Uber Hands Its Agent to 30,000 People — and Keeps Codex for the Engineers

The interesting number in OpenAI’s announcement is not 30,000. It is 1,000 — the pilot Uber ran before it. Moving from a thousand people to more than thirty thousand is not a scaling exercise; it is the point at which the governance question stops being theoretical. A thousand volunteers self-select for novelty. Thirty thousand people across operations, finance, marketing and product will do whatever is easiest, and an agent platform without guardrails inherits that.

What Actually Rolled Out

Uber’s company-wide deployment of ChatGPT Work inside ChatGPT Enterprise is live now, reaching more than 30,000 employees across its global workforce, up from the 1,000-person pilot. The agent connects to the tools those employees already live in — Google Drive, Gmail, Google Calendar and Slack — and is meant to produce finished artifacts rather than answers: research, meeting briefs, presentations, data analysis, and multi-step workflows carried across applications.

The model choice is the part worth reading closely. ChatGPT Work will run GPT-6.1 Sol as its default capability inside Uber, which OpenAI frames as balancing frontier performance against cost efficiency. That is a procurement statement, not a marketing one: at thirty thousand seats, the gap between a frontier model and a cheaper one is a budget line.

Uber’s developers and software engineering teams keep using Codex. The split is deliberate and instructive — one agent for the knowledge workforce, another for the people who ship the code.

The Real Product Is the Skills Library

What makes the announcement more than a press release is what Uber built around the model. The rollout is supported by an internal AI skills library, a set of approved integrations, and dedicated employee training. That is the layer that turns a general agent into a company-specific one: the sanctioned way to call a tool, the known-good recipe for a month-end variance analysis, the list of things not to automate.

The result is a template that reads less like a product launch and more like a maturity curve. Uber already runs a multi-agent architecture for its driver-facing Uber Assistant, routing earnings questions, onboarding questions and transactional actions to different specialized systems under three stated principles — safety, trust and low latency — at a marketplace doing 40 million trips a day across 15,000 cities. The knowledge-work rollout sits on the opposite side of the company from that surface, but it answers to the same constraints.

OpenAI and Uber say they will share adoption data, use cases and measurable outcomes as the rollout progresses, which is the number that will decide whether this becomes a template for other enterprises or remains a Uber-specific story. The pilot-to-company-wide step is where most agent platforms get their real verdict: not whether the demo works, but whether thirty thousand people keep using it after the novelty wears off.