gettyimages.com
The junior analyst two desks from the printer is doing more with ChatGPT Enterprise than the executive in the corner office — by a margin that the conventional corporate narrative about AI adoption cannot explain away.
That is the headline finding of a 69-page working paper posted to arXiv on August 12 by researchers at OpenAI, Columbia Business School, and the Wharton School. By linking more than 17 million real ChatGPT Enterprise messages from 1,764 organizations to worker job titles and seniority levels — and by connecting a subset of those organizations to public-company financial data — the paper delivers the first ground-truth picture of enterprise AI use at the scale needed to make population-level claims about who, inside adopting firms, is actually doing the work with AI.
The answer overturns received wisdom in two directions at once. Early-career workers and trainees send roughly eight to nine more weekly messages than the average active user within their own firm. Executives, founders, and partners send fewer. Marketing and communications workers and analysts also out-message executives, even though they represent a smaller slice of weekly active users. The researchers describe this as a strong negative seniority gradient in message volume — meaning the steeper your title, the less you use the tool.
Why Behavioral Logs Beat Surveys
The paper arrives at a moment when the enterprise AI conversation has been dominated by survey data — asking managers whether their employees are using AI, asking workers how often they reach for a chatbot, asking executives which workflows have been transformed. Survey-based adoption research, as the authors note, “may suffer from imperfect recall or reporting biases.”
The Chatterji-Holtz-Rakholia-Tambe-Weeratunga paper takes a different path. The primary dataset is an organization-week panel constructed from ChatGPT Enterprise account records covering adoptions between January 2024 and March 2026. The worker-level sample — the one that yields the seniority finding — covers 1,764 organizations and 17,446,551 messages at the six-month adoption horizon. No researcher manually reviewed any individual message. Job titles were mapped using SCIM identity management data, then classified by an automated system built on GPT-5-mini into broad categories covering seniority (early-career, individual contributor, senior IC, manager/director, VP, executive) and job function. A separate task classifier — a message-level system available from October 30, 2025 — assigned each message to one of 60 work task categories within a two-level taxonomy.
What the paper captures, in economic terms, is “telemetry data” — actual behavioral logs — rather than what workers say they do with AI. When you read the logs instead of the surveys, a different story emerges.
Large Firms First: The Complementary Capabilities Gap
Before addressing who uses enterprise AI within adopting organizations, the study documents which organizations have adopted in the first place. The picture is striking.
Among U.S.-based public companies linked to the Compustat database, ChatGPT Enterprise adopters are outliers relative to the general corporate population. The paper’s adopter financial profile data shows median revenue of $2.28 billion among adopters, compared to $210 million for non-adopters. Median total assets come in at $4.4 billion versus $668 million. Median employment is 2,934 workers versus 424. Median market value among adopters reaches $5 billion — approximately 16 times the $316 million median for non-adopters. Median R&D spending is $113 million among adopters, compared to $10 million among non-adopters.
These differences hold even after the researchers control for industry and firm size. Among firms in the top revenue quartile within their industry-year peer group, adoption probability is 7.2 percentage points higher than for comparable peers. For the top 5 percent, the gap widens to 11.3 percentage points.
The authors frame this through a concept borrowed from the economics of general purpose technologies — the idea that transformative innovations like steam power, electricity, and IT require “complementary capabilities” before they yield gains. Firms with larger accumulated stocks of R&D investment, sales and marketing spending, and capitalized software are all significantly more likely to adopt ChatGPT Enterprise, independent of their size. The implication, drawn from Bresnahan and Trajtenberg’s foundational 1995 work on general purpose technologies, is that access to the same underlying AI system does not produce uniform adoption because the infrastructure to deploy that system effectively — in workflows, in training, in organizational design — has to be built up first.
“Larger and more organizationally intensive firms may be better positioned both to identify valuable use cases and to deploy and integrate the technology into existing workflows,” the paper concludes on this point.
This finding has a sobering corollary for competitive dynamics. If the firms best positioned to adopt enterprise AI are also the firms already most productive, most highly valued, and most R&D-intensive, the technology may be a force for widening the gap between large and small competitors — at least in the early diffusion period — rather than a democratizing equalizer.
Usage Is Compounding From Inside
The pace of usage growth inside already-adopting organizations is the second major finding, and it complicates the simple story of “more companies joining means more usage.”
Aggregate output tokens produced by ChatGPT Enterprise customers grew sevenfold in nine months, from June 2025 to March 2026. But roughly half of that growth was not driven by new organizations joining the platform. Among firms that had already adopted before June 2025, token output grew approximately fourfold over the same period.
The researchers also note that usage accelerated in early 2026 across all adoption cohorts simultaneously. The pattern — organizations that came aboard at different times all experiencing the surge at the same time — is inconsistent with normal post-adoption expansion curves, which unfold gradually as individual firms build out use cases. “Because organizations that adopted at different times experienced this acceleration simultaneously,” the paper observes, “the increase appears to reflect developments affecting ChatGPT Enterprise customers broadly rather than only the normal expansion of use following adoption.” Something about the product improved in a way that lifted usage across the installed base at once.
Who Is Actually Using It: The Seniority Inversion
The most counterintuitive finding concerns the distribution of usage intensity within adopting firms.
Two common narratives have shaped how corporate AI adoption is discussed. The first: AI will hollow out entry-level roles because junior workers perform the most automatable tasks. The second: senior executives will be the power users, leveraging AI for strategic analysis and high-level decision support. The telemetry data contradicts both.
At the average adopting firm six months after deployment, managers and directors account for approximately 24 percent of weekly active ChatGPT Enterprise users, with individual contributors at 15 percent and senior ICs at 14 percent. Early-career workers and trainees account for 7 percent. Executives account for 10 percent. So the senior layers of the firm contribute a meaningful share of active users.
But active user share and usage intensity are different things. When the researchers measure messages per active user — how much each category of user does with the tool when they use it at all — the pattern reverses. The early-career and trainee group sends eight to nine more weekly messages than the average active user within the same firm. Managers and directors send fewer. Executives send fewer still.
The result is a firm-level picture where the population driving the most interaction with enterprise AI is the most junior one. The paper’s seniority gradient analysis notes this is “especially relevant in light of recent evidence on generative AI and early-career labor-market outcomes, because it identifies early-career workers as particularly intensive users conditional on active use.”
The finding aligns with, and now extends to real-world organizational data, a result from a landmark experimental study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond published in the Quarterly Journal of Economics (2025). That study, built on data from more than 5,000 customer-support agents, found that AI tools boosted productivity most among less experienced and lower-skilled workers — with substantial speed and quality gains for novice agents and minimal effects on the most experienced. The Chatterji et al. paper shows that this population — the cohort with the most to gain from AI access — is in practice the one choosing to use it most intensively, at enterprise scale, not just in a controlled experiment.
What Everyone Is Actually Doing With It
If early-career workers are the intensive users, what are they — and their colleagues — doing?
The task-classification analysis (drawn from the 973-organization, 8.7-million-message subsample with task classifier data available at the six-month adoption horizon) reveals that ChatGPT Enterprise has become a genuinely broad-based knowledge-work tool.
More than half of active users engage in documentation or technical writing in a given week. Close to half engage in technical digital work. Large shares also use the tool for communication drafting, research, planning, information synthesis, data analysis, legal and regulatory work, and financial tasks. The top tasks by message volume — documentation and technical writing, technical digital work, and message drafting — are common across virtually all knowledge-work roles. But specialist use is also evident: engineering and technical staff skew toward debugging; finance and accounting workers lean into financial and tax tasks; sales and marketing roles show elevated use in sales and marketing tasks.
The breadth of this task distribution is itself a finding. Generative AI in enterprises is not concentrating in a single workflow or department — it is being used across the organization in a pattern consistent with what economists call a “general purpose technology for knowledge work,” a phrase the authors borrow from Bresnahan and Trajtenberg’s broader framework for technologies like electricity and computing that, when fully diffused, show up in nearly every sector of the economy.
What Usage Data Cannot Resolve
The Chatterji et al. finding that junior workers use enterprise AI most intensively needs to be read alongside a separate, uncomfortable strand of evidence — one the paper itself cites but does not directly resolve.
Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at the Stanford Digital Economy Lab have been tracking what they call “canaries in the coal mine” for entry-level employment since 2025. Their most recent update, extending through June 2026 and drawing on administrative payroll data from ADP covering millions of U.S. workers, finds that entry-level AI-exposed employment for workers aged 22 to 25 now stands 19 percent below where it would be had it kept pace with their less-exposed peers. The divergence has widened consistently since late 2022 and shows no sign of reverting.
These two findings are not contradictory — they can both be true simultaneously. Junior workers inside firms that have already adopted ChatGPT Enterprise are using it most intensively. And employers, aware that AI can substitute for some of what entry-level workers do, may be hiring fewer junior workers in AI-exposed roles in the first place. The intensive users that the Chatterji et al. study documents are the ones who made it in. The Canaries data suggests the cohort that could have benefited from enterprise AI access is shrinking as a share of the workforce.
This gap — between the productivity gains AI offers junior workers who have access and the employment contraction happening at the entry level in AI-exposed industries — is the most significant question the data opens rather than closes.
Why You Should Read This Study Carefully
The paper merits substantive engagement, but also a calibrated skepticism. The dataset is constructed from OpenAI’s own platform — meaning OpenAI controls both the data and, effectively, the narrative. Three of the five authors (Chatterji, Rakholia, Weeratunga) are full-time OpenAI employees; the other two (Holtz of Columbia and Tambe of Wharton) contributed as paid contractors for OpenAI. The paper does not use this fact to discredit its own methodology — the privacy-preserving telemetry approach is sound, and privacy-protecting analysis of this kind has no obvious analytical bias — but the institutional alignment deserves acknowledgment. No independent researcher has yet replicated or challenged the findings, and the paper carries the “working paper” designation; results are subject to change.
The paper also covers only ChatGPT Enterprise — not usage through personal accounts, API access, or competing products from Anthropic, Microsoft, or Google. Job title coverage within included organizations is incomplete, and the seniority results describe the subset of users whose titles were successfully classified, not the complete workforce. The public-company financial analyses are restricted to U.S.-listed firms linkable to Compustat, which already skew large relative to the full population of businesses.
A parallel dataset from Microsoft — the May 2026 paper on M365 Copilot Chat usage, covering approximately 5.5 million sessions — finds broadly similar patterns: writing dominates, but task use is broad across knowledge-work categories. That convergence between two large-scale enterprise platforms supports the general picture of broad task adoption without settling the seniority gradient question, which Microsoft’s paper did not directly examine.
Adoption Is Only the Beginning of Deployment
One explicit purpose of the paper is to resist a common inference in commentary about enterprise AI: that because adoption has been rapid and usage has grown sevenfold, the productivity transformation is underway and measurable. The authors close with a reminder drawn from the GPT diffusion literature.
“Rapid adoption of generative AI by firms should therefore not be equated with immediate productivity transformation,” they write. “General purpose technologies rarely generate immediate, economy-wide gains; their impact unfolds through a slower process of co-invention in which firms discover use cases, invest in complements, and reorganize production so that a new capability becomes reliable in everyday work.” The most direct comparison point is computing and the internet: measurable economy-wide productivity gains from those technologies were delayed by years after adoption rates reached high levels, because firms needed to reorganize work around the new capability — not just install it. The authors draw on Brynjolfsson, Rock, and Syverson’s J-curve research to anchor this argument in the broader history of general purpose technologies.
“We are still in the early stages of that process,” the paper concludes. “Adoption is only the beginning of deployment.”
What This Means for Every Knowledge Worker Today
The practical stakes of this research are high enough to warrant explicit translation.
For junior workers: the data confirms that at firms where enterprise AI is available, early-career employees are already the most intensive users. That behavioral pattern is already shaping which junior workers will have the deepest AI fluency three years from now — an advantage that compounds as the technology embeds more deeply in knowledge workflows. The window to build that fluency is open now, and the Chatterji et al. data suggests that the workers who understand this are already acting on it.
For managers and executives: the seniority gradient is a signal worth interrogating. If early-career workers are sending eight to nine more messages per week than the average active user in the same firm, one of two things is happening. Either junior workers have found use cases that management has not yet recognized as productivity levers — in which case, the gap is an opportunity. Or junior workers are using the tool for lower-stakes tasks at higher volume, while senior employees are applying it more selectively to higher-stakes work — in which case, message volume is not the right metric of value. The paper cannot resolve which interpretation is correct, and it explicitly cautions against equating message volume with economic importance. Both interpretations deserve investigation at the firm level.
For corporate AI strategy: the deepening-within-existing-customers pattern — roughly half of all token growth coming from firms that had already adopted, not from new adoptions — suggests that the frontier of enterprise AI competition is not signing up more organizations. It is getting adopting organizations to build the workflows, training programs, and organizational habits that make the tool genuinely embedded in daily practice. The firms currently doing that most effectively are already the largest and most intangible-asset-rich in their industries. If adoption “widens existing firm heterogeneity,” as the authors suggest it may, the implication is that the race has already started and the leaders are already ahead.
Frequently Asked QuestionsWho uses ChatGPT Enterprise most intensively at work — junior employees or executives?
According to the Chatterji et al. working paper (arXiv:2608.12236), junior workers and trainees send roughly eight to nine more messages per week than the average active ChatGPT Enterprise user within the same firm, while executives, founders, and partners send fewer messages than the average active user. This is a finding about usage intensity among users who have already been activated on the platform — it does not measure the rate at which different seniority levels adopt the tool in the first place, which is a separate question the paper cannot answer from its data.
If AI helps junior workers most, why is entry-level employment declining in AI-exposed occupations?
These two findings can coexist. Erik Brynjolfsson and colleagues at Stanford have documented that AI-exposed workers aged 22 to 25 now face employment roughly 19 percent below trend relative to less-exposed peers, a gap that has widened consistently since late 2022. The intensive users the Chatterji et al. paper documents are those who made it into enterprise roles at adopting firms. The Canaries data suggests that the pool of workers gaining that access is contracting, particularly in tech and customer-service roles where AI substitution is most direct. These are two different measurements of two different phenomena — within-firm usage intensity versus economy-wide entry-level hiring — and they are not in conflict.
Which companies have adopted ChatGPT Enterprise, and how large are they?
The Chatterji et al. paper does not name specific adopting companies — that information would be commercially sensitive — but its Compustat-linked financial analysis of U.S. public-company adopters reveals the profile. Median revenue among adopters is $2.28 billion, compared to $210 million for non-adopters; median market capitalization is $5 billion versus $316 million for non-adopters. Adoption is especially concentrated among the top 5 percent of firms by revenue within their industry: the top 5 percent adoption gap reaches 11.3 percentage points above comparable peers. The study is limited to U.S.-based public companies and does not describe private-company or international adopter profiles.
Does rapid enterprise AI adoption mean productivity gains are already here?
Not according to the paper’s authors, who explicitly warn against that inference. Drawing on the economics of general purpose technologies, they note that transformative technologies like electricity and computing generated measurable economy-wide productivity gains only after years of organizational co-invention — firms learning where to use the technology, redesigning workflows around it, and building complementary capabilities. Usage growing sevenfold in nine months indicates widespread adoption; it does not indicate that the organizational reorganization required to capture productivity gains has been completed. “Adoption is only the beginning of deployment,” the paper concludes.