OpenAI Has Been Quietly Studying How People Actually Use AI at Work
OpenAI has spent the past year quietly building a detailed collection of research on how its own products are used at work. The body of work now spans seven published reports, starting with consumer usage patterns in September 2025 and ending with two August 2026 releases: a study of how organizations use ChatGPT Enterprise that links account records to worker roles and tasks, and a companion piece on how businesses use AI agents through ChatGPT and Codex. Together, these reports tell a data-driven story about where generative AI is actually landing inside organizations, not where vendors claim it will.
The most recent report, "How Organizations Use AI: Evidence from ChatGPT," published August 13, 2026, links ChatGPT Enterprise account records directly to usage patterns, worker roles, and specific tasks. This goes beyond an earlier consumer-focused paper, "How People Use ChatGPT," published September 15, 2025, which documented the growth of ChatGPT's consumer usage broadly. The new report moves from overall usage numbers into the detail of who, within an organization, is doing what with the tool. OpenAI Research
A day earlier, on August 12, 2026, OpenAI published "From assistance to execution: How enterprises put AI to work," focusing on agentic AI adoption through ChatGPT and Codex. An AI agent is software that can carry out a multi-step task on its own, rather than just answering a single question. The title itself signals the direction OpenAI's data is tracking: usage is shifting from conversation toward tasks that the AI handles with some independence. OpenAI
These two August reports build on a foundation laid throughout late 2025 and the first half of 2026. "The State of Enterprise AI 2025" report, published December 8, 2025, identified accelerating AI adoption, deeper integration into existing workflows, and measurable productivity gains across enterprises. A January 22, 2026 guide, "ChatGPT usage and adoption patterns at work," provided departmental and industry-level breakdowns of adoption trends and top tasks. On July 27, 2026, "How AI is expanding what people do at work" showed users taking on tasks across roles, reshaping job boundaries. And on August 5, 2026, "From asking to doing: How the world is putting ChatGPT to work" drew on OpenAI Signals data to provide country-level adoption insights. OpenAI State of Enterprise AI | ChatGPT Usage Guide | How AI is Expanding Work | From Asking to Doing
The progression across these seven publications tells its own story. The consumer paper asked how many people use ChatGPT and for what broad purposes. The enterprise state report asked whether organizations are seeing returns. The adoption patterns guide asked which departments and tasks dominate. The "expanding what people do" research asked whether roles themselves are shifting. The country-level report asked where geographically. And now the two August 2026 reports ask the most specific questions yet: what happens when you link account records to individual workers and their tasks, and what does independent AI task execution look like when enterprises deploy it through Codex?
For technology professionals, several threads in this research are worth particular attention. The linkage of Enterprise account records to usage, roles, and tasks in the August 13 report means OpenAI now has its own internal data on how organizations use its tools, at a level of detail that survey-based research cannot match. It is the difference between asking people in a questionnaire how they use a tool and reading the logs that show exactly what they did. The shift from assistance to execution described in the August 12 report aligns with the broader industry move toward AI agents, where the work being done changes from a single question-and-answer exchange to a multi-step task the AI handles with some independence.
It is worth noting that this is all vendor-published research, drawing exclusively from OpenAI's own product usage data. The approach gives it a level of detail and scale that independent academic studies struggle to achieve, but it also means the population is limited to OpenAI's customers. Organizations using competing platforms, or using open-weight models running on their own servers, are invisible to this data. The productivity gains and adoption trends described in these reports cover the OpenAI-using subset of the enterprise AI market, not the market as a whole.
That caveat does not diminish the value of what OpenAI has published. The trajectory from "How People Use ChatGPT" in September 2025 to "How Organizations Use AI" in August 2026 maps a year in which generative AI moved from a broadly understood consumer phenomenon into something organizations are measuring and building into defined workflows. The shift from assistance to execution, if the data in the August 12 report holds up under scrutiny, would point to a meaningful change in how these systems deliver value: not through individual question-and-answer cycles, but through delegated multi-step work.
For engineering leaders and platform teams deciding where to invest effort, the departmental and task-level patterns in the January 2026 adoption guide and the role-boundary findings from July 2026 offer a concrete map of where generative AI is already embedded versus where adoption is still early. The country-level data from the August 5 report adds geographic detail to what is otherwise an organizational view.
What remains absent from this body of work, at least from what has been published, is long-term outcome data: do the productivity gains documented in the December 2025 state report persist over quarters, or do they fade as the novelty wears off? Does the role expansion described in July hold up when measured against actual job redesign rather than what people self-report? These are the questions the next round of research will need to address, and the August 2026 reports' account-record approach suggests OpenAI is building the data infrastructure to answer them.
