The usage logs from Uber Technologies’ first ChatGPT trial did not read like a software rollout inside an engineering department. Of the roughly 1,000 employees in the pilot, about eight in ten were not engineers. They were people summarizing contracts, reconciling spreadsheets and assembling small automations for teams that had never filed a ticket with corporate IT.
On Thursday, Uber and OpenAI said that trial is becoming standard issue. ChatGPT Work, the business version of OpenAI’s chatbot, will be offered to about 30,000 Uber employees worldwide, up from the 1,000 who took part in the pilot. The two companies described it as one of the largest deployments of OpenAI’s enterprise product at a single customer.
Neither company would say what Uber pays per seat, how long the contract runs, or which customer ranks second. A person familiar with the arrangement said the agreement covers global staff and follows the shape of the pilot: broad access, light mandatory training, and written rules about which company data can be pasted into a prompt.
The mix of users is the part Uber’s executives point to. Early adopters spanned finance, legal, operations and recruiting, according to the company, and the most common tasks were document handling, data analysis and workflow automation. In practice that means drafting job descriptions, cleaning customer-feedback files, turning meeting notes into task lists, and writing the small scripts that move data between internal tools.
That is a different sales story from the one enterprise software vendors told for the past two years, when AI assistants were pitched mainly as coding tools and priced against engineering budgets. Coding seats are easy to defend in a review: they map to headcount and to output that can be counted. A seat for a contracts manager or a recruiter is harder to justify, which is why adoption inside large companies has been uneven even where licenses were already bought and paid for.
OpenAI has been pushing in that direction. The company has spent the past year adding admin controls, compliance certifications and connectors to corporate data sources, the features procurement departments ask about before they sign. Its enterprise business competes with Microsoft’s Copilot bundle, Anthropic’s Claude for Enterprise and Google’s Gemini offerings, all of which argue that AI belongs inside the productivity software people already use rather than in a separate tab.
Uber’s own history makes it an unusual reference customer. The company has used machine learning for years in pricing, matching and delivery-time estimates, and it runs large data teams in Bangalore and San Francisco. Its appetite for internal tooling is well documented. What it has not published, either for this rollout or earlier ones, is a productivity measure. Seat counts are easy to announce; returns are not.
Analysts said the announcement will be read as a reference win, because large employers watch what peer companies do with headcount and software budgets. A deployment spanning a workforce of that size gives OpenAI a number to put in front of other chief information officers, and a named customer in a category, ride-hailing, where the work is operationally messy and the documents are dull.
The risks sit where they usually do when a tool goes broad. Data handling is the first. Prompts that contain customer complaints, driver records or contract terms leave the company’s own systems and land in a vendor’s logs, governed by settings that few employees read. Uber said the program includes training and acceptable-use rules, and that access is tied to existing corporate identity controls.
Accuracy is the second. A fluent answer is not a checked answer, and the employees most likely to be handed the tool, recruiters, analysts, junior lawyers, are often the least equipped to spot a confident error. Companies that have run large rollouts report the same pattern: curiosity first, then a plateau, then a small group of heavy users who rebuild their working day around the tool while everyone else drifts back to email.
There is also the European labor question. A rollout of this size in Germany or the Netherlands passes through works councils, which must be consulted on software that can monitor or evaluate employees. Uber has said little about how the tool is governed in those markets, and the company’s relationship with organized labor has been contested for years.
For OpenAI, the arithmetic matters as much as the publicity. Enterprise revenue renews on contracts rather than on monthly consumer churn, and the company has been trying to show that businesses will pay list price per seat instead of negotiating for bulk discounts. A customer of Uber’s size gives it a case study with scale attached, at a moment when every large model developer is chasing the same corporate budgets.
Competitors will answer with bundles. Microsoft can fold assistant features into licenses enterprises already hold, Google can do the same through Workspace, and both can point to years of data-residency arrangements that a startup has had to build from scratch. The counterargument from OpenAI is capability: that a model sold on its own merits, without a suite attached, is the one employees actually open.
What happens after the launch will be watched more closely than the launch itself. Companies that hand out AI licenses often see usage spike and then flatten as novelty fades, and renewal decisions turn on whether the tool changed how work gets done or merely added a tab to the browser. Uber has not said whether it will publish usage or productivity data, or how it will decide whether 30,000 seats were worth the money.
