Uber gave every employee the agent and left the method open

A switch can be thrown for a whole building before anyone has agreed which job it is supposed to change. The light comes on. The work on the desk is the same work, until somebody names the difference.

TechCrunch's account of 22 October 2026 is Uber telling that story in the middle of it. Sumanth Sukumar, vice president of engineering, is the person on the record. ChatGPT Work, an enterprise-focused agent for corporate workflows, had been introduced in July. Uber had piloted it with 1,000 employees earlier this year. The account says the whole global workforce, some 30,000 people, will now be using it. He is also plain that the company has not settled how.

The pilot was not an engineering trial with a business audience watching. It ran across finance, operations and product. A majority of the early adopters, some 80 percent, were not engineers. The work he describes is concrete. They analysed operational data for trends and recommendations, localised training materials for different markets and languages, and built interactive dashboards for the day. That is a range. It is not one workflow written down for everyone else to copy.

Three tasks are not a curriculum. Operational analysis, a localisation of training material, and a dashboard for one person's day ask for different checks before anyone trusts the output. The account lists them as what early adopters did. It does not say which of the three the rest of the company is expected to start with, or who signs the result. A company that hears "not engineers" and files the story under a general introduction has skipped the part where the work was different in each function.

The line he gives for the wider step is permission, not a playbook. "Our approach is to make powerful AI tools available across Uber while giving teams room to experiment." The example he reaches for is earlier, and smaller. This year the company paired 30 AI engineers with teams and gave them two weeks to automate key parts of the job. He calls those pairings AI pods. They showed, he says, how much opportunity sits where technical skill meets someone who already understands the work. The account does not say the pods are now the way 30,000 people will work. It offers them as evidence that the combination matters.

What the account withholds is as useful as what it names. Uber did not disclose the spend. Sukumar said the company will look at how employees use the technology, at the workflows they develop, and at whether the work gets more effective, and that those outcomes will guide what happens next. OpenAI, the account says, worked closely on the integration. None of the three measures is given a figure. A workforce total is not a finding about the work.

1. Do not file the seat as the method. The tools are being made available, and he says teams still have room to experiment. Settled would look like a named workflow, an owner and a result. This account has the availability.

2. Keep the pilot's mix attached to the story. Finance, operations and product were in the 1,000, and some 80 percent of early adopters were outside engineering. The tasks are analysis, localisation and a dashboard for the day. A single engineering demo does not stand in for that mix.

3. Leave the pods in their own box. Thirty engineers, paired with teams, for two weeks, earlier this year. He treats them as a sign of what happens when the person who knows the job is in the room. The account does not say that design now covers the company.

4. Ask for the measure he already named. Use, the workflows people actually develop, and whether the work is more effective. The spend is absent. So are the results. A later number has to be one of those three, or it is a different claim.

A critical thinking session is a place to write, for one team, what "more effective" has to mean before a seat is treated as the change.