OpenAI's Own Enterprise Report Cannot Find the ROI

OpenAI published a 69-page report on August 11 on enterprise adoption of ChatGPT, and in a small table on page 35 its researchers report no statistically significant correlation between a company's revenue per employee and how much its people actually use the product, measured in messages sent and tokens consumed ^[1][2]. That finding sits underneath a week of capital movement that assumes the opposite: a staff tender offer that Bloomberg reports totaled $7 billion, a three-year-old Stockholm company doubling to a $13.3 billion valuation, and reports that Anthropic is planning a $2 trillion IPO in October, which would be the largest ever ^[13][14][15].

The report's own language is careful and worth reading twice. "Revenue per employee is not meaningfully associated with output tokens per employee or messages per active user once other controls are included," it says ^[3]. What the report does establish is a selection story: firms that already had higher revenue per employee tended to adopt ChatGPT early, and heavier users tend to be more lucrative companies in general, without the study showing that using more produces more ^[4]. That is the difference between finding that profitable firms buy software and finding that software makes firms profitable. The headline framing, according to Fortune's Emily Forlini, is a "frontier gap" in which adopters pull ahead of non-adopters, alongside exponential usage growth across seniority levels and job functions ^[5]. The intensity data on page 35 does not carry that weight ^[2].

A second buried detail complicates who is qualified to judge returns. On page 29, the most senior employees show the least intense use, with the lowest weekly messages per user, while early career employees use it by far the most ^[6]. OpenAI CFO Sarah Friar read that as a management lesson, writing on LinkedIn that "competitive advantage comes from the people closest to the work" ^[7]. Read less charitably, the executives signing the invoices have the least direct evidence of what they bought.

The usage curve has its own gap. On page 26, the total output token growth line is almost perfectly flat from roughly October 2025 to December 2025, about a quarter of no growth, before turning sharply upward in January 2026; the chart ends in March 2026 ^[8][18]. Fortune notes that Anthropic's Claude Code was becoming the default at many companies during that flat stretch ^[9]. OpenAI attributes the subsequent growth to new clients and to existing clients deepening use ^[10], and Sam Altman has been reorganizing the company around enterprise sales while cutting what it called "side quests," including the video app Sora ^[11].

The personnel move is the honest summary of all this. OpenAI named Dali Rajic chief revenue officer, replacing Denise Dresser after less than a year, with a mandate to accelerate customer adoption and help businesses measure impact ahead of the IPO ^[12]. A vendor hires someone to help customers measure impact when the customers cannot measure it themselves. Fortune also notes that OpenAI paid the academics who contributed to the report ^[16].

Worth watching: whether the next version of this report publishes an updated correlation table or quietly drops it, whether Rajic's measurement mandate produces outcome numbers before OpenAI files, and what Anthropic's October documents claim about customer returns ^[12][15].