This paper studies how firms use a leading generative AI product, ChatGPT Enterprise, by linking account records to worker roles, message-level task labels, and public-company financial data through March 2026. The authors analyze large, privacy-preserving telemetry datasets. For example, their worker-level sample at a six-month adoption horizon covers over 1,500 organizations and more than 17 million messages, allowing them to observe who used the product and what kinds of tasks it was used for.

To get these insights the researchers combined several linked samples. They decomposed growth in overall use into new firm adoption versus growing intensity among existing customers. They also matched ChatGPT Enterprise accounts to Compustat financial records for public U.S. companies, and they classified individual messages by task to see what users actually asked the system to do. The study emphasizes telemetry data as a complement to surveys, because it records actual use rather than relying on self-reports.

They report four main findings. First, usage grew rapidly: aggregate output tokens from enterprise customers rose about sevenfold between June 2025 and March 2026, with nearly a fourfold rise even inside a cohort of firms that adopted between January 2024 and June 2025. That means much growth came from heavier use within firms as well as new firms signing up. Second, among U.S. public companies, adoption was concentrated among larger, more valuable firms and those that spend more on research & development (R&D) and on selling, general, and administrative expenses (SG&A).

Third, active use inside adopting firms is spread across job functions and seniority levels, not limited to a single team. Some groups—marketing and communications, for example—sent more messages than executives, and early-career workers sent many more messages than senior employees. Fourth, the tasks covered a broad set of knowledge-work activities. The most common uses were writing, communication, and information synthesis, but researchers also saw use for research, planning, data analysis, legal and regulatory work, finance, and other applications.

These patterns matter because they show enterprise AI adoption is broad but uneven. Early adoption is concentrated in firms with greater scale and prior investment in intangible capabilities. At the same time, use within firms is dispersed across many roles and tasks. The paper highlights that getting real value from a general-purpose AI likely requires experimentation and complementary organizational investments. In other words, simply buying access to an AI tool does not guarantee effective deployment across a firm.

Important caveats are noted. The paper is a working paper and the authors say results are subject to change. The data cover OpenAI’s centrally administered workplace offering and are dominated by ChatGPT usage; they do not capture every AI product firms might use. Telemetry shows observed activity but does not by itself prove longer-term impacts on productivity, jobs, or firm performance. The authors also note that realizing value from AI often depends on follow-up investments and organizational change, which their measurements can document only indirectly.