Enterprise AI ROI: OpenAI Data Shows No Revenue-per-Employee Link

A null result on page 35

OpenAI's newest enterprise report spends most of its length on rapid growth in AI use. One small table near the middle tells a different story. On August 11, OpenAI published a 69-page study of how businesses use ChatGPT. The main message is that usage is rising fast across every job function and seniority level, and that companies leaning into AI, especially agents that take on delegated tasks, are opening a "frontier gap" over everyone else12. The table on page 35, though, shows no statistically significant relationship between a company's revenue per employee and how heavily its staff use AI, whether that is measured in messages sent or tokens consumed11.

The report puts it plainly: once other controls are added, revenue per employee "is not meaningfully associated" with output tokens per employee or messages per active user11. According to coverage of the paper, the researchers looked at more than 17 million ChatGPT Enterprise messages from over 1,500 organizations and compared usage with financial data while controlling for factors such as company size and industry13.

For a company whose business case depends on selling enterprise seats and tokens, this is an awkward thing to publish in its own report. It is also not the whole story, so it helps to be clear about what the result does and does not show.

What the finding actually says

The report does not say that AI-using companies are less productive. It says companies with high revenue per employee tend to have adopted ChatGPT early, and that heavy users tend to have higher revenue per employee overall11. In other words, large and profitable firms were more likely to sign up early. The report just can't show that using more AI made them more money1112.

This is a classic selection problem. Rich companies can afford new tools sooner, have the IT budgets to roll them out, and often employ the kind of knowledge workers who find chatbots useful right away. Raw correlations would make AI look like a profit engine. Once the researchers control for who adopts, that apparent effect goes away. Our reading is that the most honest sentence in the report is also the most important one for buyers: usage does not measure value.

The outlets covering the study agree on the facts. They differ mainly in tone. Fortune's analysis, republished widely, stressed how deeply the result was buried and how it sits beside the upbeat "frontier gap" framing1214. Allwork took a more neutral line, saying the finding challenges the assumption that more adoption automatically brings better financial results13. A Substack critique was the bluntest, arguing that if OpenAI's own data can't show AI pays for itself, CFOs should stop treating vendor-funded ROI studies as evidence15.

The credibility question

One more detail adds a caveat. Two of the report's five authors are academics: David Holtz, listed with Columbia Business School, and Prasanna Tambe, listed with Wharton. A footnote says both worked on the report as paid contractors for OpenAI12. The other three authors are OpenAI employees12. Academic co-authors usually suggest outside independence. Here, as Fortune noted, that signal is blurred, and readers have no way of knowing whether analyses that went unpublished pointed in other directions1211.

Ironically, the conflict of interest makes the null result more believable. A vendor has every reason to hide a finding like this. Reporting it anyway, even in a small table, suggests the result was solid enough that leaving it out would have been hard to defend.

The rest of the report

Other parts of the paper make the ROI question harder to answer. Executives use AI less than anyone else. The chart on page 29 shows the most senior staff sending the fewest weekly messages per user12. Early-career workers are the heaviest users, sending about eight to nine more messages a week than the average employee at their companies13. OpenAI CFO Sarah Friar picked up on this, urging leaders to learn from the employees closest to the work12. The practical problem is that the people deciding whether AI spending paid off have the least first-hand experience of the tools15.

The report also shows that total enterprise output-token growth was almost flat from about October to December 2025, then turned sharply upward in January 2026. The chart stops at March 202612. On the same day the report came out, OpenAI named Dali Rajic as its new Chief Revenue Officer, replacing Denise Dresser after less than a year in the job12. Taken together, these details point to a company under real pressure to turn usage growth into lasting enterprise revenue.

A pattern that keeps repeating

The finding fits a gap that has shown up across this year's research: workers feel more productive, but companies struggle to show results on the balance sheet. McKinsey's 2026 State of AI survey of 1,719 executives found that only 6 percent of organizations qualify as high performers, meaning they attribute at least 5 percent of EBIT to AI. Meanwhile, about 80 percent of workers report productivity gains6. PwC's survey of 4,454 CEOs found that 56 percent saw neither revenue growth nor cost cuts from AI over the previous year1. KPMG put average enterprise AI budgets at $186 million, with only 8 percent of those companies reporting tangible ROI6. A February 2026 NBER study of nearly 6,000 executives found that close to 90 percent of firms saw no AI effect on employment or productivity over three years6.

OpenAI's earlier State of Enterprise AI report, from December 2025, already pointed in this direction. It found that ChatGPT Enterprise users saved an average of 40 to 60 minutes per workday22. That time savings was self-reported by paying customers21. Workers at the 95th percentile of usage sent six times as many messages as the median worker21. That report argued benefits grow with depth of use26. The new paper looks at the company level and does not find that depth of use shows up as revenue.

The difference is probably not a contradiction. Saving time only helps the business if the time goes into output that can be measured, like more deals closed, more code shipped, or more tickets resolved. Otherwise, as one analysis of the McKinsey data put it, the gain "stays on the worker's desk"6. IBM's research shows a similar split: 79 percent of leaders see productivity gains, but only about 29 percent are confident they can measure ROI5.

Where AI is paying off

The case studies that do show returns share a pattern. Analyses of the 2026 survey data find that companies reporting real ROI usually work on narrow, high-volume use cases with a clear before-and-after cost baseline, most often in customer service, back-office automation and operations3. McKinsey's respondents most often cite cost savings in supply chain, service operations and manufacturing6. Snowflake, for example, says it removed 1,600 hours of manual grading by having AI score sales-rep calls, with an estimated 4–5x return on time and spending2.

These wins come from redesigning a specific process, not from rolling out a general-purpose chatbot to everyone. McKinsey's Michael Chui said organizational change, not tool adoption, is what sets high performers apart6. OpenAI's December report also found that roughly one in four enterprises had not yet connected the assistant to company data, a basic step toward that kind of integration21.

The counterargument

There is a fair case on the other side. AI labs themselves have very high revenue per employee. Epoch AI estimated Anthropic and OpenAI at about $9 million and $5.5 million per employee, higher than any public tech company on the Forbes Global 2000. It suggested continued growth in that figure could reflect productivity gains from AI18. Revenue per employee is also a blunt metric. It tracks labor leverage, not profit, and it can rise just because spending moves from salaries to compute20. A company could get real value from AI through better quality or lower costs without any change in revenue per head.

That case has limits, though. AI labs sell AI, so their efficiency says little about a bank or a retailer that buys it. And if revenue per employee is the wrong yardstick, the burden falls on vendors to propose a better one, not on buyers to accept usage counts in place of results.

What this means

Our reading is that the page-35 table is the most useful part of OpenAI's report for enterprise buyers. It does not prove AI is worthless. It shows that token counts, message volumes and seat activations, the numbers vendors report most readily, are not business outcomes. The companies seeing returns choose a process, set a baseline, and measure the cost or revenue effect315. Everyone else is tracking activity, and OpenAI's own data now shows that activity alone does not show up in revenue per employee.