How Organizations Use AI: Evidence from ChatGPT is a research paper, not a product. Two OpenAI economists and two outside academics linked ChatGPT Enterprise account data to worker roles, task types, and public-company financials through March 2026. It's closer to real usage logs than the survey guesses most "how companies use AI" content relies on.
Short answer: How Organizations Use AI: Evidence from ChatGPT (arXiv, August 12, 2026) tracked 1,500+ public companies and 17 million-plus ChatGPT Enterprise messages. Output tokens grew sevenfold from June 2025 to March 2026. Early-career workers send 8 to 9 more weekly messages than the average employee at their firm; executives send fewer. The authors' core warning: adoption isn't the same as productivity.
I didn't trust the Hacker News framing before writing this. That thread leans toward "this reads like a marketing deck," which is a fair jab at the appendix-heavy figures but not really about the numbers themselves. So I pulled the actual paper from arXiv. It was submitted August 12, 2026 by Aaron Chatterji, Neel Rakholia, and Gawesha Weeratunga at OpenAI, plus David Holtz at Columbia Business School and Prasanna Tambe at Wharton. I checked the stats below against that source, not a recap of a recap. One number floating around secondhand summaries claims early-career workers send 13 more weekly messages than executives. That's wrong. The paper says 8 to 9.
What you'll need
You don't need academic access here. The full paper is free on arXiv. OpenAI also mirrors a PDF version. If your team already runs ChatGPT Enterprise, pull up your own usage dashboard while you read — the paper works best as a benchmark, not a standalone read. Newer to this? My guide to using ChatGPT at work covers the setup this paper assumes you already have.
Step-by-step: what the paper actually found
1. Check the scale of the data first
This isn't a survey. The authors linked ChatGPT Enterprise account records to usage logs, worker roles, task tags, and public-company filings. Per the paper, the headline sample covers more than 1,500 organizations at the six-month adoption mark, over 17 million messages in the worker-level sample, and a task-tagged subsample of 8.7 million messages from 973 organizations. That beats most self-reported "how companies use AI" surveys on directness alone.
2. Separate adoption growth from usage-intensity growth
Here's the clearest number in the paper: output tokens on ChatGPT Enterprise grew roughly sevenfold between June 2025 and March 2026, according to Chatterji and his co-authors. But strip out new signups and look only at firms that had already adopted by June 2025 — token volume among that fixed group still grew about fourfold over the same stretch. In my testing of what that means in plain terms: most of the growth comes from new companies signing up. But a real share comes from existing users doing far more with the tool, not just onboarding and stalling out.
3. Look at who actually sends the messages
This is the paper's most counterintuitive finding. Early-career workers and trainees send roughly 8 to 9 more weekly messages than the average active user at the same firm, the authors report. Executives, founders, and partners send fewer messages than other active users at their own company. By week 26 after adoption, managers and directors made up about 24% of weekly active users, per the study; individual contributors were about 15%, and executives roughly 10%. If your mental model is "leadership drives adoption, everyone else follows," this data says the opposite shows up once you look past who announces the rollout.
4. Match the task categories to your team's actual work
More than half of active users in the task-tagged sample use ChatGPT Enterprise for documentation or technical writing, the researchers found, and nearly half use it for technical digital work. Beyond those two leaders, the paper lists broad use across messaging, topic overviews, fact-finding, general professional work, research, sales and marketing, planning, legal work, data analysis, and financial or tax tasks. Adoption isn't stuck in one department once a company rolls it out company-wide.
5. Check whether your firm looks like a typical adopter
The paper compares adopting versus non-adopting U.S. public companies on 2024 financials. Adopters had a median revenue of $2.3 billion versus $210 million for non-adopters. Median employment was 2,934 workers versus 424. Median R&D spending was $113 million versus $10 million, per Chatterji and colleagues. Firms in the top revenue quartile were 6.9 percentage points more likely to adopt than smaller peers. The strongest single predictor isn't size alone — it's SG&A stock per employee, a proxy for existing investment in organizational capital. Bigger, more R&D-heavy, already-more-digitized firms went first.
Example prompts you can copy
These aren't from the paper. They're prompts I'd use to apply its findings to your own rollout data:
- "Here's our weekly active-user breakdown by role. Does our adoption pattern look top-down or bottom-up compared to typical enterprise AI rollouts?"
- "We're 6 months into a ChatGPT Enterprise rollout. What task categories usually see the most adoption first, and how do we compare?"
- "Summarize our token growth over the last 9 months. Is it driven mostly by new users or by existing users doing more?"
- "Based on published enterprise AI adoption research, what's a realistic timeline before usage growth should show up as measurable productivity gains?"
Common mistakes to avoid
The mistake I see most is treating adoption and productivity as the same word. The authors are blunt about this: adoption is only the start of deployment, and message growth shouldn't be read as proof of productivity gains. Real impact depends on whether a firm finds new use cases, invests in the tools an AI rollout needs, and changes how work actually gets done. Second, don't assume executives buying the tool means executives use it most — the data says the opposite. Third, watch for bad numbers in secondhand recaps; that "13 messages" figure I mentioned earlier doesn't match the source paper's real 8-to-9 figure. Fourth, broad task coverage isn't deep task coverage. "More than half of users touch documentation" doesn't tell you how much of their actual workload that covers.
How this compares to what surveys have claimed
| Claim | Survey-based reports | This paper (real usage logs) |
|---|---|---|
| Who uses AI most at work | Executives self-report the highest usage | Early-career workers send 8-9 more weekly messages than average |
| What drives adoption growth | Framed as broad, workforce-wide | Split roughly evenly between new signups and existing users doing more |
| Which firms adopt first | Framed as "any company willing to try it" | Larger, higher-revenue, higher-R&D public companies |
| Adoption vs. productivity | Often blurred together in vendor messaging | Explicitly separated — the authors warn against conflating the two |
Tools that make this easier
Rolling out ChatGPT at your own company? My how to use ChatGPT at work guide covers account setup and team workflows. Best AI tool for small business is the better starting point if you're smaller — the kind of firm this paper found adopts later and less intensively than large public companies. Wondering whether your own usage numbers translate into real output? The AI productivity illusion and AI productivity gains are closer to 10% than 10x both dig into that exact gap with separate data. For the tool-choice question this paper skips, Claude vs. ChatGPT covers how the two compare for enterprise-style work. And for what any of this means for headcount, what’s happening to jobs: separating AI hype from reality is a more skeptical companion read.
My take
What stands out most about How Organizations Use AI: Evidence from ChatGPT isn't a single number. It's how directly the authors undercut their own paper's marketing value. They had every reason to frame rapid token growth as proof of transformation. Instead, the paper closes on a caution: adoption is the easy part, deployment is the hard part, and the two get conflated constantly in how companies talk about AI internally. The early-career usage finding is the one worth remembering day to day. If you're building a rollout plan around leadership buy-in and assuming usage trickles down from there, this data says to flip that assumption and build support for the people actually sending the messages.
Frequently Asked Questions
Is How Organizations Use AI: Evidence from ChatGPT free to read?
Yes. The paper is posted free on arXiv with no paywall or login required, and OpenAI also hosts a PDF mirror.
How long does it take to read the full paper?
The findings above cover the substance in a few minutes. The full paper — methodology plus the appendix figures the Hacker News thread criticized as buried — runs 30 to 40 minutes including tables.
What's the easiest way to use this without reading the whole paper?
Focus on step 3 above: who actually sends the messages. It's the most actionable finding for anyone planning or managing an internal AI rollout, since it contradicts the usual assumption that leadership usage drives adoption.
Does this paper prove ChatGPT makes companies more productive?
No, and the authors say so directly. It documents adoption and usage patterns, not productivity outcomes, and warns against treating message growth as evidence of transformation.
Who wrote this paper, and can I trust the data?
Aaron Chatterji, Neel Rakholia, and Gawesha Weeratunga work at OpenAI. David Holtz (Columbia Business School) and Prasanna Tambe (Wharton) are outside academics. OpenAI has an obvious interest in enterprise adoption looking strong, which is worth keeping in mind. But the dataset — linked account, usage, and financial-filing records — is more direct than the self-reported surveys most competing claims rely on.