How Organizations Use AI: Evidence From ChatGPT
Aaron Chatterji, David Holtz, Neel Rakholia, Prasanna Tambe, Gawesha Weeratunga · pdf · 2026-08-11 Reader 🖥️
AI Labor DisruptionAI as Normal TechnologyEnterprise AI Adoption & ReadinessMeasuring AI ImpactSkill Infrastructure (Ding)Technology Diffusion & Transformation Curves
Enterprise ChatGPT adoption and use, measured through OpenAI's own product telemetry across 1,500+ organizations and 17M+ messages, is a broad but highly uneven organizational phenomenon — concentrated among larger, more valuable, more intangible-capital-intensive firms, and diffusing unevenly across job functions, seniority levels and tasks within adopting firms — showing that organizations are still in an early, open-ended process of learning where generative AI belongs in their workflows rather than having already converted access into uniform deployment.
theoretical · expository — An empirical economics working paper (OpenAI + Columbia + Wharton authors) using administrative ChatGPT Enterprise telemetry linked to job titles, task classifications, and Compustat financials. It documents 'four stylized facts' about firm-level adoption and within-firm usage rather than arguing for a course of action, though the conclusion carries clear practical implications (adoption is not deployment; complements matter).
Structure
| Part | Role |
|---|---|
| Introduction and Related Literature (Sections 1-2) | context |
| Data and Measurement (Section 3): four linked samples — aggregate usage panel, job-title/industry sample, task-classification subsample, public-company Compustat-linked sample | evidence base / methodology |
| Fact 1 (Section 4.1): aggregate usage growth, decomposed into new adopters vs. within-cohort intensification | evidence |
| Fact 2 (Section 4.2): firm-level adoption predictors — size, revenue/employee, and pre-existing intangible/organizational capital (SG&A, R&D, capitalized software stocks) | evidence |
| Fact 3 (Section 4.3): within-firm heterogeneity in active-user composition and usage intensity by job title class, seniority, and manager status | evidence |
| Fact 4 (Section 4.4): task composition of use — breadth across ~60 task categories, with prevalence vs. message-share as distinct margins | evidence |
| Conclusion (Section 5): 'adoption is only the beginning of deployment' — general-purpose-technology framing, limitations, and implications for widening firm heterogeneity | conclusion / synthesis |
Key terms
- ChatGPT Enterprise
OpenAI's centrally administered, paid workplace product (including Codex usage bundled into the same offering during the study period) — distinct from personal ChatGPT accounts, API usage, or other subscription plans. The paper's entire evidence base is scoped to this one product surface. - Adoption vs. deployment
Adoption = a firm obtaining a ChatGPT Enterprise workspace. Deployment = the subsequent, uneven process of discovering use cases, spreading use across workers, and integrating the tool into workflows. The paper's central move is insisting these are separate events measured by separate margins (extensive vs. intensive; who vs. how much vs. for what). - Job title class / seniority level / people-manager status
LLM-inferred categorical variables (via a gpt-5-mini classifier, detailed in Appendix C) derived from raw, self-entered job title strings — not verified organizational role data, not consistently available for every user, and observed at a single point in time. - Extensive vs. intensive margin (of usage)
Extensive margin = whether/how many people use the tool at all (active-user composition, task prevalence). Intensive margin = how much active users use it (messages per active user, output tokens per employee). The paper repeatedly shows these two margins move independently or even in opposite directions across the same worker groups. - Complements / organizational and intangible capital
Pre-existing, non-AI investments — SG&A stock, R&D stock, capitalized software stock, all measured years before ChatGPT Enterprise existed — treated as proxies for a firm's accumulated capacity to identify, adapt to, and integrate a new general-purpose technology.
Propositions
- explicitAggregate ChatGPT Enterprise output tokens grew roughly sevenfold between June 2025 and March 2026, and grew roughly fourfold within the fixed cohort of firms that had already adopted by June 2025 — meaning about half of the token-consumption growth over this period occurred within already-adopting firms, not solely from new adopters.conf: high
- explicitAmong U.S.-based public companies, early ChatGPT Enterprise adopters are systematically larger, more valuable, and more R&D- and SG&A-intensive than non-adopters (e.g., median revenue $2,275.1M vs. $209.6M; median market value $4,997.2M vs. $316.4M; median employment 2,934 vs. 424 workers).conf: high
- explicitAdoption probability rises disproportionately at the top of the firm-size distribution, even within narrow industries: top-revenue-quartile firms are 6.9-7.2 percentage points more likely to adopt than the rest, and top-5%-revenue firms are 9.8-11.3 percentage points more likely, whether measured against the whole sample or within-industry-year cells.conf: high
- explicitPre-existing, employee-normalized stocks of SG&A, R&D, and capitalized software — all measured in fiscal year 2021, years before ChatGPT Enterprise existed — independently predict later ChatGPT Enterprise adoption after controlling for firm size and industry, consistent with adoption depending on prior accumulated organizational and intangible capital rather than firm scale alone.conf: highsupportsOrdinary Engineers, Not Heroic Inventors · p6 — Firm-level empirical evidence (SG&A/R&D/software stocks predicting adoption) for Ding's 'skill infrastructure' claim that GPT diffusion depends on accumulated organizational capability, not raw access.
- explicitConditional on having adopted, larger firms show lower per-employee usage intensity (messages, WAU, tokens per employee) even though larger firms were more likely to adopt in the first place — adoption and within-firm usage intensity are driven by partly opposing firm characteristics.conf: high
- explicitAmong adopting public firms, more intensive per-employee ChatGPT Enterprise use is associated with higher revenue-per-employee and market-value-per-employee, and this relationship survives adjusting for industry and firm size (though the paper explicitly declines to interpret it causally).conf: medium
- explicitSix months after adoption, active users are broadly distributed across job functions (at the average firm: engineering ~11%, executives/founders/partners ~9%, finance/accounting ~5%, marketing/comms ~5%, sales ~4% of weekly active users) and across seniority levels (managers/directors ~24%, ICs/professionals ~15%, senior ICs/principals ~14%, executives ~10%, early-career/trainees ~7%) — usage is not confined to technical roles or to one seniority tier.conf: high
- explicitUsage intensity conditional on being active shows a strong negative seniority gradient: early-career workers and trainees send roughly 8-9 more weekly messages than the average active user at the same firm, while managers, directors and executives send fewer — the seniority group with the smallest share of active users is also the most intensive per-user.conf: highextendsThe loudest warning about AI and jobs yet · p2 — Adds a concrete usage mechanism (early-career workers are the most intensive per-user AI users) that could plausibly help explain the parallel finding that early-career/entry-level jobs are shrinking under AI exposure.
- explicitActive-user composition (who uses the tool) and usage intensity (how much they use it once active) are distinct, sometimes divergent margins of diffusion: analysts and marketing/communications workers are disproportionately intensive users despite not being among the largest groups of active users.conf: high
- explicitChatGPT Enterprise use spans a long tail of roughly 60 task categories rather than concentrating in one workflow; documentation/technical writing, technical digital work, and communication dominate message volume across virtually every industry, job function and seniority level, while task prevalence (which tasks people try at least once) varies far more by industry and role than task intensity (which tasks dominate total message volume).conf: high
- implicitBecause breadth of task exposure (many active users trying many tasks at least once) and depth of message volume (a small set of task categories dominating total usage) diverge, 'AI use' inside an organization is not one phenomenon but at least two distinct diffusion curves that can mislead if conflated into a single adoption number.conf: medium
- explicitAccess to ChatGPT Enterprise does not imply uniform, effective, or immediate deployment; firms are still in an active, open-ended process of discovering where generative AI belongs in their workflows, paralleling historical general-purpose-technology diffusion (electrification, computerization) that took decades of complementary co-invention to pay off.conf: highsupportsArvind Narayanan on Why AI Isn't All That Revolutionary · p1 — Firm-level telemetry evidence for the thesis that AI's real-world impact unfolds over decades via adoption bottlenecks rather than overnight transformation.supportsAI as an Enterprise Operating System · p13 — Echoes the Solow productivity-paradox parallel this document draws — access without workflow reorganization does not by itself generate value.
- implicitBecause early and heavy adopters are already the larger, more intangible-capital-rich, higher-valuation firms, generative AI diffusion in this early period is likely to widen rather than narrow existing inter-firm heterogeneity in productivity and value, at least until AI capability and complementary organizational capacity diffuse more broadly.conf: medium
- implicitThe paper treats 'messages per active user' as a workable proxy for usage intensity and (implicitly, via the financial-outcome correlations in Section 4.2.2) as a signal of economic importance, even though it explicitly cautions in Section 4.3.2 that message volume measures intensity of use, not the value or completeness of the resulting work product.conf: medium
Arguments
ChatGPT Enterprise adoption depends not just on firm size but on pre-accumulated organizational and intangible capital — firms need 'complements' in place before they adopt. strength: medium
Explicit premises:
- SG&A, R&D, and capitalized-software stock coefficients remain positive (mostly statistically significant) after controlling for lagged size, assets/employee, PP&E/employee, and industry fixed effects.
- The pattern persists (though attenuated) when technology and high-R&D sectors are excluded from the sample.
Implicit premises:
- SG&A/R&D/software stocks measured in FY2021 are valid proxies for 'capacity to identify and integrate a new technology' rather than simply further correlates of being a large, well-run, well-capitalized firm in general.
- The direction of association plausibly runs from accumulated capability toward adoption, rather than from an unobserved third factor (e.g., general managerial quality, industry dynamism) driving both intangible investment and AI adoption simultaneously.
- Compustat-derived stock measures, built via a perpetual-inventory approximation (SG&A depreciated 20%/year, R&D 15%/year), are precise enough proxies for organizational capability that a 0.02 or even a 0.004 coefficient is economically, not just statistically, meaningful.
Inferences:
- If organizational capital this strongly predicts who adopts AI, then AI-adoption inequality across firms could persist or widen largely independent of anything AI vendors do on the product side — a market-structure implication the paper only gestures at in its conclusion rather than developing.
Adoption is only the beginning of deployment: obtaining ChatGPT Enterprise access does not translate into uniform or complete organizational integration.
Explicit premises:
- Usage intensity and task mix vary enormously by seniority, function, and industry even within adopting firms.
- Token consumption keeps growing substantially within already-adopted cohorts, not only via new adopters, suggesting continued deepening of use post-adoption.
- This pattern echoes historical general-purpose-technology diffusion (Griliches 1957; Mansfield 1961; Bresnahan & Trajtenberg 1995), which required decades of co-invention rather than delivering value immediately upon access.
Implicit premises:
- Sustained within-cohort token growth six-plus months post-adoption reflects genuine deepening of human engagement and integration, rather than being partly a mechanical artifact of model changes (e.g., contemporaneous shifts toward longer, reasoning-heavy outputs that generate more tokens per query without more human usage or value).
- The specific worker-composition and task-mix heterogeneity documented here is itself evidence of 'incomplete deployment' rather than simply evidence of a technology that has already reached its natural, differentiated steady-state use pattern across roles.
Narratives
- The electrification and early-computerization analogies (via Bresnahan & Trajtenberg, Griliches, Mansfield) — factories bolted electric motors onto old shaft-and-belt layouts and saw little productivity gain until plants were physically redesigned around distributed power; the Solow computer productivity paradox resolved only once organizations reorganized around what computers made possible.
→ Carries the paper's central 'adoption ≠ deployment, diffusion takes decades of co-invention' claim by historical analogy rather than by direct argument from the paper's own (six-month-horizon) data, which cannot itself observe decades-long diffusion.
Questions addressed
- ✅ Which firm characteristics predict ChatGPT Enterprise adoption among U.S. public companies?
Firm size, revenue-per-employee, and pre-existing intangible/organizational capital stocks (SG&A, R&D, capitalized software) all predict adoption; physical-capital intensity is negatively associated conditional on other factors. - ✅ Which workers within adopting organizations account for observed ChatGPT Enterprise use, and how intensively do they use it?
Use is broad across job functions and seniority levels, but usage intensity is uneven and shows a strong negative seniority gradient (early-career workers most intensive per-user). - ✅ What kinds of tasks does enterprise generative AI use span?
A long tail of ~60 task categories, dominated in message volume by documentation/technical writing, technical digital work, and communication, with prevalence varying more by industry/role than intensity does. - ⭕ Does ChatGPT Enterprise adoption or usage intensity causally improve firm financial performance?
Explicitly left open — the paper repeatedly states its firm-financial-characteristic associations 'should not be interpreted causally' and calls for future work connecting telemetry to productivity and organizational-change outcomes. - ⭕ Whether the sevenfold aggregate token-growth figure reflects genuinely deepening human usage/deployment, or is partly inflated by contemporaneous shifts in how many tokens a model generates per query (e.g., more verbose reasoning-style outputs) during the 2025-2026 sample window.
Not addressed; the paper treats token growth as a clean proxy for usage growth throughout Section 4.1 without discussing token-per-query drift as a confound.
Gaps
- incompleteThe paper measures usage only within OpenAI's own ChatGPT Enterprise product, explicitly excluding personal accounts, API usage, competitor AI tools (Claude, Gemini, Microsoft Copilot), and internally built AI applications. This is a significant blind spot precisely for the large, R&D-intensive, technically sophisticated firms the paper studies, since such firms are the ones most likely to run multi-vendor or custom AI stacks rather than relying solely on ChatGPT Enterprise. The paper's own footnote 10 concedes that its 'non-adopter' comparison group may include heavy users of other AI products, which would attenuate (understate) the very adopter/non-adopter differences the paper documents as facts.
Read next: Cross-vendor telemetry comparisons (e.g., a similar study using Microsoft 365 Copilot data — Counts et al. 2026, cited in the paper's own related-literature section — or firm-level surveys of the full multi-vendor AI stack) that don't restrict measurement to a single provider's product. - incompleteThe paper cannot separate token-volume growth driven by more/deeper human-initiated usage from token-volume growth driven by model-generation changes — e.g., longer reasoning traces or more verbose outputs per query as GPT models shifted toward chain-of-thought reasoning across the 2025-2026 sample window. Output tokens per message could rise mechanically even with roughly flat human engagement, inflating the headline 'sevenfold' and 'fourfold within-cohort' growth figures independent of any real deepening of organizational deployment.
Read next: Documentation or research on how output-tokens-per-query changed across OpenAI model generations over this period, or a robustness check using message counts (rather than tokens) as the primary growth metric — the paper does report messages elsewhere but leads with tokens for its headline growth claim. - incompleteThe paper documents that early-career workers are the most intensive per-user users of ChatGPT Enterprise, but does not connect this to the parallel, more alarming finding elsewhere in the literature it cites (Brynjolfsson et al. 2025a) that early-career/entry-level employment itself may be shrinking under AI exposure. It leaves entirely open whether heavy per-user usage among the early-career workers who remain employed reflects healthy skill-building and agent-native fluency, or reflects AI absorbing exactly the repetitive, information-synthesis tasks that used to be assigned to juniors — accelerating headcount reduction at that level even as remaining juniors use the tool the most.
Read next: Read this paper's task taxonomy alongside labor-market studies of AI-exposed entry-level employment (e.g., this corpus's extract-loudest-warning-ai-jobs, and the underlying Brynjolfsson et al. 2025a paper) to test whether the task categories where early-career workers are heaviest users are the same categories associated with entry-level job losses. - illogicalThe financial usage-intensity analysis (Figure 4, Table 2) is careful in prose to disclaim causality, but its visual and rhetorical presentation ('high-intensity adopters are shifted to the right') still invites readers to treat usage intensity as doing genuine explanatory work for firm value, without seriously entertaining the confound that firms with cash-rich, technically sophisticated, high-margin workforces would plausibly show both higher revenue/market-value per employee AND higher AI usage intensity for reasons unrelated to any causal effect of AI itself (e.g., general workforce quality or industry structure).
Read next: A paper or extension using instrumental variables, a natural experiment (e.g., staggered enterprise rollout timing within a firm), or panel fixed-effects design to isolate whether usage intensity has any causal financial effect, rather than firm-level cross-sectional correlation.
Related documents
- The loudest warning about AI and jobs yet
- Ordinary Engineers, Not Heroic Inventors
- Simon Wardley's Post — Industrialisation of ML: nine notes on what to expect
- Itchy Brain (Still Burning: Michael Grinich & Kent Beck on AI adoption, enterprise developer tools)
- AI productivity is burning out your best engineers
- Google's AlphaEvolve Reaches General Availability with Evolutionary Code Optimization as a Service
- Arvind Narayanan on Why AI Isn't All That Revolutionary
- 2024_final_dora_report
- Measuring AI impact, assessing readiness, and new data trends
- Why Enterprise AI Has a Leadership Problem