How are companies actually using generative AI in their day-to-day work, and what are leading organisations doing differently? Two new studies from OpenAI provide unusually detailed insights into these questions. They are based on anonymised enterprise data. The accompanying working paper alone analyses more than 17 million messages from over 1,500 organisations, six months after they introduced ChatGPT. The following findings are particularly relevant to professional research:
AI is increasingly being used not just to answer individual questions, but to carry out entire workflows.
OpenAI’s AI agent Codex is becoming increasingly common in knowledge work. Since February 2026, the number of weekly active Codex users in companies has increased 108-fold in legal functions, 41-fold in sales, 41-fold in recruitment and 26-fold in marketing. OpenAI describes this as a shift from “assistance” to “execution”: systems research information across multiple sources, work with files and carry out multi-step tasks. AI use within companies is therefore becoming increasingly agent-based.
ChatGPT is already being used across a wide range of tasks.
Writing, communication and information synthesis are among the most common use cases. More than half of active users use ChatGPT for documentation or technical writing. Other use cases include research, planning, data analysis, and regulatory and financial tasks. Business and market research are used by a large number of people, although they tend to involve fewer interactions and generate fewer messages than activities such as writing.
Usage varies by function and differs between junior and senior employees.
Analysts and employees working in marketing and communications send an above-average number of messages to ChatGPT within their organisations. Among active users, early-career employees and trainees send around eight to nine more messages per week than the average user in their organisation. Managers, directors and executives use the system less intensively and for somewhat different purposes. Their interactions are comparatively more focused on specialist topics, facts and figures, and legal and financial information.
The gap between companies is widening.
AI usage among the top 10% of companies more than tripled between January and June 2026. OpenAI itself points out, however, that token volume is only a rough indicator of depth of use and is not a measure of the economic value generated.
More advanced organisations also integrate AI more closely with internal data and tools, and make much greater use of plugins and reusable skills. For research teams, an increasingly important question is therefore likely to be how successful individual AI workflows can be documented, reviewed and turned into repeatable research processes across the team.
Verification remains essential for professional research.
At the same time, the findings highlight a key limitation of using AI for research and information retrieval. OpenAI notes that knowledge-based work is harder to automate than areas such as programming because tasks are often more difficult to specify precisely and clear criteria for verifying results may be lacking. In professional research, source verification, quality control and human judgement therefore remain crucial.
By Maria Kleiner