, productivity rose by about 14% on average, and by about 34% for the least experienced, while the effect was minimal for the most experienced. The average merged two different groups.

In the Harvard and Boston Consulting Group experiment with 758 consultants, performance improved inside the model's "jagged frontier", and participants were 19 percentage points less likely to reach the correct solution on a task outside it. The mean alone does not reveal this split.

Third: "automatable" is not "will be abolished"

Alarming headlines about jobs often rest on studies that measure how susceptible occupations are, not whether loss occurs. Frey and Osborne's 2013 study estimated the probability of computerisation for 702 detailed occupations and concluded that about 47% of American employment was "at risk". It measures how susceptible occupations are to computerisation, not how many jobs will actually disappear.

The World Economic Forum, in its 2025 report on jobs (The Future of Jobs Report), gives both sides: by 2030 employers expect 170 million jobs created and 92 million displaced, a net growth of 78 million. These are estimates built on employers' expectations, and beneath the net lie groups that are displaced and others that are created; a positive net is no comfort to someone who is out of work.

Fourth: who is estimating, and why?

Market-size forecasts rest on assumptions: the pace of adoption, average spending, and the definition of the market itself. Change one assumption and the whole result moves. This does not mean every report is biased, but someone selling a service may be inclined to portray demand for it as larger than it is. So look at the method, not the headline.

Five questions before you believe a number

1. Who published it, and what is their interest?
2. Does it measure intention or action? Susceptibility or outcome?
3. What is the comparison? Is there a control group?
4. Is it an average hiding variation?
5. Does it concern our own context? Many circulating figures were measured in markets unlike ours.

It follows that we should build our own capacity for local measurement, learning the effect of a tool in our organisations through small, controlled trials before expanding. Verifying before relaying is a matter of religious etiquette before it is a statistical method. So when you meet a gleaming number, first ask: what exactly does this number measure?