On 11 August OpenAI published How Organizations Use AI: Evidence from ChatGPT: 69 pages, more than 1,500 organisations, more than 17 million messages, data through March 2026. It is the broadest dataset on enterprise AI use anyone has published, and it comes from the supplier who can see actual usage at its customers.
Trainees work with AI the most, boards the least
Six months after a company adopts ChatGPT, usage falls as seniority rises, at every level. Early-career workers and trainees send eight to nine more messages a week than the average active user in the same firm. Managers, directors and executives send fewer. By job title, “Executive / Founder / Partner” is among the lowest categories in the study, alongside finance and accounting. Analysts and marketing and communications staff are highest.
The researchers ran this two ways: once comparing all users in the sample, once comparing each group only with colleagues in its own company. Both agree, so the gap exists inside individual organisations rather than between them.
How much that gap is worth cannot be read from this paper. The study reports only deviations from the average and never publishes the average itself, so eight messages could be a third of a typical week or twice one. The authors add their own warning: message volume measures usage intensity, not economic importance. A chief executive sending three well-framed queries may get more out of them than a trainee sending fifty.
The direction itself is clear, and it is easy to check at home: you have the admin console and the user data. If your distribution looks like the one in the study, the people approving the AI invoice work with the tool less often than anyone else in the organisation.
What the study does not measure
Two days after publication, Fortune’s Emily Forlini went through the fine print and wrote up a table showing no link between AI usage intensity and revenue per employee. Headline: “More AI doesn’t mean more money.” The table exists, but the researchers were testing something different from what the headline implies.
The table takes a company’s financial results from one year and asks whether they predict how heavily that company used ChatGPT the following year. Finances first, usage second. The question was whether already-profitable firms become heavy AI users. The authors found no clear signal and describe their own numbers as too imprecise to support a strong conclusion.
Elsewhere the same paper shows that the heaviest users have higher revenue and higher market value per employee than everyone else, including after adjusting for industry and size. That is a correlation rather than a causal link, and the direction is not even known. Did AI lift those companies’ results, or did good results let them buy the tool early, hire the people and pay for the integration? The authors lean towards the second and say so openly.
They close the question themselves in the limitations section. The findings, they write, “do not measure downstream work products, productivity effects, or changes in organizational routines”, and future research “should connect enterprise AI telemetry to measures of output, organizational change, and longer-run firm performance”. A few lines later: “The rapid adoption of generative AI by firms should therefore not be equated with immediate productivity transformation.”
The paper has four tables, and in none of them is revenue the thing being explained. In all of them the thing being explained is AI usage. OpenAI held usage data and company financial data in one place and did not calculate how the first affects the second.
That calculation is hard to do honestly. Companies choose whether to adopt, the sample covers a narrow band of US-listed firms, and three quarters of data will not separate an AI effect from a good year. The paper is marked a working draft and claims nothing beyond what it measured. Worth remembering at the next presentation where someone cites “OpenAI’s research” as evidence of return.
The same week: two departures from the leadership
The study appeared in the middle of a leadership rebuild ahead of the listing. Axios called it a “pre-IPO refresh”.
On the same 11 August, Brad Lightcap left, OpenAI’s longtime chief operating officer, to “start something new”. Two days later Denise Dresser left, previously chief executive of Slack and hired in December to build the enterprise business. She lasted eight months, and the reason given was “other opportunities”. Her replacement is Dali Rajic, president and COO of Wiz. His brief includes helping customers measure the impact of their deployments. A day later CNBC was writing about a “huge red flag” ahead of the listing.
On the Monday of that week Bloomberg reported that the company had bought back $7bn of employee shares. Fortune spoke to two former staff who each received about $10m from it, earned over years of work and paid out in a single day.
In one week a company selling AI to large enterprises replaced its head of sales and published research showing that the effect of those sales cannot yet be measured. I am not suggesting one follows from the other. I am pointing out that the measurement problem turned out to be real enough to appear in a new executive’s job description.
Impact on the Polish market
The financial section of the study contains no Polish companies. The sample is a selected set of US-listed firms, and among adopters the median revenue was $2.3bn with median headcount close to 3,000. That is a different league from most readers of this newsletter, and a different way of buying technology.
The seniority result transfers without qualification, because it is about people rather than balance sheets. It is also the one part of this study you can verify in your own organisation within an hour: pull the distribution of active users by job level from your console and see where the real usage sits.
There is one difference that works against us. The study covers companies with a ChatGPT Enterprise licence, meaning their own account, console and usage data. A large share of Polish deployments run through Microsoft licensing and a solution built by an integrator, where the client receives an invoice and the detailed usage data stays with the supplier. Without that data you cannot check the usage distribution. Access to it has to be written into the contract.
Six questions before you approve the AI budget
- Who actually uses it. What is the distribution of active users by seniority? If it looks like OpenAI’s, next year’s decision is being taken by the people with the least contact with the tool.
- The baseline. What was measured before each deployment went live? Without a number from before the start there is nothing to compare a year later, and the return discussion comes down to impressions.
- What would have happened anyway. Who estimates how much of the improvement would have arrived regardless, through the market, new clients or other changes? Without that, every good year gets credited to AI.
- Where the number came from. When someone cites a study, check what it treated as the measured outcome. Most AI research measures adoption and usage rather than financial results, and does not answer the question you are asking.
- The renewal scenario. If the measurement shows no improvement in a year, do you renew? Settle it now, while nobody is defending a decision they already made.
- The scope of measurement. Are you counting only the corporate licence, or also API tools, internally built applications and staff personal accounts? OpenAI’s study covers the first of those and says so.
Briefing
Anthropic’s investors are targeting a $2tn listing in October. Several confirmed to the Financial Times on 13 August that they expect an offering larger than SpaceX’s $1.77tn June debut, on revenue projected at $100–120bn annualised by year-end (Fortune’s summary sits outside the FT paywall). That is roughly 17 to 20 times annual revenue, on minimal profit. CNBC reports CFO Krishna Rao is holding early investor meetings without discussing valuation. So what: at that multiple the company has to keep raising revenue per customer sharply, and the simplest source of that growth is renewals with existing enterprise clients.
IBM has joined OpenAI’s top partner tier. The partnership announced on 13 August puts GPT-5.6, Codex and ChatGPT Work inside IBM Consulting Advantage. IBM plans to train and certify tens of thousands of consultants across finance, government, telecom and retail. Weeks earlier Arvind Krishna admitted IBM “did not move quickly enough”. So what: IBM has stopped being neutral about model choice. An architecture recommendation from a partner certified by one supplier is a commercial recommendation.
The median company spends $12 per employee per month on AI. The top percentile spends hundreds of times more. a16z published charts on 14 August drawn from the Ramp AI Index of 12 August, covering model subscriptions, coding agents, API and GPU cloud spend. Median spend per employee per month rose from about $3 in January 2024 to about $12 in July 2026. The top 10% of firms spend around $650, and the top 1% reach $7.5K per employee per month. a16z notes that Ramp’s data skews towards technology companies. So what: the question “what do other companies spend on AI” has no useful answer, because the spread runs to three orders of magnitude. An industry benchmark will not justify your budget in either direction.
Summary
OpenAI published the broadest study yet of enterprise AI use: 1,500 organisations, 17 million messages, data through March 2026. Its strongest result concerns seniority. Six months after adoption, trainees and early-career staff send eight to nine more messages a week than the average user in the same firm, while executives are among the categories using it least often. The gap exists inside individual organisations. The study never publishes an absolute average, so the scale of that gap cannot be read from it, and the authors warn against confusing message counts with economic importance. On return the paper says nothing: revenue is not the explained outcome in any of the four tables, and OpenAI published no calculation of how AI use affects company results despite holding both datasets in one place. Fortune’s headline about no link between AI and earnings per employee describes a table in which the researchers were testing something different. The publication landed in a week when OpenAI’s chief operating officer and head of sales both left, and the incoming sales chief was given the job of helping customers measure impact. You can check the seniority distribution in your own organisation within an hour, provided your contract gives you access to the usage data.
Stay balanced, Krzysztof
Krzysztof Goworek is founder of Quintant — AI advisory that gets enterprises from experiment to production value.