A big study found no link between how much workers use AI and their output

A working paper from OpenAI’s own researchers has produced an awkward result for an industry that sells artificial intelligence as a productivity engine. Across more than a thousand companies, how heavily employees leaned on ChatGPT showed no measurable relationship to how much revenue each worker generated. The finding sits deep inside a dense analysis, and it complicates the central promise that has justified hundreds of billions of dollars in corporate AI spending.

What the OpenAI paper measured

The study examined more than 17 million messages sent through ChatGPT Enterprise by workers at over 1,500 organizations, then lined that activity up against each firm’s financial performance. According to Fortune’s account of the 69-page paper, the researchers controlled for factors such as company size and industry before testing whether heavier use tracked with stronger results. The answer was no: revenue per employee was not meaningfully associated with the number of messages workers sent, nor with the volume of AI-generated tokens consumed per person. Firms that flooded the tool with queries were, on the financial measure the researchers chose, indistinguishable from firms that used it lightly.

Why usage and output can drift apart

The disconnect is less surprising on close inspection than the headline number suggests. Message counts measure activity, not value. A worker who fires off a hundred low-stakes prompts a day registers as a heavy user, while a colleague who asks one carefully framed question that reshapes a report barely moves the meter. Revenue per employee, meanwhile, is a blunt, top-line figure shaped by pricing, capital, management, and market conditions that dwarf any single software tool. A drafting aide or a research assistant can save real hours without those hours surfacing as a clean line on a balance sheet, especially when the savings are spread thinly across thousands of small tasks rather than concentrated in a few measurable ones.

The limits of the measurement

The study’s own framing matters as much as its result. Revenue per employee is only one lens, and a coarse one. It captures nothing about quality, speed, employee retention, or the value of work that never generated a sale in the first place. A support team that resolves tickets faster or a legal department that reviews contracts in half the time may deliver enormous benefit that a revenue-efficiency ratio simply cannot see. The absence of a correlation is not evidence of absence; it is evidence that the chosen yardstick and the chosen activity metric do not line up, which is a genuinely useful warning against measuring AI success by counting keystrokes.

The gap between adoption and returns

The result lands in a stretch when adoption has raced ahead of proof. Corporate boards have committed to sweeping AI rollouts on the expectation that heavier use would translate into leaner operations and higher margins, and OpenAI has documented the sheer scale of that use in its own reporting on how people are using ChatGPT. Adoption is not the same as impact. The pattern echoes a long-running theme in economics: new general-purpose technologies often show up everywhere before they show up in the productivity statistics, because organizations need time to redesign workflows, retrain staff, and shed the old processes the technology was supposed to replace. Simply layering a chatbot on top of existing routines rarely rewires how a company actually makes money.

Echoes of the productivity paradox

Economists have watched this movie before. The famous quip that computers were visible everywhere except in the productivity statistics captured a decade in which businesses poured money into information technology while national productivity growth stalled, only to accelerate years later once firms reorganized around the new tools. The lag was not a sign the technology was worthless; it was a sign that value from a general-purpose innovation arrives on a delay, after the hard work of restructuring is done. Artificial intelligence may be tracing the same arc. Early rollouts tend to bolt a chatbot onto processes designed for a pre-AI world, capturing scattered convenience but not the wholesale gains that come from redesigning the work itself. If that history repeats, a snapshot taken now would naturally show little connection between usage and financial output, because the reorganization that turns adoption into advantage has barely begun. The OpenAI finding, read against that backdrop, looks less like a verdict on the technology and more like a status report on how early it still is.

What it means for the AI investment case

None of this proves the tools are useless. The paper does not claim that AI destroys value or that individual workers gain nothing; it reports that raw intensity of use, at the company level, does not predict a firm’s revenue efficiency. That is a narrower and more sobering point. It suggests that the returns, if they exist, depend on how a tool is deployed rather than how often it is opened, and that executives measuring success by usage dashboards may be watching the wrong dial. For a sector whose valuations rest on the premise that AI adoption compounds into financial advantage, a study bearing the OpenAI name that finds no such compounding is a caution worth reading closely. The likelier lesson is that value comes from a handful of well-chosen applications, disciplined process changes, and the patience to let both settle in, not from sheer volume of prompts.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview