Buried in OpenAI's latest research: No correlation between AI use and revenue per employee

OpenAI’s latest enterprise report has a glossy headline and a less flattering footnote. The 69-page study, published on August 11, celebrates rapid ChatGPT adoption across seniority levels and job functions. It also leans hard on a “frontier gap” story: companies using AI are supposedly pulling ahead of those that aren’t.

But the part buried on page 35 undercuts the sales pitch. The researchers found no statistically significant link between revenue per employee and how much AI workers used, whether measured by messages sent or tokens used. The report says revenue per employee is not meaningfully associated with output tokens per employee or messages per active user once other controls are included. What it does show is that companies with higher revenue per employee were more likely to adopt ChatGPT early. That is not the same thing as proving the tool made them richer.

The report also hints at who actually uses the stuff. Executives appear to be the lightest users, and senior employees show the least weekly messages per user on the page 29 graph. Early-career employees are the heavy users. That matters, because the people making the big calls may be the least able to judge whether the software is paying off. OpenAI CFO Sarah Friar pointed to that in a LinkedIn post, arguing that competitive advantage comes from the people closest to the work.

Then there’s the enterprise growth chart. OpenAI’s total usage within enterprises went flat from about October 2025 to December 2025, right when Anthropic’s Claude Code was winning over a lot of corporate users. The line then shoots up in January 2026, and OpenAI says that came from both new customers and deeper use by existing ones. CEO Sam Altman has been reworking the company around enterprise sales and cutting back on “side quests” like Sora, and OpenAI has now hired Dali Rajic as chief revenue officer to replace Denise Dresser, who stayed less than a year.

One more wrinkle: two of the five authors were paid contractors for OpenAI, even though they’re listed with Columbia and Wharton affiliations. That doesn’t make the report useless, but it does make the varnish thinner. The loudest number in AI is still the one on the funding slide. The important number is the one companies can prove in their own books.