Evidence at a glance
What Is Expanding First Is the Scope of the Test
On October 22, 2026, Uber announced that it would roll out ChatGPT Work to more than 30,000 employees worldwide. The agent runs within ChatGPT Enterprise and is aimed at knowledge workers who need to find information in everyday tools, assemble materials, and move multi-step tasks forward. Uber had previously run a pilot with about 1,000 employees, including teams in finance, operations, and product. OpenAI describes the expansion as Uber’s first company-wide deployment of an agent for knowledge work.
The figure of more than 30,000 describes the intended reach of the rollout. It does not mean that all those employees are already using the agent, and it says nothing by itself about productivity gains. The announcement provides no active-user rate, time-saved figures, deployment costs, or return on investment. The useful question is how the pilot is being extended into more working environments, and what evidence Uber will use to judge the result. Expansion starts the test rather than completing it.
Everyday Tool Connections Create a Route into Workflows
ChatGPT Work can connect to Google Drive, Gmail, Google Calendar, and Slack. Employees can use it to research information, prepare meeting briefs, create presentations, and analyze data. The product’s proposition is that an agent can do more than respond to a single prompt: it can gather information across applications and help with work that involves several steps. OpenAI says GPT-6.1 Sol will be the default capability, and that voice and image generation are also available for different tasks.
This design places value on work across tools, but connectors alone do not create a reliable workflow. Results still depend on whether the relevant information is accessible, the task is clearly defined, and employees can review and use what the system produces. Pilot examples included analyzing operational data for trends, localizing training materials for different markets and languages, and making interactive dashboards. These examples show what teams have tried. They do not establish that the tasks are now consistently automated or that they have produced measurable business gains.
Non-Engineer Use Moves Evaluation Beyond Code Generation
TechCrunch, citing Uber engineering vice president Sumanth Sukumar, reported that around 80% of early pilot users were not engineers. That share suggests the rollout is not only about speeding up coding in development teams. It also concerns how operations, finance, marketing, and product teams handle information-heavy work. For these employees, researching material, summarizing data, adapting training content, and preparing meeting documents may be more familiar tasks than generating code.
The professional mix of pilot users is a clue about adoption, not a measure of outcomes. An individual occasionally using an agent to produce a document is different from a team establishing a workflow that is repeatable and reviewable. Uber says it will observe how employees use the tool and what workflows emerge before assessing the results. That means the company is still learning about usage, rather than presenting a validated standard way of working.
Deployment Depends on Organizational Support as Well as Models
Uber plans to support the rollout with an internal AI skills library, approved integrations, and employee training. These measures address more than whether a model can generate content. They concern how employees fit the capability into existing work: which tasks suit an agent, which integrations are approved, and how generated work can return to a team’s process. The public materials do not describe the training in detail or explain how connector permissions are configured. They therefore do not establish that data-access boundaries are already clear.
The company has not consolidated every task into one supposed all-purpose assistant. Uber’s software development and engineering teams will continue to use Codex, while ChatGPT Work is being extended to broader knowledge work. Uber had also previously put 30 AI engineers together with company teams for two weeks to explore automating parts of their work. Taken together, these efforts suggest that enterprise deployment is not just a matter of provisioning a product. It also involves mapping tasks to workflows and keeping different tools available for different kinds of work.
After Expansion, Metrics and Permissions Matter More Than Claims
Uber and OpenAI say they plan to share adoption data, use cases, and measurable outcomes as the rollout progresses. The public materials do not yet provide an actual usage rate, deployment cost, or productivity figures, and TechCrunch reported that Uber had not disclosed what the rollout would cost. The number of accounts within reach can show how far deployment has extended, but it cannot stand in for evidence of results. Technology leaders need to distinguish sustained use from one-off trials, reusable workflows from individual experiments, and acceptable output from work that still needs substantial correction.
Permissions and data handling belong in the evaluation as well. Giving an agent access to Drive, Gmail, calendars, and Slack broadens the information it can work with, but public materials do not explain how Uber configures access or handles employee data. For other companies, the actionable lesson is not to copy a rollout figure of 30,000. Treat reach as a deployment metric, and repeated use, workflow reuse, quality, and cost as outcome measures. Decide whether to extend automation to higher-impact tasks only after access boundaries and human review are clear.