OpenAI's analysis of ChatGPT Enterprise usage finds that large companies use the tool less intensively than smaller ones, but not because their engaged users are less engaged.
The gap entirely reflects penetration—how many employees use it at all—not how intensively those who do engage with it.
This suggests that most enterprise AI enablement efforts, which focus on training and engagement after employees have already adopted the tool, are addressing the wrong bottleneck; the real question is why employees never log in a second time.
What happened
OpenAI, Columbia, and Wharton published a working paper analyzing ChatGPT Enterprise usage through March 2026. Among firms that have adopted the tool, larger companies show lower usage per employee—fewer messages, fewer weekly active users, and fewer tokens per head, all statistically significant. However, a fourth measure—messages per weekly active user—showed no significant relationship to company size, suggesting the gap stems from how many people use the tool, not how intensively they use it once they do.
Why it matters
The finding reframes where the real constraint lies for enterprise AI deployment. If the problem were engagement (how intensively active users work with the tool), training and enablement programs would fix it. Instead, the data shows the problem is penetration—getting employees to log in and use the tool at all. This means most organizational effort on AI adoption is focused on the wrong phase: companies are optimizing post-onboarding when they should be investigating why employees never return after their first session.
What to watch
The paper also found that firms most likely to adopt AI early were those that had already invested heavily in organizational and managerial overhead (measured by SG&A spending in 2021), not engineering capability. Early-career workers and trainees send roughly eight to nine more messages per week than average active users, while executives send fewer. The usage pattern suggests junior staff are producing high volumes of written material while senior staff consume summaries, a dynamic for which existing review processes may be poorly designed.
Ask the AI about this article →
The paper's headline findings—usage grew sevenfold and adoption skews toward larger firms—are unsurprising. The deeper insight lies in disaggregating the enterprise AI gap. Among adopters, larger headcount predicts lower usage per employee across three metrics (messages, weekly active users, output tokens), all statistically significant. Yet the fourth measure, messages per weekly active user, shows no significant relationship to company size. This null result is the paper's most consequential finding: it isolates the constraint. The large-firm usage deficit is not a problem of engagement—active users at a 200,000-person company are as engaged as active users at a 500-person company—but entirely a problem of penetration. Few employees in large firms are using the tool at all.
This distinction divides the addressable problem space. If engagement were the bottleneck, the solution would be training, better prompts, and internal evangelism—precisely what the enablement industry has built itself around. If penetration is the bottleneck, those interventions are irrelevant; they apply only after the employee has already chosen to return, which is happening less often in large firms. The paper offers no explanation for why non-returners don't return, and that absence is the urgent question. The dataset has no record of who was offered access and declined, or of second sessions that never occurred. Whatever barriers exist are operating before any post-onboarding machinery engages, a finding the paper acknowledges briefly but does not pursue.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Astromech, an AI startup co-founded by Ben Lamm and geneticist George Church, raised $20 million in funding le…
Alibaba is guiding its AI cloud revenue toward a US$10 billion run-rate in the next quarter, signaling that it…
Ode, a venture arm of Anthropic, has acquired Casper Studios to expand its enterprise artificial intelligence…
SK hynix, the world's largest supplier of high-bandwidth memory (HBM), has published a technical roadmap for c…
Alibaba Group's June quarter results show cloud AI revenue growing 45%, while capital expenditure (investment…
Anthropic launched Claude Academy on August 20, a free learning site that explains AI fundamentals and how to…
