Uber's AI pilot shows how to roll out ChatGPT right

Thirty thousand employees at Uber now have access to ChatGPT Work, OpenAI's enterprise agent tool, after a pilot with about 1,000 workers convinced the company to scale up this week.

The AI pilot before the rollout

Uber and OpenAI announced the expanded partnership on Thursday. ChatGPT Work is OpenAI's enterprise agent product, built to handle research, analysis, and document work across a company's tools. Uber first tested it with roughly 1,000 people in finance, operations, and product before opening it to everyone, and the company has not said what it is spending on the rollout or settled on required uses. Sumanth Sukumar, Uber's VP of engineering, said about 80% of the early adopters were non-engineers, using the tool to dig through operational data, translate training materials for different markets, and build dashboards for their own work. None of that needed a developer. That is the detail small businesses tend to skip: the pilot was run and judged by the people doing the job, not by whoever installed the software.

Why no one had to use it

Sukumar described Uber's approach as making the tool available and giving teams room to experiment rather than mandating how to use it. Uber tracked what workflows people actually built and how those changed their output, then used that evidence to decide on the wider deployment. A mandate measures compliance. A pilot with no required use case measures whether the tool is actually worth the seat.

AI pods: a pattern worth stealing

Before the full rollout, Uber paired 30 AI engineers with individual teams for two weeks at a time, calling them AI pods. Each pod combined someone who knew the tool with someone who knew the job, and the pairing was time-boxed. A small business does not need 30 engineers to copy the shape of that: one person who understands the tool, one person who owns the task, two weeks, and a specific job to automate or speed up.

1. 1

   Pick one team, not everyone

   Choose a single team with a real, recurring task, not a company-wide announcement. Uber's pilot was about a thousand people out of a global workforce, not a blanket rollout.

2. 2

   Set a two-week clock

   Give the pairing a hard end date. An open-ended trial never produces a verdict. A two-week pod either changed how the work got done or it didn't.

3. 3

   Let the team pick the workflow

   Don't hand down a required use case. Tell the team what tool they have and let them decide what to point it at. The workflows that stick are the ones someone built because it solved their own problem.

4. 4

   Price the seats before you scale

   Uber didn't say what it's paying for 30,000 seats. You should know what a handful of seats cost before you ask for budget to cover the whole team, and decide on a seat count based on what the pilot actually used, not what sounded impressive.

If you're looking at a tool like this and wondering whether it fits how your team actually works, I'm happy to look at the workflow with you before you buy a single seat.