Enterprise AI Use Spreads Across Roles but Varies Widely

Enterprise generative-AI adoption is often measured through surveys, public announcements, or aggregate traffic, leaving little evidence about who actually uses a system inside a firm and for what work. The authors link ChatGPT Enterprise account records to message-level usage, worker roles, task classifications, and public-company financial data through March 2026. Their central finding is not a single productivity estimate: organizations are adopting and using the product at very different speeds, across a broad set of knowledge-work roles and tasks.

Core Contribution

The contribution is a privacy-preserving measurement framework for enterprise AI use rather than a causal evaluation of AI's effect on firms or workers. The paper builds adoption and usage samples from ChatGPT Enterprise records, then relates them to organizational characteristics, job titles, seniority, and message-level task labels. At the six-month adoption horizon, the worker-level sample contains more than 1,500 organizations and more than 17 million messages.

That scale supports several descriptive facts that are difficult to establish from public data alone. Enterprise output-token use rises through both entry by new organizations and increasing intensity among organizations already on the service. Among U.S. public firms, adopters are disproportionately larger, more valuable, and more intensive in R&D and SG&A. Within adopting organizations, active users appear in many job functions and seniority levels, while early-career workers show especially high use intensity.

Technical Approach

The sample-construction diagram makes the paper's design clear: account and activity records are filtered into firm-, worker-, and message-level analytic samples, then joined to external public-company data where identifiers permit. This creates an unusually granular view of enterprise deployment, but it also means the results describe organizations using ChatGPT Enterprise rather than enterprise AI in general.

Figure 1 documents that construction and the distinct samples used for adoption, worker-role, and task analyses.

For the worker analysis, the authors classify job titles into occupational and seniority categories. The appendix details title-classification guidelines and a taxonomy used to separate functions and career levels. For messages, the paper applies a task classifier to characterize the work being performed. The resulting categories include writing, technical work, communication, and information synthesis, allowing the authors to analyze use by task rather than treating every prompt as equivalent.

The visual evidence emphasizes distributions rather than a representative "average employee." Composition charts compare active users across worker types, while intensity plots show meaningful differences by role. Task figures similarly show a broad allocation of use across knowledge-work activities and variation across industries. This is a useful design choice: token volume alone would obscure whether a deployment is concentrated in one technical team or spread across organizational functions.

Results and Analysis

The strongest directly supported result is descriptive scale and breadth. The paper analyzes over 17 million messages from more than 1,500 organizations at six months after adoption, and the usage series shows continued growth in enterprise output tokens. The authors attribute that growth to two observable margins: more organizations adopting the product and rising usage among existing adopters. That distinction matters because it separates market expansion from deeper integration within already-adopting firms.

The public-company comparison supports a second claim: adoption is not evenly distributed across firms. The adoption sample is concentrated among companies with greater size, market value, R&D spending, and SG&A spending than other public firms. The ECDF presentation is appropriate here because it reveals distributional separation rather than relying on a single mean. But the paper does not establish that those characteristics cause adoption; they may instead mark firms with the budget, digital infrastructure, or governance capacity to procure enterprise AI.

Figure 7 shows why a narrow coding-assistant narrative would be incomplete. Usage spans writing, technical work, communication, and information synthesis, with industry-specific differences in task mix.

The worker findings are also more nuanced than a simple substitution story. Early-career workers have particularly high usage intensity, yet active use spans job functions and seniority levels. That pattern is consistent with broad experimentation and workflow learning, not proof that AI has displaced particular occupations or raised output. The paper's value is therefore empirical visibility into deployment behavior. Its evidence is substantial for describing ChatGPT Enterprise use, but not for making causal claims about productivity, employment, or firm performance.

Limits of the Evidence

The data cover one enterprise product and a selected population of adopting organizations, so they cannot represent all firms or all generative-AI tools. Public-firm financial matches further narrow part of the analysis. Message and title classifiers make large-scale measurement possible, but classification choices can affect role and task estimates. Finally, the six-month horizon captures early deployment patterns; it cannot settle whether usage persists, changes organizational design, or produces measurable financial returns.