Uber is putting ChatGPT Work, OpenAI’s enterprise agent, in front of more than 30,000 employees. The pilot had 1,000 people. OpenAI calls this Uber’s first company-wide deployment of an agentic platform for knowledge work, and Uber hasn’t said what it costs.
My view: the seat count matters least here. Handing everyone an agent is easy. What happened in the pilot, and in a side experiment Uber mentioned almost in passing, is what decides whether this pays off.
What the agent can do
ChatGPT Work launched in July as an agent inside ChatGPT Enterprise. It takes multi-step tasks and is supposed to return finished work. At Uber it connects to Google Drive, Gmail, Google Calendar and Slack, so it can pull research together, prepare meeting briefs, build presentations and analyze data across those apps.
The default model is GPT-6.1 Sol, which OpenAI describes as “balancing frontier performance and cost efficiency.” That wording is the most honest line in the announcement. At 30,000 seats, cost per task matters more than benchmark scores, and OpenAI is steering big customers toward a model that won’t blow up the bill. Voice and image generation are included. Uber’s engineers keep using Codex for coding, so ChatGPT Work is aimed at everyone else.
The pilot was mostly non-engineers
Uber VP of engineering Sumanth Sukumar told TechCrunch that about 80% of the pilot users were non-engineers. They came from finance, operations and product. They used the agent to analyze operational data and recommend changes, localize training materials for different markets and languages, and build personalized interactive dashboards.
None of that is exotic, and that’s fine. Localizing training docs across languages is tedious, hard to staff, and easy to check. It suits an agent much better than vague “strategy” work.
The pods were the better idea
Sukumar also described an earlier experiment. Uber paired 30 AI engineers with teams across the company and gave each pairing two weeks to automate key parts of a job. He said these “AI pods” showed how much opportunity opens up when technical skill meets someone who knows the job in detail.
I’d put this ahead of the rollout itself. A general agent with Slack and Gmail access is a capable blank page. Most people faced with a blank page write a summary of an email thread and stop there. The pods had a deadline, a specific job, and a person who knew where the integrations break. They show that the bottleneck is someone sitting down and redesigning one specific workflow, and better models don’t fix that.
Uber is building some support around the rollout: an internal AI skills library, approved integrations, and employee training. That helps. A library of skills is still a weaker version of an engineer sitting with the payments team for two weeks, and with 30 pod engineers against 30,000 staff, the scale doesn’t match.
Deploy first, measure later
Uber admits it hasn’t settled how employees should use the tool. Sukumar said the company will watch which workflows people build, evaluate the outcomes, and use that to guide wider deployment. OpenAI says both companies will share adoption data, use cases and “measurable outcomes” as the rollout goes on.
So the measurement plan comes after the purchase. That’s common in enterprise AI right now, and with OpenAI competing hard with Anthropic for business customers, the vendor has every reason to encourage it. I’d want the numbers before buying, not after.
When those numbers arrive, check whether they come from the 80% of non-engineers working on their own, or from teams that got time with a pod engineer.