Uber has switched on ChatGPT Work for its whole workforce, more than 30,000 people, after trying it with 1,000. The agent connects to Gmail, Google Drive, Google Calendar and Slack, and carries a task through many steps to a finished sheet, slide deck or document. In the pilot, four in five users were not engineers. They analysed operations data, translated training material for different markets, and built dashboards for their own daily work. Uber has left the exact uses open and will watch what people build.
The hard part of automating a job is knowing the job
Uber's earlier experiment was called an agentic pod. Each pod was an AI engineer working for two weeks beside someone who does a business job, in areas like legal, HR, procurement or customer support. In customer support, about 9,000 workflows built by hand were replaced by self-service automation.
The pods started with two days of watching because a job on paper is not the job. A process diagram shows the official steps. The person doing the work knows the supplier whose invoice always needs a phone call, and the column in the weekly export that is wrong every Monday. Nobody lists these things, so nobody can hand them to a programmer. An engineer has to sit and watch until they appear.
With an agent in everyone's hands, that watching step moves to the expert. They can describe the job as they do it, point the agent at the same files and inbox, and correct it when it misses an exception. Uber is not dropping the support around this: the rollout comes with a library of AI skills, approved connections to its systems and training.
The spreadsheet went the same way
Before 1979, a financial model was built on paper with pencils, and junior staff and typing pools produced the numbers a senior person asked for. When VisiCalc arrived, accountants were the first to buy it, because they could finally test a "what if" themselves without asking anyone.
The work did not disappear. Modelling became cheap, so companies asked for far more of it, and the number of accountants grew many times over in the following decades. The same pattern fits here. When a weekly report or a dashboard costs ten minutes instead of two days, a business will want many more of them, and the people who know what a useful one looks like will be asked for more, not less.
What a finance team could get back
Take a finance analyst at a hospital group who spends two working days every month pulling figures from three systems, reconciling them, writing comments and emailing department heads. Uber's pods took reporting jobs of this kind from days to minutes. If the analyst gets the same result, with an hour left for checking the numbers, that is more than 20 working days a year returned to one person, about a month. The analyst still decides which figures can be trusted and what the comments should say. The agent only does the pulling and assembling.
This also follows what we reported when ChatGPT Work launched: it is built to take on whole projects, not single questions.
The last stage of the pods is worth copying, and it costs nothing. The last days of each pod went on testing the agent with other people who do the same job, to see whether it worked beyond one person's habits. Before you automate your own routine, write down every step the next time you do it. Then ask a colleague who does the same task differently to run the result. Where their way differs from yours, you have found the exceptions the agent needs to know about.