Uber brings an AI agent to more than 30,000 employees: what to learn

What happened

On October 22, 2026, OpenAI reported that Uber began rolling out ChatGPT Work across the company. ChatGPT Work is an agent inside ChatGPT Enterprise designed for multi-step tasks that end in finished work.

According to the announcement:

• The rollout reaches more than 30,000 employees across Uber’s global workforce.
• It expands from an initial pilot of 1,000 employees.
• The agent connects with everyday tools such as Google Drive, Gmail, Google Calendar and Slack, to research, prepare meeting briefs, create presentations, analyze data and complete workflows across applications.
• The rollout is supported by an internal AI skills library, approved integrations and dedicated training for employees, in areas such as operations, finance, marketing and product.
• Uber’s development teams continue to use Codex for programming.

Both companies said they will share more about adoption, use cases and measurable outcomes as the rollout progresses. For now, the announcement includes no impact figures.

Why it matters

It is a case of AI agents taken to scale in a large company, not a lab experiment. And the order of the steps says a lot: first a bounded pilot, then the expansion, and in both cases with defined support pieces.

Three elements of the announcement are worth a close look:

1. Approved integrations. The agent does not connect to just anything. The company decides which systems it can access. That is a governance decision, not a technical one.
2. A skills library. Instead of each person inventing their own instructions, there is a shared repository of ways to work with AI. That shortens the learning curve and makes good uses repeatable.
3. Dedicated training. Giving access is not enough. The announcement mentions specific training so teams use the tool in their work.

What is still missing also stands out: published results. Uber and OpenAI say they will share them later. Not announcing impact before it has been measured is good practice.

What a company can do

You do not need 30,000 employees to apply the same logic. For a mid-sized company, the path looks like this:

1. Start with a small pilot that has an owner. One area, one process and one person in charge. Measure before expanding.
2. Decide which systems the agent can touch. Make a short list of allowed integrations and review which data is exposed.
3. Document the uses that work. An internal library, even a shared document, keeps each person from starting from scratch.
4. Human oversight on sensitive tasks. An agent that runs multi-step tasks should have review points where a person approves before it continues.

Our take

The value of this announcement lies less in the agent’s brand and more in the method: pilot, controlled integrations, shared knowledge and training. It is the same order we recommend to companies that want to move from trying agents to using them in their operations. The tool changes every few months; that method keeps working.