Here is a statistic to keep any executive awake. In McKinsey's own survey, eight in ten workers say AI has made them more productive. Yet the share of organisations reporting that AI has added to their operating profit sits at 37 percent, essentially unchanged from a year ago. The tools are clearly working for the people using them. The gains are evaporating somewhere between the desk and the income statement, and figuring out where is becoming the defining question of corporate AI.
Eric Kutcher, chair of McKinsey North America, put it bluntly this week: "We have not gotten enterprise-level productivity, with the exception of a few areas." His diagnosis is that early adopters confused activity with results. Companies handed staff access to Claude and ChatGPT and told them to use it as much as possible; some even ranked employees by how many tokens they burned, Kutcher said, "without realizing those tokens cost money." Workers duly got faster at their tasks, then filled the freed-up time with other tasks. Individual speed went up. The shape of the business did not change.
An EY survey of 1,200 chief executives lands in the same place from the top down. Half named AI as their single biggest source of productivity gains over the past year, yet 23 percent said those gains were being absorbed before they reached the bottom line, and only 16 percent had clear, real-time visibility into what AI was costing them versus returning. As Harvard's Iavor Bojinov notes, there is often a "productivity J-curve," where performance dips before it climbs, because the hard work of rewiring a process comes before the payoff. The crude error, one economist warns, is assuming time saved automatically equals money saved. It does not, unless the saved hours are converted into more output, lower cost or better decisions.
The firms that are seeing returns tend to share a discipline: they measure against a baseline before they deploy. General Motors says it starts each AI project with a specific engineering problem and a defined success metric, and only scales what demonstrably moves it. Its software testing now catches ten times as many defects, earlier, which lets engineers explore more designs rather than simply finishing the same work faster. The company is candid that not every gain can be pinned on AI, which is precisely why it bothers to establish a comparison first.
A sharper version of the argument comes from the engineering trenches. Writing from an AI accounting startup, one founder argues that the next wave of enterprise value will come from making models think less, not more. In back-office work like reconciling transactions or classifying expenses, variation is not a feature but a liability: there is one right answer, wanted the same way every time. The trick that worked for them was to stop asking the model to do the accounting and instead ask it to write deterministic code that does, something they can test, audit and rerun. They even found a small, specialised model that beat frontier systems on their task at two percent of the cost. The smartest model, it turns out, is often the wrong one.
Stitch these together and the paradox dissolves. AI reliably accelerates the work people already do, which shows up in personal productivity and almost nowhere else. The returns arrive only when a company redesigns the work itself, deciding which decisions to automate, which to keep human, and how to measure the difference. That is slower, harder and far less impressive to demo than a chatbot writing an email. It is also, increasingly, the only thing that reaches the bottom line.