Uber's First Agent Numbers: Four Rides Out of 320 Million

Uber just published the first hard number on agentic commerce, and it is far smaller than the industry narrative suggested. On its third-quarter earnings call, the company said third-party AI agents arranged four rides in the quarter. Uber Eats did slightly better: around 500 agent-placed orders during the week of Halloween.

What Uber actually disclosed

Out of roughly 320 million trips in one of its biggest weeks, agents produced four completed rides and a few hundred food orders. It is the first time Uber has reported third-party agent volume at all, according to The Verge.

CEO Dara Khosrowshahi did not frame it as a failure. "We hope that there will be more traction," he said on the call. "It is just very, very early, and we want to make sure we are not dependent on any one player, and I think we are very well positioned in this environment going forward."

The disclosure lands in a year of loud agent launches, including Meta's Muse and OpenAI's Dots, and predictions that marketplaces such as Uber and DoorDash face disintermediation: customers asking an assistant to order or book instead of opening the app. Uber also keeps building its own AI inside the product, including a Cart Assistant for Uber Eats that turns a photo of a recipe into a grocery list.

Why the transaction layer is the hard part

The roughly 125-to-1 gap between food orders and rides is the most instructive detail in the numbers. Delivery is a bounded, delay-tolerant task: pick a restaurant, accept an arrival window, wait. A ride needs live location, an exact pickup point, price confirmation and a payment handoff, and a mistake strands a person on a curb.

That is precisely where agent stacks break. Tool calls fail, sessions expire, and the confirmation step still wants a human. Most agent demos stop at the recommendation; the last mile of completing and paying for a transaction is where the volume collapses. Platform incentives compound the problem. Marketplaces have little reason to let an outside assistant own the customer relationship, and Uber said it will work with third-party AI companies without depending on any single one.

What this means for builders

Treat agentic commerce as an integration problem, not a model problem. If you ship agents that book, order or pay on someone's behalf, instrument the funnel step by step and count fulfilled actions rather than attempts. Assume a messy API, a re-authentication wall and an uncooperative platform, and design the confirmation step first instead of bolting it on.

The counter-reading matters too. Four rides is a baseline, not a verdict, and the company is now measuring it publicly. The next twelve months of these numbers will be the first real scoreboard for agents in production, and vendors who can show completed transactions rather than polished demos will have the better story.